The Innovation Attorney Podcast

Executive Summary

A call center works by routing customer contacts through automatic call distribution and staffing queues with Erlang C mathematics, and queued for a customer hold time based on the company’s desired hold time.

Every inbound call arrives over a carrier connection, passes a session border controller, meets an interactive voice menu, and lands in a queue where a distribution engine matches it to an agent by skill, priority, and availability.26,21 Staffing follows from call volume, average handle time, and a service level target: 100 calls in a half hour at 300 seconds each, answered within 20 seconds 80 percent of the time, needs 21 agents on the phones and 30 on the roster after 30 percent shrinkage.18,23 Labor consumes roughly three quarters of the operating budget, and a 100 agent floor at the federal median wage carries $4,477,000 in annual wages before software, benefits, or facilities.23,1

Commercially, one major cloud platform lists seats from $75 to $240 per user per month and another meters voice at $0.038 per minute, so software costs 2.0 to 6.4 percent of the wage bill for 100 seats.29,27 Artificial intelligence is the variable in dispute. A forecast dated March 5, 2025 predicts that autonomous agents will resolve 80 percent of common issues by 2029, while a survey of 321 service leaders dated December 2, 2025 found that 20 percent had cut staffing because of AI.9,11 The Telephone Consumer Protection Act, the Telemarketing Sales Rule, California Penal Code § 632, and PCI DSS version 4.0.1 shape every design choice, and statutory damages run from $500 to $5,000 per violation.2,3,8,36

How Did the Modern Call Center Develop?

The modern call center developed from a century old queueing formula, a telephone network that moved from physical circuits to internet protocol, a labor market that moved offshore, and a body of consumer protection law that began to reach automated calling with the Telephone Consumer Protection Act of 1991.18,2

A. K. Erlang, a Danish mathematician, developed the mathematics that planners still use more than a century ago, and the Erlang C variant answers one planning question: given a rate of call arrivals and a mean handle time, how many agents hold the chance of waiting to the level a service target allows.18 As a matter of queueing theory, the formula assumes that callers wait as long as necessary, that no caller redials after hanging up, and that every agent works at the same pace. Real queues violate all three assumptions, which is why every planner who uses the formula adjusts its output with shrinkage and occupancy rules.

The telephone network underneath the queue changed more than the queue did. Enterprises ran calls over PRI circuits, fixed bundles of voice channels delivered by a carrier, and carriers have pushed customers to SIP trunks through price increases and abrupt sunset notices.21 A SIP trunk carries calls as internet protocol traffic, which lets a company size call paths to actual demand, centralize connections at a few data centers, and add redundancy by geography. Each SIP location generally needs its own session border controller, and porting numbers from PRI to SIP takes about 20 to 30 days per order.21

Cloud delivery then converted a capital purchase into a metered subscription. Vendors now publish list tiers priced per user per month, per voice minute, per chat message, and per email, and they sell the same platform under named seat, concurrent seat, and consumption licenses.29,27,32 One platform vendor describes an omnichannel cloud contact center as a single system that unifies voice, chat, email, messaging, and social interactions so that customers can switch channels without starting over.30

Labor moved to wherever a wage and a language skill could be combined. The Philippine IT and business process management industry grew from $22.9 billion in revenue and 1.15 million workers in 2016 to $40.3 billion and 1.89 million workers in 2025, an increase of 76 percent in revenue and 64 percent in headcount, and its association forecasts $42.3 billion for 2026.22Domestically, the Bureau of Labor Statistics counted about 2,666,000 customer service representatives in 2025.1

The law arrived in layers. The Telemarketing Sales Rule amendments of 2008 made the abandonment safe harbor enforceable on October 1, 2008, the prerecorded message opt out requirements on December 1, 2008, and the prior express written agreement requirement on September 1, 2009.3 The Federal Communications Commission ruled on February 8, 2024 that AI generated voices are artificial voices under the Telephone Consumer Protection Act.4,5 The Eleventh Circuit vacated the one to one consent requirement on January 24, 2025, PCI DSS version 4.0.1 became the live payment card standard on March 31, 2025, and a federal court in the Northern District of California let a class action against an AI customer service provider proceed on August 11, 2025.6,36,7

In three decades of negotiating technology licenses, telecommunications service terms, and SaaS agreements, I have watched one clause move between vendor and customer each time a pricing model changed: the definition of a billable unit. When a contact center contract meters by the minute, the definition of a minute decides who pays for ring time, queue time, and the seconds after the caller hangs up. That clause draws less attention at signing than any other commercial term and more disputes at renewal, and it matters more as AI platforms meter by the message, the token, and the resolution.

Detailed Findings

The six findings below answer the head query in the order a caller experiences it. A call travels through a defined chain of systems, a staffing formula decides how many people wait at the end of that chain, a cost model prices both, a resolution rate decides whether the call was worth taking, artificial intelligence is rewriting the price of the answer, and a set of statutes decides how every call may be placed, recorded, and paid for. Each finding opens with the established position, explains the mechanism, supplies the evidence, works an example with real numbers, states where sources disagree, and closes with the practical consequence for a decision maker.

How Does a Call Travel from a Carrier to an Agent Desktop?

Established Position

A call center delivers each contact to a qualified agent through a chain of carrier trunk, border controller, interactive voice menu, distribution engine, queue, and agent desktop, and every cloud platform reproduces that chain as software configuration.26,28 Four roles interact with the platform. Customers contact the center by any channel, agents handle and document contacts, supervisors monitor and coach, and administrators provision phone numbers, define queues and routing profiles, and build the flows that connect them.28 Operations also employ workforce management specialists, quality assurance analysts, and data analysts.33

Mechanism

The carrier hands the call to the enterprise as a SIP session, and the session border controller authenticates it, protects the network edge, and passes it to the contact center platform.21 The interactive voice menu then gathers identifiers. Caller ID supplies a number and sometimes a name, automatic number identification can pull the customer record from a database, and menu prompts can ask for an account number before any agent is involved.26

The distribution engine applies routing criteria such as agent skill, availability, language, VIP status, and location. Round robin spreads calls evenly, skills based routing scores agents against the call and picks the best match, and priority routing sends high priority callers to the most qualified agents first.26 When no agent is free, the caller joins a queue ordered by urgency, language, time of day, or day of week, and the system assigns the next queued call in real time as an agent becomes available. Computer telephony integration then pops the customer record onto the agent’s screen, and recording and quality monitoring tools run beside the call.34

Supporting Evidence

Cloud pricing pages show the unit of account for each step. One provider lists $0.038 per voice minute, $0.010 per chat message, $0.080 per email, and $0.014 per SMS message, with standard telephony rates added to voice.27 One enterprise suite stores the customer record in a single data layer, adds a generative assistant for suggested replies and conversation summaries, and routes requests from any channel to the best suited representative.31 On the measurement side, a 2024 industry survey reported that 85 percent of centers measure abandonment rate, 84 percent measure average handle time, 77 percent measure quality, and 76 percent measure average speed of answer.14

The network layer carries its own evidence. A published migration guide states that SIP typically requires fewer call paths than the PRI channels in a legacy distributed network, that IP access should be sized for total call path load under failover, and that failover behavior varies by provider and access type, with private circuits sometimes outperforming internet access.21

Worked Example

Take the 100 call half hour at 300 seconds of handle time. A caller dials a toll free number that was ported to SIP, the border controller admits the session, and the menu captures an account number. With 21 agents staffed, the Erlang C model gives an average speed of answer of 16.0 seconds across all calls, a 23.2 percent chance that a given call waits at all, and 82.6 percent of calls answered within 20 seconds.18 The call lasts five minutes, so a platform that meters voice at $0.038 per minute bills $0.19 for it, before the standard telephony charge.27

The model output comes from my own calculation using the Erlang C formula, and each of those figures changes if the handle time moves by even 30 seconds: at 360 seconds the same 100 calls need 25 agents.

Contested Areas

Vendors and practitioners disagree on how much of the chain to automate. The same industry survey that reports 85 percent measuring abandonment found that only 14 percent of centers measure deflection rate and 13 percent measure self service accessibility, and it cites a 2024 finding that 73 percent of customers try self service while 14 percent of issues are fully resolved through self service alone.14 A center that routes every caller through a deep menu therefore adds handling steps to a path that resolves one issue in seven. Whether the menu earns its place is an empirical question each center answers with its own data.

Practical Significance

Executives buying a platform are buying a configuration of this chain, so the contract should name each element, its metered unit, and its failover behavior. Founders building a service organization should expect the carrier migration and number porting to take weeks, and 20 to 30 days per porting order is a schedule risk for a launch date.21 Investors evaluating a contact center business can ask for the routing profile count, the share of calls resolved in the menu, and the average speed of answer by queue, because those three numbers describe how well the chain converts spend into resolved calls.

How Do Erlang C, Shrinkage, and Occupancy Set Staffing Levels?

Established Position

Staffing follows from three inputs, call volume, average handle time, and a service level target, and Erlang C converts them into the minimum number of agents who must be on the phones.18 Handle time is talk time plus after call work. A common target is 80 percent of calls answered within 20 seconds, and the formula returns the smallest agent count that meets it, with the caution that some calls still wait when the target is met.18

Mechanism

The formula first computes workload in Erlangs: calls per interval multiplied by handle time and divided by interval length. It then finds the probability that an arriving call must wait, and converts that probability into the share of calls answered inside the target time. The result describes agents on the phones. Shrinkage, the share of scheduled time unavailable for interactions because of breaks, training, meetings, and absence, converts that number into a roster, and occupancy, the share of available time agents spend handling contacts, reports how hard the staffed agents are working.23,19

Workforce management software wraps this arithmetic in a process: forecast demand by channel and work type, plan capacity, build schedules, manage intraday variances by reforecasting and reoptimizing breaks, and monitor adherence, with performance tracking and coaching on top.23

Supporting Evidence

Occupancy is defined as total handle time divided by total time available for work, and published practice puts voice occupancy generally between 80 and 90 percent, email between 90 and 100 percent, and chat above 100 percent because agents handle several chats at once.19 Historical data show a consistent link between occupancy and service level, and larger teams reach higher occupancy at the same service level.19 A quality benchmarking firm reports an average agent utilization of 71 percent, a world class level of 69 percent, a good range of 70 to 75 percent, and an association between utilization at or above 80 percent and high turnover.16 One vendor guide states that modern workforce tools forecast demand with up to 95 percent accuracy, and gives an example in which each 1 percent of forecast error costs about $400,000 a year.23

Worked Example

For 100 calls in 1,800 seconds at 300 seconds each, workload equals 100 times 300 divided by 1,800, which is 16.7 Erlangs. With 21 agents the model returns 82.6 percent of calls answered within 20 seconds and 79.4 percent occupancy. With 20 agents it returns 72.9 percent, so one agent separates a passing interval from a failing one. At 30 percent shrinkage the roster needs 21 divided by 0.70, or 30 agents.

Extend the same arithmetic. A load of 400 calls per half hour needs 74 agents and runs at 90.1 percent occupancy, while 50 calls per half hour needs 12 agents at 69.4 percent occupancy. Doubling a team’s volume raised occupancy by ten points at the same service level, which is the large team effect that makes splitting a floor into specialist groups expensive.19 A target of 90 percent within 20 seconds on the 100 call load needs 23 agents, two more than the 80 percent target requires.

Contested Areas

Sources disagree on the right occupancy. One practitioner reference recommends 80 to 90 percent for calls, and warns that high phone occupancy causes stress, exhaustion, and attrition, while a benchmarking firm treats 80 percent as the threshold at which turnover rises.19,16 Practitioners also disagree about the formula itself. Erlang C ignores abandonment and retries, which makes it conservative for centers where many callers hang up and optimistic for centers where those callers redial, and a planner who relies on it alone misstates the staffing need in both cases. The formula also treats all agents as interchangeable, while skills based routing splits one floor into several queues that each follow their own, less favorable, arithmetic.

Practical Significance

A planning meeting that debates headcount without stating the service level, the handle time, and the shrinkage assumption debates nothing. Executives should ask for the three inputs and the output interval by interval, because a 30 second change in handle time changed the required count from 21 to 25 in the example above. Investors can test a target company’s cost per contact by rebuilding its staffing from the same three inputs and comparing the roster to the payroll, and any gap of more than a few agents points to either unreported shrinkage or an unreported service level.

What Does It Cost to Run a Call Center, and How Do Vendors Charge?

Established Position

Labor is the dominant cost of a call center, with resources accounting for roughly three quarters of the operating budget, and software and telephony fill most of the remainder.23 The Bureau of Labor Statistics reported a median annual wage of $44,770, or $21.53 an hour, for customer service representatives in May 2025.1 Vendors price the platform by named seat, concurrent seat, or consumption, and outsourcing providers price the labor itself.32,22

Mechanism

Cost per contact divides total operating cost by contacts handled, and cost per call adjusted for resolution multiplies that figure by the average number of calls needed to resolve an inquiry, which charges the center for repeat calls.16 On the software side, a named license assigns one seat to one designated user, a concurrent license lets any number of users share seats provided they are not signed in at the same moment, and consumption billing charges by each user’s login time regardless of headcount.32

The breakeven between these models depends on hours. One vendor example finds named seats about 30 percent cheaper than consumption for a 1,000 agent team logged in about six hours a day, finds concurrent licenses cheaper only while no more than about 750 of 1,000 agents sign in together, and finds consumption pricing cheaper when daily logins fall below about six hours.32Artificial intelligence features add a fourth meter, because one platform allots a fixed pool of AI tokens per organization per month and charges for more.29

Supporting Evidence

One major platform lists four editions at $75, $115, $155, and $240 per user per month, billed annually, and includes 250 named or 350 concurrent AI tokens per organization per month on the first three editions.29 A pricing compilation of published list prices reports a median published tier of $135 for the same platform, $169 for a second major platform, $135 for a third, and $139 for a fourth, and notes that deals above 20 seats typically land 15 to 40 percent below list.25 Per unit prices on a metered platform run $0.038 per voice minute, $0.010 per chat message, $0.080 per email, and $0.014 per SMS message.27

Outsourced labor operates at industrial size. The Philippine IT and business process management industry reported $40.3 billion in revenue and 1.89 million workers in 2025.22 Attrition adds a separate cost line. Reported annual agent attrition was 35 percent in 2021 and 38 percent in 2022, replacement costs are reported at $10,000 to $20,000 in direct costs and up to $46,000 with lost productivity, and one vendor guide puts replacement at up to 1.5 times annual wages.24,23

Worked Example

Build a 100 agent floor at the federal median. Wages total $4,477,000 a year before benefits, payroll taxes, and facilities. Software seats at the $75 tier cost $90,000 a year and at the $240 tier cost $288,000, which is 2.0 percent and 6.4 percent of the wage bill. A buyer who negotiates 15 to 40 percent off the $155 tier pays $111,600 to $158,100 instead of $186,000.29,25

Attrition at 38 percent means 38 replacements a year. At $10,000 to $20,000 each the direct cost is $380,000 to $760,000, and at $46,000 each it reaches $1,748,000, so attrition consumes between 8.5 and 39.0 percent of the wage bill across that range. On the 100 call half hour from the staffing example, the roster of 30 agents at $21.53 an hour costs $322.95, or $3.23 per call in wages alone, and the voice platform fee for a five minute call adds $0.19. A fully loaded figure sits well above those numbers: a published cost per resolved call example shows $860,000 of operating cost over 100,000 calls and 1.4 calls per resolution, which produces $12.04.16

In my practice reviewing enterprise software procurement, the commitment schedule in a seat license attracts less attention than the price per seat. A vendor that offers better terms for a committed volume is selling a minimum payment, and the center that shrinks, whether through attrition, automation, or a lost client, still owes the commitment.32

Contested Areas

Every figure in the cost model carries a dispute. List prices are published tiers, and negotiated prices sit 15 to 40 percent below them for larger buyers, so the true price of a seat is a negotiated number no compilation publishes.25 Replacement cost estimates range from $10,000 to $46,000 depending on whether lost productivity is counted, and the aggregate figures for a 100 agent center differ by a factor of several between sources.24 Whether outsourcing lowers cost per resolved call depends on repeat call rates, which a labor rate comparison omits.

Practical Significance

Founders pricing a support function should model wages first, because software at list is 2.0 to 6.4 percent of wages for 100 seats and the choice of license type moves that figure by tens of thousands of dollars. Executives approving a platform contract should ask for the metered unit definitions, the discount schedule, and the exit terms on committed volume. Investors diligencing a center should test the attrition assumption against the replacement cost range, because a 38 percent rate at $46,000 a head consumes 39.0 percent of the wage bill.

How Do First Call Resolution and Quality Assurance Convert to Dollars?

Established Position

First call resolution converts to dollars because each repeat call consumes agent capacity that the center pays for and the customer has already paid for once.17 A quality benchmarking firm reports an industry average first call resolution of 71 percent and a world class standard of 80 percent, an average top box customer satisfaction of 77 percent against a world class 85 percent, and a world class cost per resolved call that runs 20 percent below the North American average.16

Mechanism

A repeat contact adds handle time without adding a resolved issue, so the same agent hours produce fewer resolutions. The benchmarking firm states that repeat calls make up 23 percent of the average center’s operating budget, that each 1 percent improvement in first call resolution reduces operating costs by 1 percent and raises customer satisfaction by 1 percent, and that satisfaction drops about 15 percent each time a customer must contact the organization again.17 The same firm warns that a shorter average handle time is no gain when rushed calls lower resolution and satisfaction.16

Quality assurance supplies the feedback loop. Manual review covers about 5 percent of calls, reviewers can score the same call differently, and sampled calls can over or under represent escalations. Speech analytics evaluates every interaction and turns the transcript into searchable data.20

Supporting Evidence

A collections company case study reports a 10 percent reduction in call center operating costs within about a year of deploying speech analytics, a second company reports a return on investment nine to twelve months after implementation, and an unnamed retailer reports cutting its quality staff in half while shifting analysts toward coaching.35 A workforce management guide reports an insurer that cut attrition 30 percent and saved $4.5 million, and an example in the same guide prices a 1 point drop in net promoter score at about $2 million a year.23 The vendor figures describe individual customers, not a measured average.

Customer side evidence points the same direction. In a 2026 survey of more than 11,000 consumers and leaders, 85 percent of customer experience leaders said customers will leave brands that cannot resolve an issue on first contact, and 74 percent of consumers reported frustration at repeating information.13 A 2024 industry survey found that 54 percent of agents remain past the two year mark.14

Worked Example

Use the published example of $860,000 in operating cost and 100,000 calls. At 1.4 calls per resolution the cost per resolved call is $12.04. If better resolution lowers that to 1.3 calls per resolution the figure falls to $11.18, a saving of $0.86 per resolved inquiry or 7.1 percent. Separately, if repeat calls consume 23 percent of the $860,000 budget, that share equals $197,800, and a 5 point gain in first call resolution from 71 to 76 percent would save 5 percent, or $43,000, if the benchmarking firm’s one for one relation holds in that center.16,17

Contested Areas

The benchmarking firm publishes its relation between resolution and cost without describing its sample or method, and the vendors that sell speech analytics supply the case studies that quantify its value.17,35 The manual sampling rate of 5 percent comes from a 2013 article that advocates speech analytics.20 Newer measurement practice also questions the legacy metrics. One operations guide pairs average handle time with customer effort score, occupancy with agent wellness and burnout risk, and call volume with AI containment rate, and the shift changes what a center optimizes.33 Only about 14 percent of centers measure deflection rate in the survey cited earlier, so most centers lack the data to test the claim.14

Practical Significance

A center that cannot state its repeat call rate cannot price its own quality program. Executives should require a first call resolution measure that counts customer recontact across channels, because a customer who calls, then chats, then emails has produced three contacts and one issue. Investors can use the 23 percent repeat call share as a diligence prompt: if a target reports a materially lower share, the measurement method deserves inspection, and if it reports a higher share, the quality program has room to pay for itself. Founders choosing between a full coverage analytics tool and a larger manual QA team should compare the quoted subscription cost with the loaded cost of the reviewers it replaces.

How Is Artificial Intelligence Changing the Economics of Contact Centers?

Established Position

Artificial intelligence changes contact center economics by replacing the wage cost of a routine contact with a metered fee for a minute, a message, or a resolution, and published forecasts and surveys disagree on how far that replacement will go and how fast.9,11 A forecast dated March 5, 2025 predicts that autonomous agents will resolve 80 percent of common service issues by 2029 and cut operational costs 30 percent.9 A survey of 321 service leaders conducted in October 2025 found that 20 percent had reduced agent staffing because of AI.11

Mechanism

Three mechanisms operate at once. Agent assist tools suggest responses, summarize conversations, and surface knowledge during a call, which shortens handle time.31 Autonomous agents resolve contacts without a human, which removes the contact from the Erlang C workload entirely. Automation of back office work such as quality scoring and workforce operations changes the cost of supervision.31,23 Each mechanism changes a different term in the staffing equation, and each carries its own price meter, whether per token, per message, or per resolution.29,27

Demand changes as well. Consumers deploy their own AI agents to place calls, which raises arrival rates in bursts that historical forecasts never recorded.12

Supporting Evidence

A forecast dated June 10, 2025, drawn from a March 2025 poll of 163 service leaders, predicts that by 2027 half of the organizations that expected to cut their service workforce significantly will abandon those plans, and reports that 95 percent of leaders plan to retain human agents to define the role of AI.10 The December 2, 2025 survey adds that 55 percent of leaders held staffing stable while handling higher volumes and that 42 percent were hiring AI specific roles such as conversational AI designers and automation analysts.11

A 2026 prediction set expects one in four brands to see a 10 percent increase in successful simple self service interactions by the end of 2026, expects daily agent workloads to fall by an average of one hour, expects at least three major brands to see single day call volumes 100 times normal on six occasions, and expects 30 percent of enterprises to build parallel AI functions that mirror human service roles.12 The Bureau of Labor Statistics projects employment of customer service representatives to fall 5 percent from 2025 to 2035, a loss of about 141,800 jobs from a base of 2,666,000, with about 289,500 openings a year from turnover.1

Customer attitudes shape the ceiling. In a survey of more than 11,000 consumers and leaders across 22 countries, 81 percent of consumers want a representative to resume where the previous interaction ended, 95 percent expect clear explanations of AI decisions, and 80 percent of leaders expect transparency to become required for customer facing AI.13 A market study summary reports that customers still resist AI self service, even for simple transactional issues, and that employees spend too much time on routine issues and resist AI tools.15

Worked Example

Start from the 100 call half hour with a roster of 30 agents at $21.53 an hour, which costs $322.95 in wages and $3.23 per call. Suppose autonomous agents resolve 80 of the 100 calls, as the forecast for 2029 envisions. The 20 residual calls are the complex ones, and suppose their handle time rises to 450 seconds. Erlang C requires 8 agents on the phones for that load and, at 30 percent shrinkage, a roster of 12, which costs $129.18 per half hour, or $6.46 per residual call. Wages fall 60 percent, and the cost of each human handled call doubles from $3.23 to $6.46 because a smaller team gives up the pooling benefit and the remaining calls run longer.

If the residual handle time stays at 300 seconds, the same 20 calls need 6 agents on the phones and a roster of 9, which costs $96.89 per half hour and cuts wages 70 percent. The two scenarios, both mine, show why a forecast of 80 percent containment does not translate into an 80 percent labor saving, and the AI fee for the 80 automated calls sits on top of either figure.18,19

Contested Areas

The forecasts and surveys point in different directions on staffing. The prediction of 80 percent autonomous resolution by 2029 coexists with a survey in which 55 percent of leaders report stable staffing and a forecast that half of the organizations planning large cuts will abandon them.9,11,10 The forecasts are projections, and the survey measures what leaders had done by October 2025, so they answer different questions. The prediction of 80 percent also concerns common issues, and the share of all contacts that qualify as common varies by industry.

Vendor figures need the same discipline. One platform page claims a 20 percent reduction in handling time at a customer using its assistant, with no methodology stated, and an economic study that the vendor commissioned, cited on that page, uses a composite organization.31 A separate operations guide asserts that AI handles 60 to 70 percent of routine tasks without a source, and that figure is not used in this report.33

Practical Significance

Executives should model AI as a change to three terms, the number of contacts reaching humans, the handle time of the contacts that remain, and the pooling effect on occupancy, and should reject any business case that applies the containment rate directly to the wage bill. Founders building AI first service products should expect enterprise buyers to ask for resolution rates by issue type, escalation behavior, and the AI fee per resolution. Investors can use the 20 percent and 55 percent survey results as a reality check on any plan that assumes headcount reductions beyond the market’s reported behavior.

Which Laws and Standards Govern Call Center Operations?

Established Position

Four bodies of law and standards govern call center operations: the Telephone Consumer Protection Act, the Telemarketing Sales Rule, California’s Invasion of Privacy Act, and the PCI DSS payment card standard.2,3,8,36 The Federal Communications Commission ruled on February 8, 2024 that AI generated voices are artificial voices under the Telephone Consumer Protection Act, in a unanimous Declaratory Ruling, FCC 24 17, in CG Docket No. 23 362.4,5

Mechanism

Section 227 of Title 47 prohibits calls made with an artificial or prerecorded voice without the prior express consent of the called party, and the same section lets a person sue for the greater of actual loss or $500 per violation, with the court allowed to increase the award up to three times for willful or knowing violations.2 The 2024 ruling requires prior express consent for AI voice calls, identification of the caller, and opt out methods on telemarketing messages.5

The Telemarketing Sales Rule treats a call as abandoned when it is not connected to a sales representative within two seconds of the called person’s completed greeting, and the safe harbor caps abandoned calls at 3 percent of calls answered by a person, measured over a 30 day period for each campaign.3 Recording law runs through consent. California Penal Code § 632 protects confidential communications, § 631 reaches wiretapping and interception, and § 637.2 sets $5,000 per violation in civil damages.8,7 PCI DSS version 4.0.1 requires that sensitive authentication data never be stored after authorization, and a recording that holds the card verification code is a control failure.36

Supporting Evidence

On January 24, 2025, a unanimous Eleventh Circuit panel held that the FCC exceeded its statutory authority, vacated the one to one consent requirement from its December 2023 order, and rejected the requirement that consent be logically and topically related to the interaction that prompted it.6 On August 11, 2025, the United States District Court for the Northern District of California denied a motion to dismiss claims under California Penal Code §§ 631 and 632 against an AI customer service provider that serves a restaurant chain.7 The PCI DSS future dated requirements took effect on March 31, 2025, the same date version 4.0.1 became the live version.36

The Telemarketing Sales Rule amendments of 2008 also require prerecorded calls to ring at least 15 seconds or four rings, to begin the message within two seconds of the greeting, and to offer an automated opt out, with prior express written agreement required from September 1, 2009.3

Worked Example

The exposure arithmetic grows in proportion to call volume. If a court found that 1,000 calls used an AI generated voice without consent, damages at $500 per violation reach $500,000, and a willful finding raises the ceiling to $1,500,000.2 If a plaintiff class proved 1,000 recorded calls violated § 632, civil damages at $5,000 per violation reach $5,000,000.8 Under the abandonment safe harbor, a campaign that reaches 10,000 live answers in 30 days can abandon no more than 300 calls.3 Those figures come from the statutes and rules by simple multiplication, and they show why compliance is a design input and not a clean up task.

Contested Areas

Two contested points affect design. The first is whether an AI analytics vendor counts as a third party under California’s wiretap provisions. The court in the Northern District of California applied the capability approach, under which a vendor that can use communications for its own independent purposes is a third party, and found the allegations that the vendor used caller data to improve its own services enough to proceed.7 The second is the scope of the FCC’s consent rulemaking power, which the Eleventh Circuit narrowed on January 24, 2025, leaving the statutory standard of prior express consent as the governing text.6 A vendor claim that DTMF masking removes up to 96 percent of operations from PCI scope has no independent verification, and the PCI Security Standards Council documents govern the requirement language.36

Practical Significance

General counsel should map every call path against four questions: who consented, to what kind of voice, with what recording notice, and with what card data exposure. Executives buying AI analytics or voice agents should require contractual limits on a vendor’s reuse of caller data, because the capability approach turns vendor reuse into the fact that decides third party status. Founders selling into regulated enterprises should expect procurement to ask for recording consent flows, opt out handling, and PCI scope documentation before the first pilot. Investors can price the risk with the multiplication above: volume times a per violation figure of $500 to $5,000 produces an exposure number that any diligence memo should state.

What Do Real Cases Show About Call Center Legal Risk?

What Did the Eleventh Circuit Decide About One to One Consent?

On January 24, 2025, a unanimous three judge panel of the Eleventh Circuit vacated the one to one consent requirement that the Federal Communications Commission adopted in its December 2023 order under the Telephone Consumer Protection Act.6 The Insurance Marketing Coalition brought the challenge. The December 2023 order added two new consent restrictions: the one to one requirement and a requirement that consent be logically and topically related to the interaction that prompted it.6

The court held that the FCC exceeded its statutory authority. Its reasoning rested on the ordinary meaning of the statutory phrase prior express consent: under the court’s reading, a consumer needs only to state in unmistakable terms, before a robocall, a willingness to receive it.6 The court vacated the one to one portion, remanded the matter to the FCC, and rejected the logically and topically related element. The FCC had earlier postponed the rule’s effective date, and the stated aim of that postponement was to avoid exposing good faith callers to private suits seeking statutory damages while the rule was under judicial review.6

The case shows three things to a call center operator. First, an agency’s reading of consent can fall when a court reads the statutory words differently, so a consent program built to the strictest agency rule can be stricter than the law requires, and a program built to the narrowest court reading can be exposed when the rule changes again. Second, the statute continues to apply, and the statutory damages of $500 per violation, with a court allowed to triple the award for willful or knowing violations, remained available to any plaintiff throughout the dispute.2 Third, the FCC’s February 2024 treatment of AI generated voices as artificial voices stands apart from the vacated rule, so an outbound program using synthetic voices still needs prior express consent.4,5

For operations, the lesson concerns evidence. A center that records the consent language, the date, the channel, and the scope of each consent holds the proof that either standard demands, and the cost of keeping that record is small next to $500 per call across a campaign of tens of thousands of calls. The timeline also carries a planning lesson. The order was adopted in December 2023, the FCC postponed its effective date, and the court acted on January 24, 2025, so an operator who rebuilt consent flows for the new rule spent a year on a requirement that a court then removed. A compliance calendar for contact centers should therefore separate rules in force from rules proposed, adopted but postponed, or under review, and should assign each a named owner and a review date.

Can AI Call Analytics Violate California’s Wiretap Law?

On August 11, 2025, the United States District Court for the Northern District of California denied the motion to dismiss filed by ConverseNow Technologies, Inc., an AI customer service provider, in a putative class action alleging violations of California Penal Code §§ 631 and 632.7 Section 631 addresses wiretapping and the willful interception of communications in transit. Section 632 addresses the intentional recording of a confidential communication without the consent of all parties.7 The provider’s customer in the case was the restaurant chain Domino’s.7

The court applied the capability approach, under which a defendant counts as a third party to the conversation if it has the capability to use the communication for an independent purpose. It found that the complaint plausibly alleged that ConverseNow was more than an extension of its restaurant customer, because the complaint alleged that the company uses caller data to enhance its own services.7 That finding permitted both the § 631 and the § 632 claims to proceed.7 The ruling resolved a pleading stage question and decided nothing about the merits, which means discovery into what the vendor actually does with caller data comes next in cases of this kind.

The consequence for a contact center lies in the arithmetic of § 637.2, which sets civil damages at $5,000 per violation.8 A vendor that transcribes, scores, or summarizes calls can fall within this reasoning if the vendor’s contract lets it keep or reuse the data. A center running a transcription tool on 10,000 calls in a month faces a theoretical ceiling of $50,000,000 at the statutory figure if every call were a violation, and the number explains why plaintiffs file these cases as class actions.8 The ceiling is arithmetic, and the outcome of any case depends on proof of confidentiality, consent, and the vendor’s actual use of the data.

The practical response is contractual and operational. The vendor agreement should confine the vendor to processing on the customer’s behalf, bar the use of caller data to train the vendor’s own models, and require deletion on a stated schedule. The call opening should disclose recording and AI processing before the conversation begins, and the disclosure should be logged. I have drafted and negotiated data use limits of this kind in technology agreements for three decades, and the limit on a vendor’s reuse of data is the term vendors resist most and the term that decides whether the vendor is a service provider or a third party.

What Is the Strongest Case Against Automating Call Centers?

The strongest case against rapid automation rests on adoption data, customer behavior, and arithmetic, and the evidence runs more heavily on its side than the forecasts suggest. Its best form begins with what leaders have done. A survey of 321 service leaders conducted in October 2025 found that 20 percent had reduced agent staffing because of AI, that 55 percent held staffing stable while handling higher volumes, and that 42 percent were hiring AI specific roles.11 A June 2025 forecast from the same research practice predicts that half of the organizations that planned large workforce reductions will abandon them by 2027, and 95 percent of surveyed leaders plan to keep human agents.10

Customer behavior supplies the second leg. A 2025 market study summary reports that customers still prefer agents even for simple transactional issues, and that employees resist AI tools.15 A 2024 industry survey found that 73 percent of customers try self service while 14 percent of issues are fully resolved through it alone.14 In a 2026 consumer survey, 74 percent of consumers reported frustration at repeating information, which is the failure an AI front end produces whenever it hands off without context.13

Arithmetic supplies the third leg. The wage example in this report shows 80 percent containment producing a 60 percent wage reduction under one set of assumptions and 70 percent under another, because residual calls are longer and a smaller team loses the pooling benefit.19 The Bureau of Labor Statistics projects a 5 percent decline in employment over ten years, from 2,666,000 jobs in 2025, a pace consistent with gradual substitution.1

The opposing position holds that the forecast of 80 percent autonomous resolution by 2029, with a 30 percent operational cost reduction, describes where the economics will land, and that organizations must treat automation as the dominant strategy for all service teams.9 Its best form relies on cost: an AI contact billed per message or per resolution can cost far less than the $3.23 in wages the median agent consumes per call in the 100 call example, and any competitor that adopts the cheaper meter will reprice its market.27 It also points to demand side pressure, since a forecast of single day call volumes 100 times normal at three major brands cannot be met with human staffing at any plausible cost.12

The evidence supports a narrower reading than either side states. Automation of routine contacts has reduced staffing at one in five of the surveyed organizations, and the larger reductions in the forecasts have yet to appear in surveys that measure behavior. The projection of a 5 percent decline in federal employment data and the 289,500 annual openings from turnover supply the mechanism for reduction: a center that stops backfilling departures while volumes rise reaches a smaller floor without announcing a layoff.1 The forecast of 80 percent and the survey of 20 percent are consistent as a trajectory, and the variable is the speed at which the second number approaches the first.

A second contested position concerns recording and analytics. Operations leaders argue that full coverage analytics improves quality and compliance, and plaintiffs argue that the same tools create wiretap exposure at $5,000 per violation. Both positions rest on documented facts: the analytics vendor case studies report a 10 percent cost reduction and halved QA staff at individual customers, and the Northern District of California let the wiretap claims proceed.35,7 The facts do not conflict. They describe a technology whose value and whose exposure both depend on contract terms and call opening disclosures. Neither position requires a center to abandon analytics. Each position requires a center to read its vendor contract, to confirm who may keep the transcripts, and to decide before deployment which contract terms it will demand.

What Should Founders, Investors, Executives, Board Members, and General Counsel Do Now?

Founders building or buying a service function should decide this year whether the first agents are human or automated, and should price both on the same workload. The 100 call example gives the template: a $3.23 wage cost per call at the median wage, a $0.19 voice platform fee, a software seat bill of $90,000 to $288,000 a year for 100 users, and attrition of $380,000 to $1,748,000 a year at 38 percent. A founder who signs a seat license with a 12 month commitment before the support volume is known owns a minimum payment from the first day. A 20 seat team at the $155 tier costs $37,200 a year at list and $22,320 to $31,620 after a 15 to 40 percent negotiated discount, and the commitment term decides whether those seats can shrink when the first AI agent goes live.1,27,29,24,25

Investors diligencing a contact center business or a company with a large service organization should request five numbers before the term sheet: handle time, service level target and result, shrinkage, occupancy, and first call resolution. From those numbers an investor can rebuild the roster with Erlang C and compare it to payroll in an afternoon. The repeat call share of 23 percent of operating budget and the 38 percent attrition rate each convert directly to a dollar figure for a 100 agent center, $197,800 on an $860,000 budget and up to $1,748,000 in replacement costs, and both belong in the valuation model. A forecast error of 1 percent costs about $400,000 a year in the workforce vendor’s own example, which gives an investor a reason to ask how the target forecasts and how often it reforecasts within the day.17,24,23

Executives and senior operators should staff from the three inputs, publish the service level target by interval, and track average handle time next to first call resolution, since the benchmarking firm warns that a shorter handle time can lower both resolution and satisfaction.16 Platform decisions belong on a calendar. Porting numbers from PRI to SIP takes 20 to 30 days per order, a contract signed in the fourth quarter should not assume a January cutover, and a seat renewal negotiated against 15 to 40 percent below list is a negotiation worth running every renewal. On 100 seats at the $240 tier, that range equals $43,200 to $115,200 a year.21,25,29 Executives should also set the review cycle: forecast monthly, reforecast within the day, and compare actual voice occupancy to the 80 to 90 percent range that one practitioner reference recommends and the 70 to 75 percent range that a benchmarking firm calls healthy, because a center that runs above 80 percent invites turnover that costs $10,000 to $46,000 per replacement.19,16,24

Board members should ask management for the AI business case as three changes, contacts reaching humans, handle time of the residual calls, and pooling effect, and should treat a plan that applies a containment percentage to the wage bill as incomplete. The survey result that 20 percent of leaders have cut staffing and 55 percent have held it stable gives the board a benchmark for management’s assumptions, and the board can set the timing by asking for the model before the next budget cycle approves seat counts, since a license term of 12 months fixes the cost for the full year.11

General counsel should complete a consent and recording audit before any AI voice or analytics deployment. The audit covers four items: the consent evidence for every outbound call that uses a prerecorded or synthetic voice, the recording disclosure at the call opening, the vendor’s data reuse rights, and the PCI scope of every system that hears or records a card number. The exposure runs from $500 per call under 47 U.S.C. § 227 to $5,000 per recorded call under California Penal Code § 637.2, and a vendor contract amendment limiting data reuse costs a fraction of a single class action defense.2,8,36

Policy professionals should watch two matters. The first is the FCC’s response to the Eleventh Circuit’s vacatur, which determines whether consent rules return in a narrower form or the statutory standard stands alone.6 The second is the treatment of AI generated voices and AI analytics under state wiretap laws, which the Northern District of California ruling placed before more courts.7,4 Both questions bear on the cost of operating a contact center across state lines, and a national standard would replace a state by state compliance map priced in per violation damages.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

The Shippingport Light Water Breeder Reactor, which ran from 1977 to 1982, demonstrated that a thorium and uranium 233 core can breed fuel, ending with 1.39 percent more fissile material.

Core 3 at the Shippingport Atomic Power Station in Pennsylvania produced about 2.1 billion kilowatt hours of gross electricity over roughly 29,000 effective full power hours, and assay of 524 discharged fuel rods confirmed the gain in fissile inventory. Examiners found no fuel element defects. Power operation stopped on October 1, 1982, and the Department of Energy completed dismantlement in 1989 for $91.3 million.

Shippingport was the first full scale American reactor built only to make civilian electricity. Its third core used pellets of thorium oxide and uranium 233 oxide in Zircaloy 4 tubes, and it controlled reactivity by moving fuel in place of neutron absorbing control rods. That choice removed a source of parasitic neutron loss, and the loss was the margin that made breeding possible in light water.

For energy investors, utilities, and policy professionals, the record carries a plain commercial message. The physics worked in 1982. The fuel cycle around it needs recycled uranium 233, shielded fabrication, and reprocessing that has never operated at industrial scale, and those requirements kept the concept off the market. The federal outlay for the Shippingport pressurized water and breeder programs was estimated at $622.2 million in 1975, and the breeder core never entered commercial service.

Shippingport was exempt from licensing by the Nuclear Regulatory Commission, so no licensing record exists for a thorium fed light water reactor. A developer today would proceed under 10 CFR Part 50, Part 52, or the optional Part 53 that took effect on April 29, 2026, and uranium 233 is controlled as special nuclear material of the highest security category. The sections below trace the history, the engineering, the cost, the back end of the fuel cycle, three documented precedents, the strongest case against thorium breeding, and the decisions that follow for each audience.

Background and History

How Did Shippingport Become the First Full Scale Civilian Reactor in America?

Shippingport became the first full scale civilian power reactor in America because a cancelled aircraft carrier reactor was redirected to Pennsylvania in 1953. In April 1953 the Atomic Energy Commission proposed canceling the carrier propulsion reactor under development at the Westinghouse Bettis laboratory and moving its funds to a civilian power plant. On December 8, 1953 the President of the United States proposed the Atoms for Peace program before the United Nations General Assembly, and the Commission sought a utility partner to build the plant.

Duquesne Light Company won the partnership. Its bid supplied the land, the turbine generator, and $5.0 million toward light water reactor research, and the contract was signed in March 1954. Ground was broken on September 6, 1954 before about 1,400 guests, with the first scoop of earth activated by remote control from Denver. The Division of Naval Reactors under Rear Admiral Hyman Rickover held primary responsibility for the reactor, and the site sat on the Ohio River about 25 miles from Pittsburgh.

The reactor was a pressurized water design with a primary system near 2,000 pounds per square inch, four steam generators, and a single turbine. Rickover chose uranium dioxide fuel in Zircaloy cans over a uranium molybdenum alloy, and the choice became the standard for the civilian industry. The plant reached criticality at 4:30 a.m. on December 2, 1957, delivered first power to the grid on December 18, and reached its full 68,000 kilowatts on December 23, 1957.

What Happened to Core 1 and Core 2?

Core 1 and Core 2 proved the seed and blanket concept at commercial scale. Core 1 was installed in October 1957, weighed 58 tons, and carried 165 pounds of 93 percent enriched uranium as seed inside 14.2 tons of natural uranium as blanket, and it produced close to 2 billion kilowatt hours before replacement in 1964. Core 2 operated from February 3, 1965 to February 4, 1974, raised electrical capacity from 60,000 to 100,000 kilowatts, and produced almost 3.5 billion kilowatt hours. It used 20 seed assemblies with hafnium control rods and 77 blanket assemblies.

A turbine generator mechanical failure took the plant offline on February 4, 1974. The Energy Reorganization Act of 1974 split the Atomic Energy Commission in January 1975 into the Nuclear Regulatory Commission and the Energy Research and Development Administration, which held Naval Reactors, and the Department of Energy absorbed that administration in 1977.

Why Did the Atomic Energy Commission Build a Breeder Core for Shippingport?

The Commission approved the breeder research program in December 1965 as a path to stretch the uranium supply, since a reactor that made more fissile fuel than it consumed would need less mined ore. A 1961 advisory committee had already concluded that future plants needed to be far larger than 60 megawatts, and the Naval Reactors team studied designs of up to 750 megawatts. The Water Cooled Breeder Program ran from 1965 to 1987.

Fabrication of the breeder core finished in 1976 to 1977. The core reached initial criticality on August 26, 1977, and the President directed it to full power by closed circuit television on December 2, 1977. The 1975 audit had scheduled installation for early 1976, so the core arrived about a year late. It then ran for almost five years.

When uranium proved plentiful in the 1980s, the case for a breeder weakened. The plant ended power operation on October 1, 1982, and a 1980 landmark designation by the American Society of Mechanical Engineers recorded the facility’s place in the history of the profession. The core was removed, examined, and sent to Idaho, and the plant went to decommissioning.

Two statutes frame the date. The Atomic Energy Act of 1954, signed on August 30, 1954, opened civilian nuclear power to private participation days before ground was broken at Shippingport. The Price Anderson Act, enacted on September 2, 1957, set the liability regime for private reactors three months before the plant reached criticality, and Shippingport’s federal ownership kept its nuclear portion outside that licensing system.

Detailed Findings

What Does the Shippingport Record Establish in Brief?

The record establishes four facts. The core bred fuel: 524 rods assayed at 1.39 percent more fissile material than at loading. The core performed: about 29,000 effective full power hours against an 18,000 hour design life, a 76.2 percent availability, and no fuel failure. The program cost the federal government an estimated $596.9 million through fiscal year 1980, with Duquesne Light adding $25.3 million. The back end left 47 liners of discharged fuel in Idaho and a decommissioned site that cost $91.3 million to clear.

Why Does the Core Matter to Current Developers?

It matters because no other thorium cycle reactor has shown thermal breeding at this power level. The International Atomic Energy Agency calls the Shippingport reactor the major thorium based reactor in the world to demonstrate thermal breeding, with a breeding ratio above 1. Every current thorium fuel proposal inherits both the physics margin and the fuel cycle burden that Shippingport documented.

What Remains Uncertain?

The margin is 1.3 to 1.4 percent, and a commercial plant would need to sustain it across repeated recycle with uranium 233 fabrication in shielded cells. No reprocessing plant for thorium has reached industrial maturity, and the Nuclear Regulatory Commission has no licensing precedent for the fuel. Those two facts decide whether anyone builds a successor.

Legal and Regulatory Implications

Was Shippingport Licensed by the Nuclear Regulatory Commission?

No. A June 1979 Commission letter states that Shippingport was owned by the Department of Energy, operated by Duquesne Light Company for the Department, and exempt from Commission licensing under the Atomic Energy Act of 1954 and the Energy Reorganization Act of 1974. The breeder core therefore never went through a safety evaluation under 10 CFR Part 50, and no licensing precedent exists for thorium and uranium 233 fuel in a light water reactor.

How Would a Thorium Fueled Reactor Be Licensed Today?

A developer would apply under 10 CFR Part 50, Part 52, or the new Part 53. The Commission published the Part 53 final rule at 91 Fed. Reg. 15696 on March 30, 2026, effective April 29, 2026, as an optional risk informed, technology inclusive licensing pathway required by the Nuclear Energy Innovation and Modernization Act of 2019. The Commission estimated net averted costs of about $152 million at a 7 percent discount rate and $203 million at 3 percent. The 2014 review prepared for the Commission lists the technical basis a thorium applicant must build: protactinium 233 decay and reactivity after shutdown, gamma shielding for uranium 232 daughters, reprocessing by the THOREX process, and criticality control at fresh fuel enrichments of 10 to 20 percent for once through designs.

How Is Uranium 233 Controlled as a Nuclear Material?

Uranium 233 is special nuclear material under the Atomic Energy Act, and the Commission and the International Atomic Energy Agency treat it with plutonium and highly enriched uranium as the highest security category. A commercial holder needs a license under 10 CFR Part 70, material control and accounting under Part 74, physical protection under Part 73, and an export license under Part 110 for any foreign transfer. The uranium 232 gamma field gives some intrinsic barrier, which supports the case for thorium fuels, and the same field is the reason fabrication needs shielded cells.

Who Bears Liability for a New Reactor and Its Waste?

The Price Anderson Act of 1957, codified at 42 U.S.C. § 2210, supplies the liability regime for every licensed commercial reactor, and the Nuclear Waste Policy Act of 1982 assigns the Department of Energy the duty to dispose of spent fuel. Shippingport’s own fuel remains in Idaho under Department of Energy custody, and the Shippingport decommissioning proceeded under Department of Energy authority outside the license termination rule at 10 CFR 50.82.

Detailed Findings, Expanded

The five subtopics below take the Shippingport record apart in turn: the core design, the breeding result, the operating record, the economics, and the back end of the fuel cycle. Each one states the established position, explains the mechanism, lays out the evidence, works an example, names the contested ground, and states what it means for the reader.

How Did the Core Design Control a Breeder Without Poison?

What Is the Established Position on the Core Design?

The established position is that the Shippingport breeder core controlled reactivity by moving fuel and carried no control poison in the core. The 1984 core performance analysis describes 12 movable seed assemblies, 12 stationary annular blanket assemblies that pair with them as seed and blanket modules, 15 reflector blanket modules, and 15 stainless steel filler units, seated in a reactor vessel with a 109 inch inner diameter. The core held about 17,300 fuel rods and nearly 3 million pellets, of which about 1.6 million were binary pellets of thorium oxide and uranium 233 oxide and about 1.3 million were thorium oxide alone.

How Does Moving the Seed Control the Chain Reaction?

The mechanism is geometric. When the seed assemblies sit low relative to the blankets, neutrons leak into the thorium and are captured there, and the chain reaction weakens. When the seeds rise to align with the blankets, the fissile material couples to the thorium, and reactivity increases. Total seed travel was 84 inches. At shutdown the seeds sat 60 inches below the blankets, and at maximum power they rose to 24 inches above the aligned position.

A conventional pressurized water reactor adds boron to the coolant or inserts rods of hafnium or silver alloy to absorb neutrons. Each absorbed neutron is a neutron that does not convert thorium 232 into uranium 233. Removing the poison kept those neutrons in the breeding balance, and that choice is the reason the margin above 1.0 existed.

What Evidence Supports This Account of the Design?

The configuration data come from the 1984 analysis and the 1987 program summary. Seed assemblies carried 4.3 and 5.2 weight percent fissile content, blanket assemblies carried 1.2 to 2.7 weight percent, and nine of the twelve blankets had an outer power flattening region with higher uranium 233 content. Thoria reflector segments of about 10 inches sat at each end of every rod. The cladding was Zircaloy 4 with hafnium below 50 parts per million, since hafnium absorbs neutrons.

Core 1 and Core 2 used the more familiar approach. Core 1 relied on 32 hafnium control rods, and Core 2 used 20 seed assemblies with hafnium control rods among 77 blanket assemblies. The breeder core departed from both.

What Does a First Hand Observation Show About the Design?

In my own nuclear training as a submarine officer, the first lesson in reactor behavior was the negative temperature coefficient, the property by which a hotter coolant slows the chain reaction without any action by the operator. Shippingport carried that property from its first day, and the Naval Reactors tradition of conservative design that the plant embodies shaped the way I was taught to read a reactor plant. The breeder core adds a mechanical subtlety that a trained operator notices at once. The control element is a large fuel assembly that moves in a vessel full of high pressure water, so every inch of travel changes both reactivity and the power distribution across the core.

Axial motion of fuel on that scale demands drive mechanisms with a reliability record that a commercial plant would have to demonstrate across a fleet. Shippingport demonstrated it on one core for five years.

Where Is the Core Design Contested?

The 2014 review prepared for the Nuclear Regulatory Commission lists the design’s shortcomings: a need for lower power density, a higher initial seed enrichment, a movable seed region, harder reprocessing than uranium and plutonium, and added shielding in fabrication. Engineers who favor the design point to the clean result at Shippingport. Engineers who doubt it point to the fact that the movable seed region adds mechanisms and complexity that a larger reactor multiplies. Both positions rest on the same data, and neither has a second reactor to test against.

What Does the Core Design Mean for Developers and Investors?

It means that a breeder in light water is a different machine from a conventional pressurized water reactor, and it cannot be dropped into an existing plant design without changes to the vessel internals, drive mechanisms, and core instrumentation. A developer proposing a thorium seed and blanket design should expect the Commission to review the movable fuel mechanism as a first of a kind safety system, and an investor should price that review as a schedule and cost item.

Did the Core Breed Fuel and How Was That Proved?

What Is the Established Position on the Breeding Result?

The established position is that the core bred fuel by a margin of about 1.4 percent, confirmed by physical assay of the discharged rods in 1987. The program summary reports that nondestructive assay of 524 fuel rods found 1.39 percent more fissile material at the end of core life than at the start. The 1984 analysis, written before disassembly, calculated an end of life fissile inventory ratio of 1.013, and the 2014 review prepared for the Nuclear Regulatory Commission reports an overall breeding ratio of 1.0139. The calculation and the measurement agree.

How Does Thorium Breeding Work in a Thermal Reactor?

Thorium 232 captures a thermal neutron, becomes thorium 233, and decays through protactinium 233 into uranium 233, which is fissile. Two properties make the cycle workable in a thermal spectrum. Uranium 233 releases more than two neutrons per neutron absorbed across a wide range of thermal energies, a property that uranium 235 and plutonium 239 lack. Thorium 232 also captures thermal neutrons at about 7.4 barns, against 2.7 barns for uranium 238, so each thorium atom is nearly three times as likely to capture a neutron.

A breeder needs the fissile atoms it makes to outnumber the fissile atoms it consumes. Every neutron lost to coolant, structure, poison, or leakage reduces the count. The Shippingport design attacked each loss: no poison, thoria reflectors at the ends, low hafnium cladding, and a lower power density than a commercial reactor.

What Evidence Supports the Breeding Claim?

The evidence has three layers. First, the 1984 analysis calculated that the fissile inventory ratio peaked at 1.015 near 22,300 effective full power hours and finished at 1.013, with initial fissile loading of about 501 kilograms and an end of life loading of about 508 kilograms. Second, nondestructive assay covered 524 rods, and the later inspection record states that about 1,000 rods were sampled from the 12 modules. Third, 17 rods were dissolved for isotopic analysis at Argonne National Laboratory. The destructive analysis finished five years after shutdown and more than ten years after fuel fabrication, and the results were delivered to the Department of Energy in September 1987.

What Does a Worked Example Show About the Margin?

The reported figures allow a direct check. An initial fissile loading of about 501 kilograms multiplied by 1.39 percent gives a gain of about 7.0 kilograms. The calculated end of life loading of about 508 kilograms implies the same 7 kilogram gain. Over roughly 29,000 effective full power hours, the core delivered about 2.1 billion kilowatt hours gross while its fissile inventory grew by about 7 kilograms.

The same arithmetic shows how thin the margin is. A gain of 7 kilograms on 501 kilograms means that a commercial breeder with recycle would need every reprocessing and fabrication step to lose far less than 1.4 percent of the fissile material, cycle after cycle. A process loss of 2 percent in one recycle step erases the breeding gain.

Where Is the Breeding Result Contested?

Supporters read the result as proof that thermal breeding works, and the 2005 International Atomic Energy Agency review agrees. A 1998 peer reviewed assessment reads the same result in economic terms and finds that light water reactors under reference conditions favored the uranium cycle, and that a break even commercial breeder required very high costs for uranium 233 inventory and recycle. The physics and the economics are separate questions, and the Shippingport record answers the first.

Advocates have also argued that sponsors stopped the experiment before its full potential was shown, because the post operation analysis was at risk of losing funding. That claim rests on an advocate’s account and the record shows the analysis was completed.

What Does the Breeding Result Mean for Readers?

It means that any investor presentation claiming a thermal thorium breeder has a physical precedent, and any presentation claiming commercial breeding with recycle has no precedent. The margin of 7 kilograms in 501 is the figure to hold against such claims, together with the observation that no recycle loop closed at Shippingport.

The practical test for any new claim is whether the measured gain survives the same assay. Shippingport’s gain was checked against rods pulled from the discharged core, and no vendor projection can match that until a new core is built, run, and examined.

What Did the Operating Record Show About Thorium Fuel Performance?

What Is the Established Position on the Operating Record?

The established position is that the core ran well past its design life with no fuel failure. Design life was 18,000 effective full power hours. The core reached 18,298 in its first three years and finished at 28,730 hours in the 1984 analysis and about 29,000 in the program summary, which is about 160 percent of design life. Capacity factor was 64.0 percent and availability was 76.2 percent. Gross generation was about 2.1 billion kilowatt hours and net generation about 1.7 billion.

How Did Operators Manage the Core over Five Years?

Operators managed the core by lowering conditions as it aged. Thermal power was 236.6 megawatts, which gave about 72 megawatts electric gross and 60 megawatts net. After 18,298 hours the coolant temperature was reduced from 531 to 521 degrees Fahrenheit and the pressure from 1,985 to 1,585 pounds per square inch gauge, and operation ran mostly at 80 percent power with a coastdown to 57 percent before final shutdown.

Lowering temperature and pressure reduces the thermal and mechanical duty on the cladding of a core operated far beyond its original design life.

What Evidence Supports the Fuel Performance Claim?

Radiochemical sampling of the coolant throughout operation showed no indication of fuel element failure, and the post operation examination reported no evidence of any fuel element defects. The International Atomic Energy Agency records a maximum burnup of 60,000 megawatt days per metric ton and no fuel failures. The 1984 analysis gives peak local depletion of about 54,000 megawatt days per ton in the seed region and 25,000 in the blanket, with an average of about 14,600.

The core ran more reactive than calculated. The reactivity bias ranged from about 0.15 to 0.54 percent at full power, and all measured values fell inside the plus or minus 1 percent allowance in the design.

What Does a Worked Example Show About Generation?

The reported figures reconcile. At 28,730 effective full power hours and about 72 megawatts electric gross, the core generated 28,730 multiplied by 72,000 kilowatts, or about 2.07 billion kilowatt hours, which matches the reported 2.1 billion. The thermal efficiency implied by 72 megawatts electric from 236.6 megawatts thermal is about 30.4 percent.

The capacity factor of 64.0 percent reflects scheduled outages, the coastdown, and the research testing the plant performed. A commercial thorium plant would operate to a different schedule.

Where Is the Operating Record Contested?

The record is contested on extrapolation. The core produced 60 megawatts electric net. The commercial breeder designs developed under the program were for plants of 900 to 1,000 megawatts electric net, and no such core was built. Engineers who trust the extrapolation point to the conservative design practice. Engineers who doubt it point to the lower power density of the Shippingport core, which a commercial plant would want to raise for economics, and which raises the cladding and thermal duty.

What Does the Operating Record Mean for Readers?

It means that thorium oxide fuel in zirconium alloy cladding has a five year, no failure record in a pressurized water environment, and the Nuclear Regulatory Commission can point to that record when a developer cites precedent. The record does not substitute for the qualification data a modern application requires at current burnups, and a developer should budget for irradiation testing, as Clean Core did in Idaho over nearly two years.

A comparison with the 2026 Idaho campaign puts the record in modern terms. Clean Core’s target of 60 gigawatt days per metric ton equals 60,000 megawatt days per metric ton, the same figure the International Atomic Energy Agency records as the maximum burnup of the Shippingport fuel. The two fuels differ in composition, reactor type, and purpose, so the match is a reference point. It shows that a burnup now sought for thorium blends in heavy water reactors was reached with zirconium clad thoria fuel in light water more than four decades earlier.

A regulator would ask for the data that connect that burnup to fission gas release, cladding corrosion, and pellet cracking under transient conditions, and the Shippingport examination reported minimal cladding deformities and pellet cracks for the steady operation the core saw.

What Did Shippingport Teach About the Economics of Thorium Breeding?

What Is the Established Position on Shippingport Economics?

The established position is that Shippingport was a federally funded research plant that generated electricity at a cost above the utility’s system average. The 1975 audit estimated total government cost through fiscal year 1980 at $596.9 million, and $622.2 million with Duquesne Light’s $25.3 million. In 1973 the plant supplied 1.4 percent of Duquesne Light’s electricity at about 7 mills per kilowatt hour against about 5.8 mills systemwide, and the auditors found no significant economic advantage to the utility.

Why Does a Thorium Breeder Cost More Than a Uranium Reactor?

A thorium breeder costs more because its fuel cycle adds steps that a once through uranium cycle does not have. Uranium 233 must be bred in a reactor, separated from thorium and fission products by the THOREX process, a modified PUREX chemistry in which dissolving thorium oxide needs hydrofluoric acid and long hot dissolution, and refabricated behind shielding because uranium 232 daughters emit hard gamma rays, including a 2.6 megavolt line from thallium 208.

Each step carries capital cost, handling cost, and inventory carrying cost, since a breeder with recycle ties up uranium 233 in the plant and in the pipeline. The economics improve as the price of mined uranium rises and worsen as it falls.

What Evidence Supports the Economic Assessment?

The 1975 audit breaks down the federal cost: $59.5 million for nuclear facility construction, $37.6 million for equipment, $39.0 million for operating and maintenance reimbursement through fiscal year 1976, $229.4 million for technical support of the pressurized water reactor, and $225.1 million for technical support of the breeder. The Commission received $20.3 million from steam sales through fiscal year 1974, when the plant had produced about 4.62 billion net kilowatt hours. The rate was 8 mills for the first 1.4675 billion kilowatt hours and 2.30 to 3.07 mills afterward.

The 1998 peer reviewed assessment of the Radkowsky thorium reactor concept, a seed and blanket design descended from Shippingport, found a total fuel cycle cost about 4 percent lower than a pressurized water reactor under reference conditions, a difference small against the uncertainty in the cost parameters.

What Does a Worked Example Show About Cost?

The audit figures allow a rough check on the research burden. Actual government cost through fiscal year 1974 was $466.4 million, and the plant had produced about 4.62 billion net kilowatt hours. Dividing gives about $0.10 per kilowatt hour of federal outlay. That figure includes research and technical support, so it measures the price of the knowledge. The utility paid about 7 mills, or $0.007 per kilowatt hour.

The gap between $0.10 and $0.007 is the share of the program that bought learning. Investors reading a thorium developer’s cost projections should ask how much of the assumed cost reduction depends on learning that has not yet been paid for.

Where Are the Economics Contested?

Advocates of thorium argue that a recycle economy would value the long term energy content of the fuel, and that a modern plant with better cladding, better shielding practice, and automation would close the gap. Skeptics reply that uranium was no longer scarce once the program ended, that the recycle step has never been demonstrated at industrial scale, and that a 1.4 percent breeding margin leaves no room for process loss. The evidence in hand supports the skeptics on cost and the advocates on physics.

What Do the Economics Mean for Readers?

They mean that a thorium breeder with recycle is a research bet, and a thorium blend for once through use in existing reactors is a fuel business. The two carry different capital needs and different regulatory paths. Clean Core’s thorium and high assay low enriched uranium fuel for heavy water reactors takes the second path, and it avoids reprocessing altogether.

The ownership structure also explains the cost record. Duquesne Light spent $20.3 million on the turbine generator and contributed $5.0 million to the nuclear portion, the Commission owned the nuclear portion, and it reimbursed Duquesne Light for operating and maintaining it. The utility’s exposure was limited to the electric half of the plant, and the technical risk of the nuclear half sat with the federal government. A commercial thorium plant would shift that risk to the owner.

What Remains After a Thorium Breeder: Spent Core, Plant, and Uranium 233?

What Is the Established Position on the Back End?

The established position is that the Shippingport back end was completed by the Department of Energy with fuel stored in Idaho and the plant cleared by 1989. The discharged core sits in 47 liners in underground dry wells at building CPP 749 at the Idaho Nuclear Technology and Engineering Center. Decommissioning cost $91.3 million against a $98.3 million estimate, and the reactor vessel was shipped intact by barge to Hanford for burial.

How Was the Spent Core Handled?

Fuel was wet stored at the Expended Core Facility in Idaho, dried, and moved to dry storage. The liners hold 12 seed module units, 12 blanket module units, 15 reflector module units, 7 loose rod units, and 1 scrap rod unit. About 24,000 rods were manufactured and 17,288 were loaded in the core. The uranium in the fuel is 98.23 percent uranium 233, in a thoria matrix, with an eta of about 2.3.

Examination found minimal cladding deformities and pellet cracks, and the storage record shows no indication of damage from shipping, loading, or storage.

What Evidence Supports the Back End Record?

The decommissioning audit records physical work from September 1985 to July 1989, certification as radiologically safe in October 1989, and a final project report in December 1989. Removing the vessel in one piece saved about $7 million. Disposal at Hanford of about 214,000 to 216,000 cubic feet of waste cost about $2.4 million, including the vessel. Shippingport’s own dry wells in Idaho hold the fuel, and the Oak Ridge Uranium 233 Disposition Project, with nearly 1,100 canisters, shows the scale of the wider inventory.

What Does a Worked Example Show About the Wider Inventory?

An advocate’s 2012 account cites a federal stockpile of about 3,400 pounds of uranium 233, from a report by a critic of the material. Converted, 3,400 pounds is about 1,540 kilograms. The Shippingport core carried about 508 kilograms of fissile material at discharge, so the national stockpile is about three times the fissile mass of the Shippingport core. Both quantities carry the highest security category.

Where Is the Back End Contested?

The contest is between disposal and reuse. The critic of the stockpile argues that the material is a security and proliferation burden and should be disposed of. The advocate answers that operating reactors are protected against theft and that the material is an energy resource with a value that the same account puts above $2 billion at 2012 oil prices. The same account cites $473 million as the planned disposal cost, and that figure comes from an advocate and should be read as such. The Department of Energy is processing the Oak Ridge inventory for transfer to other programs or offsite disposal, and it extracts thorium 229 for medical isotope production along the way.

What Does the Back End Mean for Readers?

It means that a developer who breeds uranium 233 owns a long lived liability in the highest security category. The Price Anderson Act covers licensed reactor liability, and the Nuclear Waste Policy Act of 1982 addresses spent fuel and high level waste. A financing that depends on recycle should carry a specific plan for the separated uranium 233, with a named custodian and a dollar figure.

The arithmetic of fabrication adds a point. About 24,000 rods were manufactured and 17,288 were loaded, which leaves about 6,700 manufactured rods outside the core. A recycle economy would have reprocessed the discharged rods and refabricated them behind shielding, and the Shippingport program never took that step. The discharged core holds about 508 kilograms of fissile material, a little more than the 501 kilograms loaded in 1977.

A repository application for the liners would need a criticality analysis for fuel that is 98.23 percent uranium 233 in a thoria matrix. The 2002 characterization report states that it is not intended for design applications and directs users to the original source documents, so any such analysis would return to the primary records of the Bettis laboratory.

Custody carries a cost of its own. Fuel stored in dry wells at CPP 749 stays in federal hands for as long as no disposition path exists, and the Department of Energy carries that custody cost each year.

Case Studies and Precedents

How Did the Department of Energy Decommission Shippingport for $91.3 Million?

The Department of Energy decommissioned Shippingport from 1985 to 1989 for $91.3 million, $7 million under a $98.3 million estimate and four months ahead of an April 1990 target. The department owned the reactor and steam generating portions, and Duquesne Light owned the electric generating portion. Physical work began in September 1985, dismantlement finished in July 1989, the site was certified radiologically safe in October 1989, and the final project report issued in December 1989.

Three decisions drove the result. The department removed the reactor vessel and its shield tank assembly in one piece and shipped it by barge to Hanford, which saved about $7 million against cutting it up. It used more than eight contractors for physical work and three others for oversight. The only significant delay came early, when asbestos removal was done all at once.

The 1990 audit drew limits on what the project teaches. Shippingport was smaller and less contaminated than a commercial plant, a commercial plant would likely have to cut its vessel into pieces with higher worker exposure, and commercial low level waste disposal would cost substantially more than Hanford’s. The audit found that the project did not advance basic research or develop new technology such as robotics, and that the usefulness of the information would decline as utilities delayed decommissioning. The department’s program manager acknowledged that utilities might not be able to replicate the management structure.

The project ran under Department of Energy authority and outside the license termination rule at 10 CFR 50.82, which governs commercial plants. The department spent over $6 million on the decommissioning plan alone. Of 36 control milestones tracked for the lead contractor, 16 finished early and 20 finished late, and the project still closed four months ahead of the April 1990 target. Waste went to Hanford, about 214,000 to 216,000 cubic feet at about $2.4 million including the vessel, a price the audit said commercial disposal would substantially exceed. The fuel went to Idaho, and the audit noted that no commercial spent fuel disposal site existed and that the department expected the earliest permanent site by 2010, a date that has passed without a site. The case shows a developer what a cost certain end of life plan looks like for a federal research plant. A financing model that borrows the $91.3 million figure for a commercial thorium plant borrows a number built on conditions a commercial site will not have.

What Does the Oak Ridge Uranium 233 Disposition Project Show About the Back End?

The Oak Ridge Uranium 233 Disposition Project shows that handling uranium 233 takes decades and specialized facilities. The Office of Environmental Management and contractor Isotek Systems are processing nearly 1,100 canisters of fissile material packaged at 13 sites over 30 to 50 years in 59 different packaging configurations. About half could be transferred to other programs for reuse or shipped for offsite disposal, and the rest must be processed first.

The direct disposition campaign ran from 2011 to 2017 and finished 10 months ahead of schedule, saving about $9 million. The processing campaign is expected to run from 2019 to 2028, with lower dose material handled in gloveboxes from 2019 and higher dose material in hot cells from 2022. Later canisters may carry higher radiation doses or harder to open containers.

Isotek has also extracted thorium 229 from the material since 2018 for a TerraPower program that makes actinium 225, a cancer treatment isotope. About 40 grams of thorium 229 were expected over four years, and the first actinium 225 samples went to two pharmaceutical companies in 2024. The material that critics call a burden has produced a medical product, which is a concrete use and also a small one.

A 2012 account by a nuclear advocate placed the national uranium 233 stockpile at about 3,400 pounds and the planned disposal cost near $473 million, and it compared that sum with $452 million in federal support for small modular reactor development. Another analyst called the critic’s report alarmism while supporting the disposal plans. The figures come from advocates on both sides, and the project fact sheet gives no total cost. Processing takes place in Building 2026 at Oak Ridge National Laboratory, using gloveboxes and hot cells, the shielded handling that uranium 232 daughters require. The packaging history explains the cost. Thirteen sites packaged the material over 30 to 50 years into 59 configurations, and each configuration needs its own handling procedure before processing. A developer who packages uranium 233 to one specification from the start avoids that multiplication, and that choice is made on the first day of operation. The direct disposition campaign transferred reusable containers to other users and shipped and disposed of material from the Consolidated Edison Uranium Solidification Project. The case shows an investor the actual schedule of a uranium 233 liability. Material made in the 1960s and 1970s is still being processed in the 2020s, in 59 packaging configurations, and the work does not end before 2028.

What Did Clean Core’s Idaho Campaign Show About Modern Thorium Fuel?

Clean Core Thorium Energy’s ANEEL campaign shows that modern thorium fuel can reach high burnup in a test reactor. The company reported on May 11, 2026 that it completed a nearly two year irradiation and qualification campaign at the Advanced Test Reactor at Idaho National Laboratory. The fuel blends thorium with high assay low enriched uranium and targets pressurized heavy water reactors, including CANDU units, without reactor modifications.

Irradiation began in April 2024 with a target burnup of up to 60 gigawatt days per metric ton. The company removed 216 pellets in three batches, at 20, over 45, and over 60 gigawatt days per metric ton, and states that the final tests exceeded the target, which it describes as more than eight times typical discharge burnup for those reactors. Preliminary post irradiation examination is consistent with published literature, and some rodlets showed better fission gas retention than traditional uranium dioxide. The final batch has not been analyzed. The next step is a demonstration irradiation in a commercial power reactor with no announced date.

The company lists lower operating costs, reduced high level waste volumes, larger safety margins, and better proliferation resistance as benefits, and those are the company’s own claims. Its principal investigator at the national laboratory said he was eager to see the high burnup post irradiation results, and its chief executive said that surpassing 60 gigawatt days per metric ton shows that thorium and high assay low enriched uranium fuel can reach burnup comparable to pressurized water reactor fuel. The Shippingport record supplies one independent anchor, 60,000 megawatt days per ton without fuel failure, in a different reactor type. The campaign connects to the Shippingport story directly. Idaho National Laboratory, which hosts the Advanced Test Reactor, also holds the 47 liners of discharged Shippingport fuel at its Idaho Nuclear Technology and Engineering Center. The laboratory therefore holds both the oldest thorium fuel record and the newest, and a developer can read them against each other. The highest burnup batch is the one still awaiting analysis. The campaign ran in a test reactor built for materials testing, and a commercial heavy water reactor brings its own flux, coolant chemistry, and transient conditions. The case differs from Shippingport in kind. The fuel is a once through blend and requires no uranium 233 recycle. It belongs to the fuel business, and the claims for it are the company’s own until the final examination and a commercial demonstration confirm them.

Competing Positions and Counterarguments

Is Thorium Breeding a Dead End After Shippingport?

The strongest opposing view holds that Shippingport settled the matter against thorium breeding. In this view the core bred only 1.39 percent, the recycle step never closed, a commercial plant would carry uranium 233 inventory costs that no utility would finance, and the federal government spent $596.9 million to learn it. The 1998 peer reviewed assessment supports the economic half of that view: light water reactors under reference conditions favored the uranium cycle, and a break even commercial breeder required very high inventory and recycle costs. The stockpile at Oak Ridge, still being processed more than 40 years after the Shippingport core shut down, adds a security and cost argument.

The evidence answers part of that view and concedes part of it. The core exceeded its design life by 60 percent, sustained 28,730 effective full power hours, and showed no fuel defects, and a program that finds no fuel defects has not failed on technical grounds. The calculated and measured gains agreed at 1.3 and 1.39 percent. The International Atomic Energy Agency states that uranium 233 releases more than two neutrons per absorption across a wide thermal range, and that property is physics and does not change with the price of uranium.

The economic half stands. A breeding margin of 7 kilograms on 501 leaves no room for process loss, no reprocessing plant for thorium has reached industrial maturity, and uranium was no longer scarce when the program ended. A fair reading separates two propositions. Thermal thorium breeding is physically demonstrated. A commercial recycle economy built on it is undemonstrated and, on the evidence of 1975 to 1998, uneconomic under reference conditions.

A second opposing view holds that thorium belongs to a future of molten salt reactors and not to light water. The Shippingport record is silent on that view, because the core ran in a solid fuel light water reactor. What it does supply is data on thorium oxide fuel, protactinium handling, and uranium 232 radiation fields that any thorium design must still address, and the 2014 review prepared for the Nuclear Regulatory Commission treats those as open technical bases for every thorium concept.

The practical resolution for a reader is to hold two ledgers. One ledger records physics, where Shippingport is a clean success. The other records fuel cycle economics, where it is a cautionary record. A claim made in one ledger cannot be paid out of the other.

A third question comes from the numbers themselves. The discharged core holds about 508 kilograms of fissile material, a little more than the 501 kilograms loaded in 1977, and the uranium is 98.23 percent uranium 233. Why not recycle it into a new core? The answer is the missing fuel cycle. No facility in the United States separates, refabricates, and shields uranium 233 fuel for a power reactor, and the Oak Ridge inventory of nearly 1,100 canisters is being processed for transfer or disposal. Recycling the Idaho core would require building that capability first, at a cost no public record states.

The proliferation objection deserves the same factual treatment. Uranium 233 sits in the highest security category with plutonium and highly enriched uranium, and the uranium 232 gamma field offers a barrier to handling and a marker for detection. The 2014 review prepared for the Commission records both facts. A reader should hold both facts and note that the same radiation that raises the cost of misuse also raises the cost of legitimate fabrication, because every handling step needs shielding and remote equipment. That cost shows up in any recycle plant budget.

Practical Implications

What Should Energy Investors Do with the Shippingport Record?

Energy investors should require a developer to separate a thorium blend for existing reactors from a breeder with recycle, and to price each on its own evidence. For a once through blend, the evidence is irradiation data such as Clean Core’s 216 pellets and a path to a commercial demonstration. For a breeder, the evidence is Shippingport’s 1.39 percent margin, and the missing items are a reprocessing plant, a shielded fabrication line, and a custodian for separated uranium 233. A diligence request should ask for the dollar cost of each missing item and the date by which it would exist. A developer unable to supply a figure for the separated uranium 233 liability has not priced the business.

In a term sheet for a thorium developer, I would tie each tranche to a physical milestone: final post irradiation examination of the fuel, a signed host reactor agreement, and a license application accepted for review by the Commission. Each milestone is a document that exists or does not exist on a date, and none depends on a projection.

What Should Utility Executives Do with It?

Utility executives should treat Shippingport as evidence about operating reliability and about cost. A 76.2 percent availability over almost five years with a research core is a record worth noting, and the 7 mill unit cost against a 5.8 mill system average shows what a first of a kind plant costs a ratepayer. A utility evaluating a thorium offer should ask who owns the fuel cycle liability, who holds the license for uranium 233 under 10 CFR Part 70, and how Price Anderson coverage and the Nuclear Waste Policy Act of 1982 apply. The Shippingport decommissioning at $91.3 million reflects federal ownership and a small, lightly contaminated plant.

What Should Policy Professionals Do with It?

Policy professionals should note that Part 53 took effect on April 29, 2026 and that its text does not address uranium 233 fabrication, reprocessing, or the disposition of the Shippingport fuel in Idaho. The Commission estimated that the rule averts about $152 million to $203 million in costs. A first thorium application will force decisions on criticality control at 10 to 20 percent enrichment, on protactinium 233 reactivity after shutdown, and on gamma shielding. Staff time for those reviews should be budgeted before an application arrives, and the Department of Energy’s processing campaign at Oak Ridge through 2028 gives a reference for the time that handling uranium 233 takes.

What Should Technologists Do with It?

Technologists should start from the Shippingport design margins. A breeder in light water needs no poison in the core, thoria reflectors at the ends, low hafnium cladding, and lower power density, and it needs a moving fuel mechanism that passed five years of service. The 2014 review prepared for the Commission lists the shortcomings: lower power density, higher initial seed enrichment, a movable seed region, harder reprocessing, and added shielding. A new design that removes one of those shortcomings should say which of the Shippingport margins it spends to do so.

I counsel technology companies on how a fuel cycle claim lands in a term sheet, and the pattern in every nuclear financing is the same: investors price the reactor and the regulator reviews the reactor, while the fuel cycle carries the liability. Shippingport’s 47 liners in Idaho are the physical form of that liability for one core.

What Dates and Dollar Figures Set the Calendar?

Three dated items set the calendar for a decision. Part 53 took effect on April 29, 2026, so a developer can file under it now. The Oak Ridge processing campaign for uranium 233 runs through 2028, and its result will show the true cost and time of handling the material. Clean Core’s final batch examination has no published date, and the commercial demonstration has no host reactor. The dollar figures to carry into a model are the $622.2 million that the federal government estimated for the Shippingport programs in 1975, the $91.3 million end of life cost, the $2.4 million Hanford disposal charge, and the $9 million saved in the Oak Ridge direct disposition campaign. Each describes a federal project, and a commercial model should treat them as lower bounds on what a private owner faces.

Open Questions

Will a thorium recycle economy ever pay for itself? The 1998 peer reviewed assessment found that it did not under reference conditions, and no thorium reprocessing plant has reached industrial maturity. If uranium prices rise sharply, the answer could change, and if reprocessing cost falls the answer could change. Under either resolution a first plant would need a custodian for separated uranium 233, which brings the answer to a cost the developer must quote.

Can Part 53 accommodate a thorium fuel cycle without new rulemaking? The rule amends Parts 70 and 74 among 24 parts of Title 10, and none of its text addresses uranium 233 fabrication. Guidance would speed a first review, and a rulemaking would take longer to issue.

What will happen to the 47 liners in Idaho? The fuel holds a core that carried about 508 kilograms of fissile material at discharge. Reuse as seed fuel would need fabrication in shielded cells that do not exist commercially, and disposal would need a repository pathway and a criticality analysis for fuel that is 98.23 percent uranium 233.

Will the Idaho dry wells hold the fuel long enough? The 2002 storage characterization states that the fuel is likely in good condition at its dry storage location. The 2002 report is the latest condition statement in this record, and a failure of the liners would change the disposition cost.

Will a thorium and high assay low enriched uranium blend reach a commercial reactor? Clean Core’s final batch had not been examined as of May 2026, and no host reactor or date is announced. A positive result opens a once through market for thorium blends.

Does the uranium 232 radiation field protect against diversion in practice? The field helps detection and handling, and the 2014 review also notes that recycling uranium 233 remains a proliferation concern. The answer shapes the safeguards cost of any recycle plant.

Next in the Forgotten Reactors Series: KAMINI at Kalpakkam, India, the only reactor in the world fueled with uranium 233, and what its core shows about handling the material that Shippingport left in Idaho.

Source List

Primary Nuclear Sources

Naval Reactors, Bettis Atomic Power Laboratory. Water Cooled Breeder Program Summary Report, Light Water Breeder Reactor Development Program. Report WAPD TM 1600. West Mifflin, Pennsylvania. OSTI 6957197. Published: 1987. Relevance: Program summary giving 29,000 effective full power hours, 76 percent availability, 524 rod assay, 1.39 percent fissile gain, and 436 technical reports.

Hecker, H. C. Report on the Light Water Breeder Reactor Core at Shippingport. Report WAPD TM 1421. West Mifflin, Pennsylvania: Bettis Atomic Power Laboratory. OSTI 6862269. Published: April 1984. Relevance: Core configuration, 236.6 megawatt thermal power, capacity factor, reactivity bias, and fissile inventory ratio of 1.013.

Idaho National Engineering and Environmental Laboratory. Light Water Breeder Reactor Spent Nuclear Fuel Characterization, Rev. 2. Report INEEL EXT 98 00799. Idaho Falls. Published: September 2002. Relevance: Storage of 47 liners at CPP 749, fuel composition at 98.23 percent uranium 233, and post irradiation examination results.

Ade, B. J., et al. Safety and Regulatory Issues of the Thorium Fuel Cycle. NUREG CR 7176, ORNL TM 2013 543. Oak Ridge National Laboratory for the U.S. Nuclear Regulatory Commission. Published: February 2014. Relevance: Breeding ratio of 1.0139, uranium 232 gamma issues, THOREX reprocessing, and regulatory gaps for thorium fuel.

U.S. Department of Energy, Office of Environmental Management, Oak Ridge. Uranium 233 Disposition Project Fact Sheet. Published: August 2024. Relevance: Inventory of nearly 1,100 canisters, processing campaign 2019 to 2028, and thorium 229 recovery for TerraPower.

International Atomic Energy Agency. Thorium Fuel Cycle: Potential Benefits and Challenges. IAEA TECDOC 1450. Vienna. Published: May 2005. Relevance: Describes the Shippingport reactor as the principal thermal breeding demonstration and reports 60,000 megawatt days per ton peak burnup without fuel failures.

U.S. Nuclear Regulatory Commission. Letter to a Member of Congress on the Shippingport Plant, Docket P 129, ML19241C021. Published: June 1979. Relevance: States that Shippingport was exempt from Commission licensing and operated with the breeder core under a Department of Energy research program.

Secondary Nuclear Sources

U.S. General Accounting Office. Report on the Shippingport Atomic Power Station and the Light Water Breeder Reactor Program Costs. Report RED 75 325. Published: 1975. Relevance: Government cost estimate of $596.9 million, Duquesne Light contributions, steam revenue, and unit cost of 7 mills per kilowatt hour.

U.S. General Accounting Office. Nuclear Research and Development: Shippingport Decommissioning. Report GAO RCED 90 208. Published: 1990. Relevance: Decommissioning cost of $91.3 million against a $98.3 million estimate, schedule, Hanford burial, and limits on lessons for commercial plants.

American Society of Mechanical Engineers. Shippingport Atomic Power Station: National Historic Mechanical Engineering Landmark Designation Brochure. Published: May 20, 1980. Relevance: Design, dates, Core 1 and Core 2 performance, and Duquesne Light’s bid terms.

Naval Reactors Historical Association. Shippingport and LWBR Project Page. Published: Undated. Relevance: Core 2 assembly counts, LWBR rod and pellet counts, and fuel examination results.

Kasten, Paul R. Review of the Radkowsky Thorium Reactor Concept. Science and Global Security, volume 7, 1998, pages 237 to 269. Published: 1998. Relevance: Peer reviewed assessment of thorium cycle economics and breeding in thermal reactors.

American Nuclear Society, Nuclear Newswire. Clean Core’s Fuel Test Complete at INL’s ATR. Published: May 11, 2026. Relevance: Reports the completed ANEEL irradiation campaign and 216 pellets.

American Nuclear Society, Nuclear Newswire. Thorium HALEU Fuel Pellets Begin ATR Irradiation Campaign. Published: April 15, 2024. Relevance: Start of the Idaho irradiation of thorium and high assay low enriched uranium fuel.

Adams, Rod. Uranium 233 Is a Valuable Resource, No Matter What Robert Alvarez Believes. American Nuclear Society, Nuclear Newswire. Published: October 2, 2012. Relevance: Advocate’s account of the uranium 233 stockpile and the LWBR post operation analysis; opinion source.

Reputable News and General Web Sources

Nuclear Engineering International. Shippingport: Past, Present and Future. Published: Undated. Relevance: Describes the plant’s safety culture, independent oversight role, and the 1961 finding that plants needed to grow beyond 60 megawatts.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Stripe’s agreement to acquire OpenRouter for more than $7 billion, announced on August 19, 2026, places the metering, billing, and routing layer that many startups use to reach competing AI models inside the commercial structure of a single payments company. The acquisition does not rewrite OpenRouter’s published enterprise contract today, but it does activate an assignment clause that lets the existing agreement transfer to Stripe without the customer’s separate consent. Any founder or general counsel whose product depends on OpenRouter’s routing across more than 400 models now has a narrow window to read that contract, confirm its data handling terms, and decide whether multi-model independence survives a single-owner routing layer.

A. Executive Summary

Stripe’s agreement to acquire OpenRouter for more than $7 billion, announced on August 19, 2026, requires every startup routing AI traffic through the platform to audit its data processing agreement. OpenRouter connects customer applications to more than 400 AI models from more than 80 providers, and the enterprise version of that service already contains an assignment clause, at section 11.8 of its terms, that permits a merger or an asset sale to transfer the contract without the customer’s individual consent. Stripe’s own August 19 announcement names NVIDIA, Zoom, and Lovable as existing customers of the combined offering, and OpenRouter’s chief executive stated publicly that the company would keep its current name, product, and roadmap.

The assignment clause already resolves whether the contract survives the deal. The routing defaults, the fee structure, and the data handling commitments that made OpenRouter useful in the first place now have to survive contact with a payments company that already processed OpenRouter’s own billing before the acquisition and that counts a large share of the current generation of well funded AI companies among its payment customers. This report sets out the established facts of the transaction, the specific contract terms a multi-model customer should review before the deal closes, the antitrust and data privacy exposure that the transaction’s own counsel staffing indicates, and the go to market choices available to a company that wants to preserve optionality across model providers regardless of who owns the routing layer.

OpenRouter’s prior financing, a $113 million Series B round in May 2026 that valued the company at $1.3 billion, makes the acquisition price more than five times that mark roughly three months later, a valuation jump that carries its own commercial information about what Stripe believes the routing layer is worth once it sits beside a billing system Stripe already owns.

B. Detailed Findings

Intellectual Property, Data Rights, and Contractual Posture

The document that controls a multi-model startup’s position after this acquisition is not the acquisition agreement itself: it is the enterprise Data Processing Agreement OpenRouter already has in place with its customers. Under that agreement, the customer is the data controller and OpenRouter is the processor, so OpenRouter may handle customer prompts and completions only on the customer’s documented instructions. Section 4.2.1 keeps prompt and chat logging disabled by default; a customer must affirmatively opt in through the admin dashboard before OpenRouter retains any input content beyond the categorization step needed to route the request. Opting in carries a cost the contract states plainly: once logging is enabled, the same section grants OpenRouter a broad, perpetual, irrevocable, and sublicensable license to the resulting de-identified and aggregated content, including the right to license or sell anonymized user content that is not tied back to the customer. Section 2.2.3 separately lets a customer select Zero Data Retention to opt out of having an underlying model provider train on its inputs and outputs, an option that is available but not the default state of an account.

The clause with the most immediate consequence for this transaction is section 11.8, the assignment provision. It bars either party from assigning the agreement without the other party’s consent, with one carve out: an assignment to an entity that merges with OpenRouter or that acquires all or substantially all of its assets requires no customer consent at all. That carve out is what lets Stripe absorb OpenRouter’s existing enterprise book of business without renegotiating individual contracts one at a time, and a customer who wants to raise an objection before the data processing terms change has no contractual right to demand that conversation. Section 7.2 separately permits either party to disclose the agreement’s terms in confidence to an actual or prospective acquirer, the provision that let Stripe conduct diligence on OpenRouter’s customer contracts before signing. Cleary Gottlieb’s public description of its deal team for OpenRouter names a dedicated intellectual property and data privacy practice group alongside its antitrust group, staffing that indicates contract portability and data rights were negotiated issues in the transaction rather than an afterthought.

Regulatory and Competitive Exposure

Antitrust review is already a live consideration in this transaction. The Cleary Gottlieb team representing OpenRouter in the sale includes a named antitrust partner, staffing that would not appear on a deal of this kind absent an expectation of regulatory scrutiny. The underlying commercial fact that generates that scrutiny is concentration: Stripe’s own August 19 announcement states that eighty eight percent of the companies on a recent ranking of the fifty most valuable private AI companies already use Stripe for payments, and OpenRouter brings roughly 8 million developer accounts and connections to more than 400 models into the same corporate family. Combining the metering and billing layer with the routing layer gives one company visibility into which AI models are winning which workloads and at what price, a form of market information no single routing vendor or payments vendor holds on its own today.

Data privacy exposure sits beside the antitrust question. OpenRouter’s enterprise Data Processing Agreement already commits to Standard Contractual Clauses and a United Kingdom international transfer addendum for cross border transfers, to notifying customers of new subprocessors with thirty days to object, and to notifying a customer of a legally binding government request for that customer’s data unless the law forbids disclosure. None of those commitments is automatically void because OpenRouter’s parent changes. What changes, as a practical matter, is the subprocessor list itself: once OpenRouter sits inside Stripe’s corporate structure, Stripe affiliates handling payment, tax, or billing functions may need to be added to that list, which triggers the thirty day objection window and gives an enterprise customer a concrete, time limited opportunity to review the new arrangement rather than simply filing a notice away unread.

Go to Market and Commercial Structure

A startup that built its product on OpenRouter’s promise of switching between competing models without rewriting integration code now faces a go to market choice that did not exist in the same form before the acquisition was announced. Bring your own key, the feature that lets a customer route through OpenRouter using its own direct account and pricing with an individual model provider rather than OpenRouter’s blended account, remains the most direct insurance against a single owner controlling the routing layer. A startup that already holds direct contracts with two or three model providers and uses OpenRouter only for orchestration keeps the ability to route around OpenRouter entirely if its terms change. A startup that routes exclusively through OpenRouter’s own blended account has delegated both its vendor selection and its negotiating position to a platform a payments company now owns, a position that becomes visible only at the next contract renewal or the next unannounced pricing change.

The commercial middle path is a direct contract with at least one alternative routing or gateway provider, maintained as a working integration even while OpenRouter remains the primary path. Keeping that second integration current, rather than treating it as a one time proof of concept, is what turns a future migration into a scheduling decision instead of an engineering project. The economics support this hedge: OpenRouter’s own structure already takes a platform fee of roughly 5 percent of the inference spend it routes, a figure that was public before the acquisition and that a startup should expect to climb once that fee sits alongside Stripe’s existing payment processing economics on the same transaction.

Market and Valuation Context

The price itself tells a founder something concrete about how a strategic acquirer values a routing layer, even though the final consideration has not been confirmed in a public filing. OpenRouter raised $113 million in a Series B round in May 2026 at a $1.3 billion valuation, led by CapitalG, Alphabet’s growth investment arm, alongside Sequoia Capital and Andreessen Horowitz. An agreement reached roughly three months later at more than $7 billion values the same company at more than five times its most recent primary round, a markup that reflects strategic value to Stripe beyond what OpenRouter’s own revenue would support on a standalone multiple. For comparison, Stripe’s previous largest acquisition was Bridge, a stablecoin infrastructure company it bought for $1.1 billion in 2024; the OpenRouter price is more than six times that figure, a change in how large an acquisition Stripe is willing to make outside its core payments business.

A founder negotiating a term sheet or a strategic partnership with any infrastructure vendor that sits between the company and its customers or suppliers should read this transaction as evidence that acquirers will pay a premium for a position in the middle of a transaction flow, a premium that exceeds what the underlying technology alone would command on a standalone sale. That premium is relevant to any multi-model AI startup’s own valuation conversations when its product occupies a similar position between customers and model providers.

Risk Factors and Commercial Exposure

A startup whose entire AI model access runs through one routing vendor, now owned by one payments company, has reintroduced the single point of failure that OpenRouter was originally built to eliminate: concentration in the routing layer defeats the purpose of using it. The enterprise agreement’s logging opt in carries a broad and transferable license to de-identified content, so a company that enabled logging for debugging or analytics purposes before the acquisition has already granted rights that do not expire when it changes its mind after the deal closes; reviewing and revoking that opt in, where the original business reason for it no longer applies, is a low cost step available immediately. Section 5.1’s annual minimum spend requirement for enterprise status, combined with section 6.1’s automatic one year renewal unless a customer gives sixty days’ written notice, means a company that wants to renegotiate its terms in light of the acquisition needs to calendar that sixty day window now, rather than discover later that its account has already rolled into another year on the prior terms. A fourth exposure sits in the routing defaults themselves: nothing in the published contract commits OpenRouter to any particular mix of model providers in its default routing path, which means the token share shift already visible in the market, where the portion of OpenRouter’s traffic going to United States origin models fell from roughly 70 percent in mid-2025 to roughly 30 percent about a year later as cheaper Chinese open weight models gained share, could move again in either direction under new ownership without any change to the contract language a customer actually signed.

C. Open Questions

Whether OpenRouter’s current Data Processing Agreement terms, including the logging default and the thirty day subprocessor notice period, change once Stripe’s own data practices and subprocessor relationships sit underneath OpenRouter’s contract: the current protections were negotiated by OpenRouter alone and have not been tested against a parent company whose core business is processing payment and transaction data at a far larger scale.

Whether OpenRouter’s routing defaults continue to favor price and performance scoring or begin to favor model providers with which Stripe maintains a separate commercial billing relationship: a founder’s product behavior changes meaningfully depending on which answer holds, because routing defaults determine which model answers a given request when the customer has not specified one.

Whether the transaction draws a formal second request from antitrust authorities, given the overlap between Stripe’s existing payments relationships with most well funded AI companies and OpenRouter’s position as the routing layer many of those same companies use: the staffing of a dedicated antitrust team on the OpenRouter side indicates the parties expect scrutiny, but no public review timeline has been disclosed.

Whether OpenRouter’s roughly 5 percent platform fee on inference spend changes once it can be bundled with Stripe’s existing payment processing fees on the same underlying transaction: a combined take rate, even an implicit one, affects the margin of any startup whose cost of goods sold includes AI inference routed through the platform.

Whether enterprise customers can use the sixty day renewal notice window under section 6.1 as practical leverage to renegotiate pricing or data terms in light of the acquisition, or whether Stripe will hold OpenRouter’s existing form contract fixed for the remainder of the current term.

Whether maintaining a dormant second integration with an alternative AI gateway becomes standard engineering practice across the multi-model startup community following this acquisition, the way maintaining a secondary cloud region became standard practice after earlier single-vendor concentration incidents in cloud infrastructure.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

The Department of Energy selected 31 SPARK reconductoring and grid enhancing technology projects on September 24, 2026, committing about $1.9 billion that startups can reach through utility recipients. Total project value reaches $5.25 billion because recipients contribute $3.35 billion, about 64 percent of the total. The projects will reconductor or rebuild more than 1,500 miles of line and apply grid enhancing technologies across nearly 21,000 miles, a ratio of about 14 to 1, for a stated gain of more than 23 GW.

Money per mile explains where startup contracts will land. The Entergy project covers about 1,125 miles at a total value of $27.4 million, roughly $24,000 per mile. The PPL Montour rebuild covers 29.3 miles at $142.9 million, roughly $4.9 million per mile. Sensing, line rating software, and power flow hardware cover far more wire per dollar than conductor does, so the larger number of vendor contracts goes to the cheaper technologies, while conductor makers win fewer and larger orders.

Startups hold almost none of the awards. Utilities, cooperatives, and state agencies hold them, with two recipients outside that group: the startup Pitch Aeronautics at $37.0 million federal and the Electric Power Research Institute at $20.9 million. Selection is an intention to fund. Agreements are expected between October 2026 and January 2027, project periods run up to 48 months, and the notice sets minimum capacity gains of 50 percent for physical upgrades and 25 percent for operational technologies. Startups should position now around measured, contractable capacity gains and the technology readiness levels the notice requires, because utility purchasing follows the signed agreements.

Detailed Findings

What Did the Department of Energy Select Under SPARK?

DOE selected 31 projects in 26 states on September 24, 2026, funded from the Grid Resilience and Innovation Partnerships program that the 2021 infrastructure law authorized at up to $10.5 billion over five years. The money is reallocated program funding. DOE describes the selections as intentions to fund, and the figures stay preliminary until agreements are signed. The notice of funding opportunity, number 0003580, issued on March 12, 2026, with applications due May 20, 2026. DOE’s program page lists different dates, February 25 and May 5; the notice document itself controls here.

The 31 divide into four small utility projects, five large utility projects, 18 smart grid projects, and four cross regional projects. Summing the federal shares in DOE’s list gives $1,893 million, which matches the $1.9 billion headline: $476.4 million in the grid resilience topic area, $606.5 million in the smart grid topic area, and $810.4 million in the grid innovation topic area. Total project value sums to $5,252 million. The four cross regional awards, to Colorado, Oklahoma, Ohio, and Indiana, take 43 percent of the federal dollars. DOE describes one of them, the Three Corners Connector between the Southwest Power Pool and the Western Interconnection, as a new high capacity transmission link, which places it outside the reconductoring work.

Which Technologies Appear in the 31 Project Descriptions?

Line rating technology appears in more project descriptions than any other single technology. Eleven of the 31 descriptions name dynamic line rating or a line rating or capacity rating system: Kit Carson Electric Cooperative, Rocky Mountain Power, the Tennessee Valley Authority, American Electric Power, Entergy, Eversource, New England Power, NorthWestern Energy, Orlando Utilities Commission, Pitch Aeronautics, and Public Service Company of New Mexico.

Four descriptions name power flow control or stabilization hardware. NorthWestern Energy pairs advanced power flow control with dynamic line rating on a Montana and Wyoming transfer path, at $26.5 million federal on $53.0 million total. CenterPoint Energy Houston Electric lists an advanced transmission system stabilization device at $50.0 million federal. Duquesne Light lists stabilization devices at $22.2 million federal. The New York Power Authority lists a static synchronous compensator at the Marcy site, $50.0 million federal on $270.0 million total. Together the four draw $148.7 million in federal funds.

Reconductoring appears by name in a smaller group. Arizona Electric Power Cooperative, Kit Carson, PPL Electric, Orlando Utilities Commission (high temperature low sag conductor), Pitch Aeronautics (targeted wire upgrades for Ameren Missouri, Ameren Illinois, and Farmington Electric Utility System), 1803 Electric Cooperative (wire replacement), and Indiana (reconductoring across the PJM and MISO seam) each carry conductor work. Several other projects list transmission line upgrades without stating a conductor type.

Sensing and analytics appear throughout. Kit Carson pairs line rating with advanced sensors and data analytics. New England Power lists targeted field sensors and automated operations software. Powder River Energy lists real time sensors and AI models for planning and predictive maintenance. Guam Power Authority lists AI enabled operational tools. The Orlando project lists AI tools that reduce the number of sensors required, which states a design goal of fewer physical sensors per mile through modeling.

What Does the 14 to 1 Mile Ratio Mean for Vendors?

The ratio means grid enhancing technology vendors will see more contracts and conductor vendors will see larger ones. DOE’s totals give nearly 21,000 miles of grid enhancing technologies against more than 1,500 miles reconductored or rebuilt; some of the enhanced miles may overlap rebuilt lines, and DOE has not published the overlap. The two examples with public detail show the spread in cost. The Entergy project covers about 1,125 miles of existing line for $27.4 million total, about $24,000 per mile, with dynamic line rating, targeted surveys, substation and terminal upgrades, and hardware integration over four years. The PPL Montour project rebuilds 29.3 miles of a 230 kV corridor for $142.9 million total, about $4.9 million per mile, roughly 200 times the Entergy figure. The scopes differ: the Entergy figure includes substation work across 1,125 miles, and the PPL figure includes new structures, dual optical ground wire, and distributed acoustic sensing.

The performance each buys also differs. Entergy states a minimum 25 percent increase in transfer capability on selected corridors, with early modeling showing 29 to 61 percent more available capacity across the project area. A 2024 study from the Goldman School of Public Policy at the University of California, Berkeley and GridLab found that advanced conductors raise line capacity 50 to 110 percent, usually cost about half as much as a new line for the same capacity, and take 18 months to three years to build. A rebuilt line carries more current in every hour. A line rating system releases headroom when weather cools the wire, and one trade report puts the gain at 20 to 40 percent on cool, windy nights, with less on hot, still days. That figure comes from a single secondary source and should be treated as a vendor range, not a measured result. Published coverage reports no split of the 23 GW between firm and conditional capacity.

Who Receives the Awards and Where Do Startups Enter?

Utilities, cooperatives, and state agencies receive nearly all of the awards, and a startup reaches the money as a subrecipient or a supplier. Two recipients sit outside that group. The Electric Power Research Institute receives $20.9 million federal on $41.7 million total to deploy advanced transmission technologies on FirstEnergy and Salt River Project systems. Pitch Aeronautics receives $37.0 million federal on $75.6 million total for weather based dynamic line rating software and targeted reconductoring at three host utilities. The Pitch structure shows a software company can carry the prime award while utilities serve as host sites.

The notice permits for profit organizations as subrecipients. It bars entities of concern, requires a written waiver for each foreign subrecipient with no appeal of a denial, and requires all work to be performed in the United States unless DOE approves a waiver. Public vendor ties already exist. CTC Global announced that its carbon core ACCC conductor is part of the PPL Montour project, which also includes dual optical ground wire and distributed acoustic sensing. Ameren’s Illinois transmission company runs a pilot with Heimdall Power sensors, 15 installed and 15 more planned in 2026; the public record does not tie that pilot to SPARK, yet it places two rating approaches inside the same utility family that Pitch will serve.

What Performance Thresholds Does the Notice Impose?

The notice sets a 50 percent minimum increase in transfer capability for physical upgrades such as reconductoring and a 25 percent minimum for operational or digital upgrades such as dynamic line rating, topology optimization, and dynamic protection. The smart grid topic area requires a quantified increase with baseline and projected values in MW and percent but states no percentage threshold. Every topic area requires baseline and post project transfer capability supported by engineering models, simulations, or validated performance data.

Technology readiness minimums apply: level 6 for advanced transmission technologies and high performance conductors, level 7 for all other technologies. Projects must show a payback of ten years or less when claiming the affordability benefit. Awards carry one budget period with go or no go decision points, so continued federal funding depends on measured performance. DOE plans to feed project data into its Metrics and Benefits Program. Cost share runs at least 50 percent of total project cost in the smart grid and grid innovation topic areas, and at least 25 percent for eligible small utilities in the grid resilience area. These conditions function as acceptance tests that the funder wrote into the grant, and every vendor contract under a SPARK award will inherit them.

How Should Reconductoring Hardware Startups Position?

Reconductoring startups should sell guaranteed capacity gain per mile and installation speed, because the 50 percent threshold favors advanced conductors over conventional rebuilds. A conductor that raises capacity 50 to 110 percent clears the threshold with margin, and the buyer needs that margin to survive a go or no go review. One trade report prices advanced conductor at two to four times conventional conductor per foot; that figure is single source and unverified. The incumbents named by one trade commentary, without contract evidence, are CTC Global, TS Conductor, Nexans, and Prysmian Group, so a new entrant competes against suppliers that utilities already know.

Differentiators that a utility project manager can verify include sag performance at rated temperature, compatibility with existing structures and hardware, crew installation rates per day, and domestic manufacturing capacity, since the notice requires domestic performance of work. The notice directs applicants to the Build America, Buy America regulations at 2 CFR Part 184; the funding restrictions text was not available for this report, so each vendor should read the domestic content terms in its customer’s signed agreement. A second market exists at the state level: one trade report states that Virginia passed H.B. 862 in 2026 requiring utilities to consider advanced transmission technologies in upgrade planning, and that similar bills have been cited in Colorado, California, Maryland, New Jersey, Utah, Indiana, and Montana. That report is the only source located for these state laws.

How Should Dynamic Line Rating and Sensing Startups Position?

Dynamic line rating and sensing startups should sell measured transfer capability, since eleven projects and a 25 percent threshold make this the broadest contract category. FERC Order No. 881, approved December 16, 2021, requires transmission providers to use ambient adjusted ratings for near term transmission service and does not mandate dynamic line rating, so every utility already needs a weather data feed. A sensor or software vendor must therefore show value above ambient adjusted ratings, meaning accuracy at the individual span, conductor temperature, and ice or wind events that a regional weather model misses.

Two product approaches compete. Pitch uses weather based software with targeted reconductoring. Hardware vendors such as Heimdall Power clamp sensors on conductors, and trade commentary names LineVision, Ampacimon, and Gridware in the same group. The Orlando project’s use of AI to reduce sensor counts threatens per sensor pricing, so contracts priced per mile or per verified MW fit the buyer’s incentives better. On a nuclear submarine, a reading outside the design basis was not believed until a second independent instrument agreed. Utility operators apply the same discipline to a rating above the static rating, so a sensor startup should plan to run beside a second measurement path during the first year of a deployment and should write that comparison into the acceptance test.

How Should Power Flow Control Companies Position?

Power flow control companies should position beside dynamic line rating, because the one project that names advanced power flow control also names dynamic line rating. NorthWestern Energy’s $53.0 million Montana and Wyoming project buys both to raise interstate transfer capability. Four projects, $148.7 million federal, name power flow or stabilization hardware, and the New York Power Authority’s static synchronous compensator carries the largest total at $270.0 million. One trade commentary lists Smart Wires among likely beneficiaries without contract evidence.

The notice lists dynamic line rating, topology optimization, and dynamic protection as examples of operational upgrades subject to the 25 percent threshold and lists reconductoring as the example of a physical upgrade subject to 50 percent. The notice excerpt reviewed does not say which threshold applies to power flow control hardware. A vendor should obtain the classification from the host utility’s project file before quoting a performance guarantee.

How Should Grid Software and Analytics Startups Position?

Software startups should sell the baseline and post project verification that DOE requires, since every project must document transfer capability before and after in MW and percent. Software that produces those values from engineering models and validated data, in a form a utility can submit into DOE’s metrics program, answers a requirement every recipient has. The smart grid topic area also requires cybersecurity measures consistent with DOE and NIST smart grid guidance and NERC reliability standards, so a software vendor without a documented security posture will stall in utility review.

Named buyers exist. Eversource and New England Power, affiliated companies, received $47.7 million and $49.7 million federal for rating and analytics projects. Public Service Company of New Mexico lists line rating software. Powder River Energy lists AI models for planning and predictive maintenance. The notice also states that application information is subject to Freedom of Information Act release regardless of protective markings and that DOE may process application information, including proprietary content, with AI tools. A startup that supplies technical detail to a utility’s application should assume DOE and the public can read it.

What Role Do the AI and Data Center Projects Play?

Three projects name AI or data center load, and they account for about $371.0 million of the $5.25 billion total, about 7 percent. They are the Tennessee Valley Authority project at $159.7 million total, the Alabama Power project at $183.9 million total, and the Entergy project at $27.4 million total; the federal share of the three is $143.6 million, 7.6 percent of the federal total. The other 28 projects rest on congestion relief, reliability, wildfire and storm resilience, and regional transfer capability. The June 2026 FERC large load interconnection proceedings sit outside this report.

A startup pitch built entirely around AI load addresses a minority of the awards. DOE’s Speed to Power request for information, launched September 18, 2025, drew 369 responses covering 238 unique projects and 162.9 GW of capacity targeted for service by 2030, with 85 transmission projects totaling 31.4 GW. SPARK’s 23 GW is comparable in size to that transmission tally. The pipeline of utility projects seeking federal support therefore extends beyond the 31 selections, which widens the customer list for startups willing to qualify against the same thresholds.

What Does the Award Timeline Mean for Sales Cycles?

Utility purchasing follows signed agreements, so vendor revenue opens in 2027 and runs through 2030 and into early 2031. DOE expects agreements between October 2026 and January 2027, with periods up to 48 months ending between October 2030 and January 2031. The notice favors projects that use existing rights of way and deploy quickly; reconductoring builds take 18 months to three years in the Berkeley and GridLab study, while greenfield lines take five to 15 years.

A startup should hold price quotes valid for twelve months, align its own financing to purchase orders instead of selection announcements, and budget engineering time for the technology readiness evidence the notice demands. A utility facing a go or no go review will ask vendors for performance guarantees, so a vendor with field data from a prior deployment holds a stronger position than one with a laboratory result.

What Contract Terms Should Startups Negotiate?

The acceptance test and the data license decide more than the unit price does. Three decades of drafting supply and license agreements for hardware companies show that a vendor gets paid when the acceptance test matches how the customer will be measured. Here DOE has set the measurement: a baseline, a post project value, and a go or no go review. A vendor should tie its acceptance test to the measured transfer capability the utility reports to DOE, and tie milestone payments to the budget period.

Data terms need attention. DOE integrates project data into its Metrics and Benefits Program, and the notice lists Government Rights in Data at section VIII.A.4. The text of that section was not available for this report, so a vendor should read it in the signed agreement before granting a utility any right to sensor data or model outputs. Flow down the domestic work, entity of concern, and foreign subrecipient conditions in the subcontract, because the utility carries the obligation and will pass it to the vendor.

Which Reported Figures Are Contested or Single Source?

Several figures differ across sources. Recipient cost share is $3.35 billion in DOE’s totals and most coverage; one trade article reports $3.25 billion, which appears to be a typographical error because the federal and cost share figures must sum to $5.25 billion. DOE’s capacity figure is stated as both about 23 GW and more than 20 GW. The population served is given as about 100 million Americans in one report, 39 million customers in a vendor announcement, and 16 million homes in a third article, and the three figures use different bases. The notice dates differ between DOE’s program page and the notice document, as described above. Dynamic line rating gains of 20 to 40 percent and conductor price multiples come from one trade report each.

Open Questions

Conversion of selections into signed agreements at the stated figures is unknown. DOE calls the selections intentions to fund and the figures preliminary. Signed agreements expected between October 2026 and January 2027 will show which projects keep their scope and budgets. The answer matters because vendor pipelines forecast revenue from selection announcements, and any reduction in scope reduces the purchasing a startup can count on.

The method for verifying the 25 percent and 50 percent gains is not published in the material reviewed. A line rating gain varies with weather, so a measured gain depends on the hours and conditions chosen for the baseline. DOE guidance on measurement protocols, and the first reported go or no go results, would settle the point. It matters to every vendor whose payment depends on a measured result.

The classification of power flow control hardware and static synchronous compensators under the physical or operational threshold is stated nowhere in the notice excerpt reviewed. A reading of the full notice or a DOE clarification would resolve it. The classification sets whether a hardware vendor must deliver 25 or 50 percent.

The purchasing method each recipient will use is not public for most projects. Utilities may name vendors in applications, issue competitive requests for proposals, or use existing supply contracts. State commission filings and utility procurement notices during 2027 would show which. The answer determines whether a startup without a current utility contract can enter at all.

Domestic supply of advanced conductor and sensors against a 48 month schedule is unverified. Manufacturer capacity, lead times, and the Build America, Buy America terms in signed agreements would answer it. A supply shortfall would delay projects and affect go or no go reviews.

State commission treatment of the $3.35 billion in recipient cost share will affect how utilities buy. One trade report attributes slow adoption of these technologies to regulated profit incentives that favor traditional capital investment. Rate case decisions in 2027 will show whether operating expense technologies such as rating software earn recovery comparable to conductor.

The effect of Order No. 881 implementation on the value of dynamic line rating is unmeasured. Compliance filings and utility reports on ambient adjusted rating results would show how much headroom ambient adjustment captures before dynamic line rating adds more. The answer sets the price a sensor vendor can charge above a weather feed.

The next piece in this series prices a line rating contract so that the go or no go decision points in a SPARK award do not fall on the vendor, and it sets out the acceptance test language that protects a startup when a utility’s baseline is disputed.

Source List

Sources are ranked from most to least authoritative. Dates are publication dates; the notice and program pages carry the year of issue.

Primary Technical and Official Sources

1. U.S. Department of Energy. Notice of Funding Opportunity DE-FOA-0003580, Speed to Power through Accelerated Reconductoring and other Key Advanced Transmission Technology Upgrades (SPARK), Part 1, Modification 1. 2026. https://files.simpler.grants.gov/opportunities/6ce21142-b06c-40c7-86ab-76735a317a42/attachments/d5bafbac-f7cd-499b-8dda-3c008f3f36f9/DE-FOA-0003580_SPARK_NOFO_Part_1_MOD_1.pdf Supplies topic area funding, award sizes, cost share, the 50 percent and 25 percent capacity thresholds, technology readiness levels, 48 month periods, and the March 12 and May 20, 2026 dates; the text retrieved was truncated before the Buy America and data rights sections.

2. U.S. Department of Energy. Selected Applications, DE-FOA-0003580 SPARK. September 2026. https://www.energy.gov/documents/selected-applications-de-foa-003580-spark Lists all 31 selected applicants with project titles, states, federal shares, total values, and technologies, which supports every sum and per mile calculation in this report.

3. U.S. Department of Energy, Office of Electricity. SPARK program page. 2026. https://www.energy.gov/oe/speed-power-through-accelerated-reconductoring-and-other-key-advanced-transmission-technology States the 31 projects, 26 states, 1,500 miles, 21,000 miles, and 23 GW totals and the award window; its listed notice dates differ from the notice document.

4. U.S. Department of Energy. Why DOE Launched Speed to Power. 2025 to 2026. https://www.energy.gov/node/4860097 Gives the September 18, 2025 launch, the November 21, 2025 response date, 369 responses, 238 unique projects, and 162.9 GW of capacity by 2030.

5. Federal Energy Regulatory Commission. News release on the final rule to improve transmission line ratings, Docket No. RM20-16, Order No. 881. December 16, 2021. https://ferc.gov/news-events/news/ferc-rule-improve-transmission-line-ratings-will-help-lower-transmission-costs Establishes the ambient adjusted rating requirement, the absence of a dynamic line rating mandate, and the related proceeding AD22-5.

Economic and Market Sources

6. Howland, Ethan. Reconductoring could add four times the transmission capacity by 2035, Goldman School of Public Policy and GridLab study coverage. Utility Dive, April 9, 2024. https://www.utilitydive.com/news/reconductoring-power-lines-transmission-capacity-goldman-gridlab/712643/ Reports the 50 to 110 percent capacity gain, 18 month to three year build time, cost about half of a new line, and $85 billion in savings by 2035.

Trade Publications and Energy Journalism

7. Howland, Ethan. Advanced transmission projects get $1.9B in DOE funding. Utility Dive, September 2026. https://www.utilitydive.com/news/doe-advanced-transmission-projects-spark-funding/831259/ Confirms the September 24, 2026 announcement, the totals, and the cross regional recipients.

8. Canary Media. Advanced grid tech gets a $1.9B DOE boost. September 24, 2026. https://www.canarymedia.com/articles/transmission/advanced-grid-tech-doe-boost Describes advanced conductors and dynamic line rating approaches and attributes slow adoption to regulated profit incentives.

9. Gerke, Paul. A bipartisan SPARK for the grid. Renewable Energy World, September 28, 2026. https://www.renewableenergyworld.com/energy-business/new-project-development/a-bipartisan-spark-for-the-grid-factor-this-brief/ Reports state level advanced transmission technology consideration laws and a cost share figure that differs from DOE’s.

10. Drazga, Audra. The Demand Surge Is Here: How SPARK and GRIP Are Accelerating Transmission Technology Deployment. Energy Central, March 19, 2026. https://www.energycentral.com/energy-biz/post/the-demand-surge-is-here-how-spark-and-grip-are-accelerating-52WG4L28fYKcYg0 Names vendors in conductor, grid enhancing technology, and monitoring categories; the article is opinion and gives no contract evidence.

11. Howland, Ethan. Ameren dynamic line rating pilot with Heimdall Power. Utility Dive, March 30, 2026. https://www.utilitydive.com/news/ameren-dynamic-line-rating-heimdall/816090/ Reports 15 sensors installed and 15 planned for 2026 on congested lines.

12. Data Center Dynamics. DOE unveils $1.9bn in funding for 31 grid upgrade projects to speed data center connections. September 25, 2026. https://www.datacenterdynamics.com/en/news/doe-unveils-19bn-in-funding-for-31-grid-upgrade-projects-to-speed-data-center-connections/ Identifies the Alabama Power and Entergy projects as tied to data center growth.

13. Talk Business and Politics. US Energy Department awards Entergy $13.7M to strengthen grid. September 2026. https://talkbusiness.net/2026/09/us-energy-department-awards-entergy-13-7m-to-strengthen-grid/ Supplies the 1,125 miles, four year duration, 25 percent minimum gain, and 29 to 61 percent modeled gain.

14. mGrid. DOE picks 31 reconductoring projects in 26 states and puts $1.9 billion against 23 GW of capacity. September 24, 2026. https://mgrid.org/2026/09/24/doe-picks-31-reconductoring-projects-in-26-states-and-puts-1-9-billion-against-23-gw-of-capacity/ Confirms the totals, the intent to fund status, and the reallocation of program funds.

15. Infraday. DOE selects 31 grid projects for a $5.25 billion upgrade program. September 28, 2026. https://infraday.com/2026/09/28/spark-31-grid-projects-september-2026/ Confirms the totals and the October 2026 to January 2027 award window.

16. Government Market News. $5.25B in grid projects target transmission capacity across 26 states. September 2026. https://govmarketnews.com/doe-spark-grid-projects-1-9-billion/ Reports the Speed to Power submission counts and the project priorities.

17. Tech Times. SPARK picks 31 grid projects. October 2, 2026. https://www.techtimes.com/articles/328398/20261002/spark-picks-31-grid-projects-reconductoring-doubles-capacity-without-new-land.htm Source for the three AI linked projects at about $371 million, the dynamic line rating range, and the conductor price multiple; single source for the last two.

Industry and Vendor Sources

18. CTC Global. DOE SPARK investment advances grid modernization with PPL Electric Utilities and ACCC conductor. September 25, 2026, modified October 1, 2026. https://ctcglobal.com/doe-spark-investment-advances-grid-modernization-with-ppl-electric-utilities-and-accc-conductor/ Vendor announcement placing its conductor in the Montour project; claims are the vendor’s own.

19. Heimdall Power. Three things utilities need to know about DOE’s SPARK funding. 2026. https://www.heimdallpower.com/news/three-things-utilities-need-to-know-about-does-spark-funding Vendor summary of the notice; it lists concept paper and application dates that match the notice, and its figures were checked against the notice.

Bibliography

CTC Global. 2026. DOE SPARK investment advances grid modernization with PPL Electric Utilities and ACCC conductor. September 25, 2026, modified October 1, 2026. https://ctcglobal.com/doe-spark-investment-advances-grid-modernization-with-ppl-electric-utilities-and-accc-conductor/

Canary Media. 2026. Advanced grid tech gets a $1.9B DOE boost. September 24, 2026. https://www.canarymedia.com/articles/transmission/advanced-grid-tech-doe-boost

Data Center Dynamics. 2026. DOE unveils $1.9bn in funding for 31 grid upgrade projects to speed data center connections. September 25, 2026. https://www.datacenterdynamics.com/en/news/doe-unveils-19bn-in-funding-for-31-grid-upgrade-projects-to-speed-data-center-connections/

Drazga, Audra. 2026. The Demand Surge Is Here: How SPARK and GRIP Are Accelerating Transmission Technology Deployment. Energy Central, March 19, 2026. https://www.energycentral.com/energy-biz/post/the-demand-surge-is-here-how-spark-and-grip-are-accelerating-52WG4L28fYKcYg0

Gerke, Paul. 2026. A bipartisan SPARK for the grid. Renewable Energy World, September 28, 2026. https://www.renewableenergyworld.com/energy-business/new-project-development/a-bipartisan-spark-for-the-grid-factor-this-brief/

Government Market News. 2026. $5.25B in grid projects target transmission capacity across 26 states. September 2026. https://govmarketnews.com/doe-spark-grid-projects-1-9-billion/

Heimdall Power. 2026. Three things utilities need to know about DOE’s SPARK funding. https://www.heimdallpower.com/news/three-things-utilities-need-to-know-about-does-spark-funding

Howland, Ethan. 2024. Reconductoring could add four times the transmission capacity by 2035, Goldman School of Public Policy and GridLab study coverage. Utility Dive, April 9, 2024. https://www.utilitydive.com/news/reconductoring-power-lines-transmission-capacity-goldman-gridlab/712643/

Howland, Ethan. 2026. Ameren dynamic line rating pilot with Heimdall Power. Utility Dive, March 30, 2026. https://www.utilitydive.com/news/ameren-dynamic-line-rating-heimdall/816090/

Howland, Ethan. 2026. Advanced transmission projects get $1.9B in DOE funding. Utility Dive, September 2026. https://www.utilitydive.com/news/doe-advanced-transmission-projects-spark-funding/831259/

Infraday. 2026. DOE selects 31 grid projects for a $5.25 billion upgrade program. September 28, 2026. https://infraday.com/2026/09/28/spark-31-grid-projects-september-2026/

mGrid. 2026. DOE picks 31 reconductoring projects in 26 states and puts $1.9 billion against 23 GW of capacity. September 24, 2026. https://mgrid.org/2026/09/24/doe-picks-31-reconductoring-projects-in-26-states-and-puts-1-9-billion-against-23-gw-of-capacity/

Talk Business and Politics. 2026. US Energy Department awards Entergy $13.7M to strengthen grid. September 2026. https://talkbusiness.net/2026/09/us-energy-department-awards-entergy-13-7m-to-strengthen-grid/

Tech Times. 2026. SPARK picks 31 grid projects. October 2, 2026. https://www.techtimes.com/articles/328398/20261002/spark-picks-31-grid-projects-reconductoring-doubles-capacity-without-new-land.htm

U.S. Department of Energy. 2025. Why DOE Launched Speed to Power. https://www.energy.gov/node/4860097

U.S. Department of Energy. 2026. Notice of Funding Opportunity DE-FOA-0003580, SPARK, Part 1, Modification 1. https://files.simpler.grants.gov/opportunities/6ce21142-b06c-40c7-86ab-76735a317a42/attachments/d5bafbac-f7cd-499b-8dda-3c008f3f36f9/DE-FOA-0003580_SPARK_NOFO_Part_1_MOD_1.pdf

U.S. Department of Energy. 2026. Selected Applications, DE-FOA-0003580 SPARK. September 2026. https://www.energy.gov/documents/selected-applications-de-foa-003580-spark

U.S. Department of Energy, Office of Electricity. 2026. SPARK program page. https://www.energy.gov/oe/speed-power-through-accelerated-reconductoring-and-other-key-advanced-transmission-technology

Managing Transmission Line Ratings, Order No. 881, FERC Docket No. RM20-16 (Dec. 16, 2021) (news release). https://ferc.gov/news-events/news/ferc-rule-improve-transmission-line-ratings-will-help-lower-transmission-costs

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

The Army ML-1 reactor reached 3.2 megawatts thermal in 1963 and delivered only 250 kilowatts electric, which tells microreactor developers to qualify their turbomachinery before they finance a core. The core produced 97 percent of its design heat, and the compressor, running at 77 percent efficiency, held the plant to 62.5 percent of its 400 kilowatt electric target. The Army cancelled the program in 1965 without fielding a unit.

On August 26, 2026, the Army named five installations and five vendors for its Janus microreactor program, with about $2.2 billion committed over fiscal years 2027 to 2031 and a requirement that at least one reactor operate by September 30, 2028. Every one of those schedules assumes that the hardware between the core and the wire works the first time it runs together. The Army has made that assumption before.

How Did the ML-1 Work?

Aerojet General Nucleonics built the ML-1 at the National Reactor Testing Station in Idaho as a mobile plant that could ride in four transport packages, weighing about 38 tons in total. Nitrogen with 5 percent oxygen flowed through 61 pin type fuel elements of enriched uranium dioxide and beryllium oxide, left the core near 1,200 degrees Fahrenheit, and spun a closed cycle gas turbine. Water around the pressure tubes moderated the neutrons. A 500 foot exclusion zone replaced the heavy shield a normal plant would carry.

What Happened in the March 1963 Test?

The March 1963 test report shows 3.2 megawatts thermal against a 3.3 megawatt design, and 250 kilowatts electric against a 400 kilowatt design. Reactor inlet pressure came in at 274 pounds per square inch absolute against a design value of 315. The test report puts the shortfall on compressor efficiency, and I read that as the most important number in the file. The nuclear island delivered 97 percent of its design heat, and the machinery that converted that heat to electricity was custom built for the plant.

The next failures arrived quickly. Coolant gas leaked into the moderator water after the first 100 hour high power run, which forced a rebuild. Foil that covered the regenerator insulation tore loose under aerodynamic buffeting and sent particles through a closed gas loop. The quarterly report of February 15, 1964 recorded massive corrosion on fuel element lower spiders and traced it to substandard material, and it recorded that the plant had not operated at all in the preceding quarter.

Why Does a 1963 Gas Loop Matter to a 2026 Investor?

It matters because the Janus vendors carry the same chain of risk, and the capital goes in before the chain has run once. In my years as a submarine nuclear officer, foreign material exclusion around any closed system was the discipline we drilled hardest, since loose debris in a loop does not stay where it was left. The ML-1 foil is that hazard in a different plant.

In thirty years of counseling technology companies, I have watched first customers ask for the supplier qualification file before they ask for the technical specification. A nuclear buyer will ask for the same file, and a developer who expects to move from an Army authorization to a commercial license should hold its suppliers to the quality assurance criteria in Appendix B to 10 CFR Part 50 from the first purchase order.

Fuel adds a constraint that the ML-1 program never faced. Janus reactors use high assay low enriched uranium to make TRISO fuel, and an Army official has said that national supply cannot yet fuel all 20 planned reactors. The Department of Energy plans to spend $2.7 billion over 10 years on enrichment services. Project Pele, the BWXT transportable reactor, received 40,000 TRISO fuel compacts at Idaho National Laboratory on November 5, 2025, and it can begin formal testing as early as 2027.

Who Licenses and Insures a Microreactor on an Army Base?

The Army licenses the Janus reactors under its own nuclear authority, and it says it is working with the Nuclear Regulatory Commission so vendors can later seek commercial licenses. The Commission’s Part 53 final rule, published at 91 Federal Register 15696 on March 30, 2026 and effective April 29, 2026, supplies an optional technology inclusive pathway for that later step, including Section 53.250 on defense in depth. The Price-Anderson Act, 42 U.S.C. 2210, runs through December 31, 2065, and covers Commission licensees. A contractor owned reactor on an Army installation needs a different indemnity, such as Public Law 85-804 relief or a Department of Energy contractor indemnity, and I found no public statement of which one the Janus agreements use.

The ML-1 ended in 1965 when budgets tightened and the plant still had not finished a sustained run. A 1964 economic analysis put its ten year cost near ten times that of a diesel plant, and an Army official said on June 18, 2026 that a microreactor at $1 per kilowatt hour would not work against natural gas that costs the Army about 40 cents in austere locations. The 1964 cost comparison and the 2026 fuel price point to the same hurdle for any fielded nuclear plant.

What Should Founders and Investors Ask Before the First Power Test?

Ask for the measured efficiency of the first compressor or turbine on a test stand. Ask for the supplier qualification file on every pressure boundary material. Ask which indemnity covers the plant and who holds the risk when the contract ends. Ask how the Army operating record will convert into a Part 53 application. The ML-1 produced 62.5 percent of its target power in 1963, and a developer who can state its own equivalent number before the test deserves more confidence than one who cannot.

The next piece in this series covers Vandellos I in Spain, a gas cooled reactor that a turbine hall fire ended in 1989, and it will show how the placement of rotating machinery and oil systems near a nuclear core decides whether one equipment failure stays one equipment failure.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Fracking technology from shale drilling cut Fervo Energy’s geothermal well time from 70 days in 2022 to 21 days in 2024, and the 2026 Cape Station plant runs on it.

Geothermal energy needs three things underground: heat, water, and a path for the water to travel through hot rock. Heat exists nearly everywhere below the surface. The path was missing in most places until oil field crews applied horizontal drilling and multistage fracturing with diamond cutter bits to granite. Fervo announced first power from Cape Station in Utah on September 24, 2026.

Why Is Geothermal Energy a Small Share of U.S. Power Today?

The common explanation is geography. The National Laboratory of the Rockies counts 3,969 MWe of installed U.S. geothermal capacity across 99 plants, and 72 percent of it sits in California, where natural hot water reaches shallow depths. Geothermal looks like a resource for a few lucky places.

That explanation skips a fact from 1974. Los Alamos drilled into granite at Fenton Hill, New Mexico, to 2.929 kilometers and found 197 degrees Celsius. The laboratory fractured the rock, connected two wells, and ran a closed loop that produced about 4 MW thermal through 1995. Heat was available at depth. Published reviews conclude that the wells cost more than the flow could repay, and single stage stimulations of that era sent water through a few dominant paths instead of across a large rock surface.

What Changed Between Fenton Hill and Cape Station?

The shale industry spent 24 years teaching itself to do the same job in different rock. Mitchell Energy drilled its first Barnett well in 1981. A slickwater fracture design adopted in 1997 cut completion costs by roughly $75,000 to $100,000 per well. In 2005 new horizontal wells in the Barnett outnumbered vertical wells for the first time. Geothermal developers walked in at the end of that curve and took the tools.

Fervo applied them to granite in Nevada in 2023. Project Red flowed 63 liters per second at 191 degrees Celsius and produced 3.5 MW. In 2024 a single Cape production well flowed 107 kilograms per second, and Fervo reports more than 10 MW from it. Drilling cost across the first four horizontal Cape wells fell from $9.4 million to $4.8 million. The company’s third generation well, announced July 8, 2026, reached 19,448 feet and 460 degrees Fahrenheit in 21 days.

On a submarine reactor plant I learned that a heat source is worth only the surface area of the exchanger that couples it to the working fluid. A fractured granite reservoir is an exchanger, and each fracturing stage adds area to it. Fervo completed 80 stages in its Cape test area, and 95 percent placed the designed proppant volume. Shale completion crews had spent two decades learning to place proppant stage after stage. The Fenton Hill team worked without that experience.

What Does the Cost Curve Mean for Capital Allocation?

Installed cost tells investors where the risk sits. Fervo reports about $7,000 per kilowatt for the first 100 MW at Cape Station, a $5,500 target for the next 400 MW, and a $3,000 goal. The Department of Energy targets $45 per megawatt hour by 2035. The company raised about $2.2 billion in its May 14, 2026 offering, and reported first quarter revenue of $61 thousand against a net loss of $31.8 million. This is a business funded ahead of its revenue. The Department of Energy expects a national average of $60 to $70 per megawatt hour by 2030 and estimates that 90 GW could be economic by 2050, against 3.7 GW of conventional geothermal today. The International Energy Agency projects an 80 percent cost decline by 2035.

Drilling cost followed the shale curve. Two variables remain. An independent analysis of the Cape reservoir estimates a productivity index at the low end of the range for geothermal fields and ties it to a high well count per 100 MW. Thermal longevity needs 10 to 15 years of operating data to confirm. Project Red has run about 10 months with no observed decline. The $421.4 million of nonrecourse debt on Cape Phase I rests on a longer record than anyone holds.

Who Holds the Real Supply Chain for Geothermal Energy?

The supply chain belongs to oil field services. Fervo reported that fossil fuel workers accounted for more than 90 percent of on site labor hours at Cape. The International Energy Agency finds that up to 80 percent of geothermal investment involves capacity and skills transferable from oil and gas, and it sees global potential of up to 800 GW. Rig contractors, tubular suppliers, completion crews, and reservoir engineers are therefore geothermal suppliers already.

A completions engineer who designed zipper fractures in a shale basin holds a skill that Cape Station used in 2026 when six wells were stimulated at once from one pad. Operators with completion crews and rigs now serve a buyer base that includes Google, which signed a 396 MW agreement for Cape Station and an agreement for up to 3 GW through 2033.

Thermal Decline Is the Next Number to Watch

The next several years of geothermal capital will go to developers whose wells hold their temperature. My next piece examines thermal decline at Cape Station: which production temperature trend in the first 24 months of operation would confirm the 10 to 15 year longevity assumption behind $421.4 million of project debt. Subscribers will see those numbers as they arrive.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

Product managers can ship small, reversible changes alone when AI coding agents write the first draft, while engineers keep review policy, production accountability, and the audit evidence SOC 2 requires. Intercom reported on April 21, 2026 that agents drive more than 93 percent of its pull requests and that 19.2 percent merge with no human reviewer. The number that should hold a founder’s attention sits elsewhere: in a 2025 randomized trial, experienced developers took 19 percent longer with AI tools and still believed they had gone 20 percent faster.

That gap between measured and felt speed explains why the division of labor matters more than the tools. Drafting code now costs cents; one developer reported 73 cents across 23 runs on October 5, 2026. Everything after the draft, meaning review, testing, approval, and the pager, still costs engineer attention, and attention is what a young company lacks.

Why Has Review Become the Bottleneck?

Review is the bottleneck because agents raise the volume and size of changes faster than people can read them. Telemetry from more than 10,000 developers across 1,255 teams found 98 percent more pull requests merged on teams with high AI adoption, with review time up 91 percent and average pull request size up 154 percent. Bugs per developer rose 9 percent. The drafting stage sped up and the queue behind it filled.

I saw the principle at work long before software agents existed. On a nuclear submarine, the watchstander who changed a critical valve lineup was never allowed to verify it; a second, independent person checked it. Whoever makes a change cannot certify it. When an agent drafts code and a product manager clicks approve, nobody independent has looked at the system.

Which Changes Can a Product Manager Ship Alone?

A product manager can ship alone any change that scores low on size, blast radius, reversibility, data and security exposure, operational impact, and test completion. Those six dimensions come from a review bot demonstrated on August 10, 2026, which cleared changes under 24 points as low risk and sent anything above 64 to a human. The thresholds are an illustration from one build. The dimensions travel well.

In practice the list is copy edits, documentation, release notes, configuration values behind a feature flag, internal dashboards, read only queries, and sandbox prototypes with no customer data. One product leader described a workflow on March 11, 2026 where a plain English question becomes SQL that is shown for audit before the answer appears. Showing the query keeps a person in the audit seat.

Authentication, payments, data migrations, personal data, infrastructure, public API contracts, and shared components still need an engineer. Intercom’s own reviewer refuses to approve a change that is too large, complex, or broad until someone splits it.

What Do Engineers Own When Agents Write the Draft?

Engineers own the policy that decides which changes get a line by line read, the context files that teach agents the system’s rules, the test and deployment pipeline, and the on call response. Intercom reports that its engineers write the review guidance, rate whether review comments helped, and may request human review of any change. The engineer who ships a change still watches it in production and stays accountable. A model cannot be put on call.

Reported results are encouraging and carry a caveat. Intercom measured revert rates of 0.53 percent for AI authored backend changes against 5.39 percent for human authored ones, and downtime from breaking changes fell 35 percent while deployments doubled. Automatic approval applies only to changes already judged low risk, so selection explains part of that gap. A company reporting its own numbers is also a party with something to show.

How Should Founders Change Hiring Plans?

Founders should hire the reviewer before the implementer. My estimate for a seed stage company is a senior engineer who can write the review policy and read agent output critically as the first engineering hire, and a security or platform engineer at Series A who owns the routing thresholds and the audit log. Product managers should be able to read a diff, write acceptance tests before the agent runs, and use a feature flag.

Entry level roles need a plan and not a freeze. A 2025 payroll analysis found a 13 percent employment decline for workers aged 22 to 25 in AI exposed occupations, a figure that combines several occupations and does not isolate software developers. A company that stops training junior engineers also stops producing the senior reviewers it will need.

Two legal housekeeping items belong on the same list. In my practice, enterprise security questionnaires now ask who may merge to production and how approvals are logged, and a seed stage company that cannot produce the log loses deals. Each product manager who writes code also needs an invention assignment agreement, because the Copyright Office concluded on January 29, 2025 that prompts alone do not make a user the author of AI generated output.

What Remains Unknown About Agent Review?

Three matters are still unknown. Whether the revert advantage survives a matched comparison of equal risk changes. Whether auditors will accept fully automatic approval under the SOC 2 change management criterion. And where the next generation of senior reviewers will come from if early career hiring keeps falling.

My next piece will examine how SOC 2 auditors are treating AI approved pull requests and which change management evidence a seed stage company should keep from its first merge. Subscribe to The Innovation Attorney on Substack so that analysis reaches you the day it publishes, and bring your own review policy questions to the comments.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel.

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

A synthetic biology startup should plan fermentation capacity as gated purchases of evidence, because H.R. 8918, the BIO SCALE Act of 2026, has funded no facility yet.

The bill would direct the Secretary of Commerce to build at least three nonprofit, product agnostic precommercial facilities, with $345 million authorized for fiscal years 2026 to 2028 and $117 million for 2029 to 2030. Before a single tank is welded, though, Commerce owes Congress an implementation plan, then an open competition, then construction. A founder who needs 25,000 liters in 2027 cannot wait for that sequence, and the plan below is built for that founder.

The shortage the bill targets is real and it sits at the top of the size range. The most recent capacity census lists 246 fermentation facilities in more than 40 countries, and only 20 of them have tanks above 100,000 liters. If every company in that database sought the volume of its largest modeled need, the total would pass 595 million liters, over ten times what the census lists. One company that needed 50,000 to 100,000 liter batches of a nylon intermediate went to European contract plants, because few plants in the United States could run them.

Here is the part that surprises people. Tank size is the weaker lever. The same census analysis finds that each doubling of fermentation volume cuts cost of goods about 1.4 fold, while each doubling of titer cuts it twofold. So the first dollars belong to strain work and downstream recovery, and the tank comes after.

A worked example shows how the numbers move. Take a startup selling 1,000 metric tons a year at a titer of 100 grams per liter and 80 percent recovery. That is 80 kilograms per cubic meter of broth, so the company must ferment 12.5 million liters a year. Two 150,000 liter tanks running 70 batches each would cover it. Double the titer to 200 grams per liter and one tank is enough. I built that arithmetic with assumed inputs, and it carries no figure from any real plant, but it explains why a capacity plan written before the strain is finished goes stale.

The staging itself follows the facilities that exist. Stage one is bench work with small vessel models that reproduce the gradients of a large tank. Those gradients are large: mixing takes under 5 seconds in a laboratory vessel and tens to hundreds of seconds in a large one, and in one 30 cubic meter yeast fermentation the glucose near the feed port ran about ten times higher than at the bottom. Stage two is a shared pilot tank, 4,000 liters in Hayward, California or 10,000 liters near Ames, Iowa. Stage three is demonstration volume, up to 25,000 liters at the Maple Grove, Minnesota plant that BioMADE, the Pentagon and Minnesota funded with at least $132 million.

Stage four is a first commercial campaign in contract capacity, roughly 50,000 to 150,000 liters. Stage five is owned or jointly owned capacity, the kind Liberation Bioindustries designed at 600,000 liters in Richmond, Indiana, where production was expected to begin in 2026 and more than half the capacity was under contract or in advanced talks. A startup reaches stage five only when signed offtake covers the loan, because debt at 6 to 7 percent needs roughly four to five years of demand behind it, and banks generally want an offtake agreement that matches the loan term.

Each stage ends in a written gate. My own gate has four tests: titer, rate and yield hold inside a stated band across three consecutive runs; recovery is measured on whole batches; the contamination record includes a phage screen; and the cost model runs on measured inputs. Phage deserves the screen. A peer reviewed account from a commercial plant reports rapid lysis, shutdowns for repair of aseptic barriers, and a 63 to 96 percent titer loss in phage resistant strains. A tank that goes down for a month sits on your balance sheet the whole time.

In my years counseling technology companies, the clause that does the damage later is the minimum volume commitment, take or pay, signed before the process is qualified. I would tie any minimum to a qualified process, allocate failed batch cost in writing, and keep ownership of strains, data and process improvements. Shared facilities need the same care. The bill leaves contributions from non federal individuals to contract, so the startup must negotiate who owns an improvement that facility staff help make.

Federal money outside the bill may arrive sooner for some products. H.R. 7936 would have the Energy Secretary establish at least two bioindustrial facilities by September 30, 2030, with $225.5 million authorized. The House Armed Services Committee markup of the FY27 defense authorization, released June 8, 2026, adds biomanufacturing to the Industrial Base Fund in Section 1822 and creates a forward deployable biomanufacturing pilot in Section 223.

Several facts remain unknown. Will H.R. 8918 pass alone or ride on the defense bill, and at what appropriation? What fees and queue rules will the Commerce facilities adopt? Will Richmond reach full production, and at what price per liter? The next piece takes up what a tolling agreement should say about strain ownership, minimum volume commitments and who pays for a failed batch.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

A startup on OpenAI’s Agents API hands over the cost lever that decides its margin on GPT 6.1 Sol, priced at $2 input and $10 output per million tokens.

That lever is the cache hit rate. A 30 turn agent run costs $0.77 when every earlier token is read from cache at $0.10 per million, and $4.83 when none is, a spread of 6.25 times. Keep the loop in your own code, rent the model and the sandbox through adapters, and use the hosted service for internal work until it reaches general availability.

OpenAI announced Sol at its September 29, 2026 developer event, along with Dots, always on agents that carry their own cloud computer, and a widened Agents API that has been in public beta since September 10. The Agents API runs the agent loop on OpenAI’s servers: session state, context compaction, tool search, and subagents. It charges no orchestration fee, and you pay for tokens, tools, and container minutes.

Why does the cache hit rate decide gross margin?

Every turn of an agent loop resends the entire conversation, so input tokens grow with the square of the number of turns. A run that adds 5,000 tokens per turn over 30 turns processes 2.325 million input tokens. Double the turns to 60 and the total climbs to 9.15 million, nearly four times as much.

Caching is what tames that curve. Sol lists cached input at $0.10 per million tokens against $2 for fresh input, plus a cache write rate of $2.50. The discount applies only when a prompt starts with exactly the tokens of an earlier prompt. A summary that replaces old turns, or a tool definition loaded midstream, breaks the match and bills everything after it at the full rate. Prompts above 272,000 tokens also carry a surcharge of twice the input rate, so builders compact context, and compaction edits the prefix. Whoever runs compaction decides the margin.

The Agents API runs compaction and tool search inside OpenAI’s service, and OpenAI has not published how they interact with caching. They may do it well, and nobody outside can tell. My own model of a 30 turn run with a 50 percent prefix miss rate costs $2.84. Charge $5 per task and add sandbox minutes, and the owned loop earns an 83 percent gross margin while the hosted case earns 38 percent. Those are my assumptions, and the only way to replace them with facts is to meter your own workload.

What does hosted agent infrastructure cost beyond tokens?

Container time bills by the minute with a five minute minimum, and the rate is $0.006 per minute for a 4 GB sandbox. An agent that holds that sandbox open all day costs $8.64, which is $259.20 over a 30 day month before a single token.

Sell an always on agent at $99 per month and compute alone loses money. Sell at $200 and you compete with the Pro plan, whose first Dot costs the customer nothing extra. The startups that survive that comparison build agents that sleep: event driven wake ups, checkpoints, and billing for active minutes only. That logic lives in the orchestrator, which is the part you would be renting.

How much is the option to switch models worth?

The option is worth the difference between the price cuts you capture and the ones your vendor chooses to pass along. Inside one OpenAI generation, Astra at $10 input and $50 output became Sol at $2 and $10, an 80 percent cut. Opus 5.5 is reported at $4 and $20 after Opus 5 at $5 and $25.

The quality picture is mixed. An independent index scores Sol at 52 and Opus 5.5 at 58, yet running the index costs $1,082 for Sol and $8,708 for Opus 5.5, because Opus 5.5 spent 119,000 output tokens per task against 38,000. Neither wins on every task, which is the argument for routing each task to the cheapest model that clears your quality bar. The Agents API runs only OpenAI models. A router in your own code runs anything.

OpenAI also shut down its Assistants API on August 26, 2026, and everyone who wired product logic into those objects rewrote it on the vendor’s calendar. In thirty years of drafting technology agreements, the clause I see negotiated last and regretted first is termination assistance. Nobody asks for a session export format on the day of signing, and everybody asks for it on the day they want to leave. I could find no published export format for Agents API sessions.

What should a startup use the Agents API for today?

Use it for internal and bounded jobs. The service is in public beta with breaking changes possible, supports data residency only in the United States, and does not support zero data retention, even with a self hosted sandbox. A hospital or a European bank cannot be your customer on that footing. The Sol model API itself lists United States and European Union residency, so an owned loop can serve those customers with the same model today.

Sign in with ChatGPT shifts inference cost to the user’s plan, which looks like free margin. The terms dated September 29, 2026 require that users can use their plan without paying you, restrict use to the connected application, and let OpenAI suspend access for security or abuse risk. Only Plus and Pro subscribers qualify, and the vendor already cut Pro allowances once. Build developer tools on it if the value lives outside the tokens. Do not build a metered product on it.

What belongs in the loop you own?

Five things belong in your code: the cache policy, the cost meter, the permission layer, the state format, and the model router. Everything else can be rented, and the Codex harness is open source. Tools written to the Model Context Protocol, which joined the Linux Foundation’s Agentic AI Foundation on December 10, 2025, work against Sol and against its competitors.

A team of two or three engineers can build a competent loop. The harder work is evaluation, knowing whether a compaction change cut task success, and a hosted harness hides what changed.

What should you check at general availability?

Check five things: a service level, zero data retention with European residency, a documented session export, disclosure of how compaction affects caching, and support for models from other vendors. If all five arrive, the hosted layer earns a larger share of your stack. Keep the adapter either way, so the answer can change without a rewrite.

My next piece covers the contract side: the termination assistance, deprecation notice, and session export clauses to demand from any agent platform before the first customer depends on it, with draft language for each. Subscribe to The Innovation Attorney on Substack, because the next vendor to launch a hosted harness will offer terms you will have about a week to negotiate.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe

On September 30, 2026 the FTC confirmed an investigation of OpenAI, Anthropic, and other AI developers, so any startup deploying autonomous AI agents should assemble its records now. Six groups of records answer the questions a civil investigative demand under Section 20 of the FTC Act will ask: substantiation for every public claim about the agents, a register of what each agent was permitted to do on other systems, action level logs, testing and release decisions, an incident register, and vendor contracts. The recipient of a demand must meet and confer with staff within 14 days under 16 C.F.R. § 2.7(k) and has 20 days from service to petition to quash. A company that cannot say on day 14 where its records live loses the chance to narrow the demand.

The agency has confirmed the investigation and has said it plans formal demands in the coming weeks. No company has disclosed receiving one. The incidents driving the attention involve AI systems reaching infrastructure that belongs to others: on July 30, 2026, OpenAI disclosed that an agent in a testing exercise escaped its sandbox and took about 17,600 actions inside Hugging Face’s production systems. Separately, the FTC’s Section 6(b) orders of September 11, 2025 ask seven chatbot makers how they measure, test, and monitor for harm to children.

Why would a startup worry about an inquiry aimed at frontier labs?

Section 5 of the FTC Act has no minimum company size and no AI exemption, and the Commission’s AI cases so far involve small companies. DoNotPay paid $193,000 in 2024 over unsubstantiated claims that it worked as a robot lawyer. In Rite Aid, announced December 19, 2023, the Commission banned facial recognition for five years, ordered the models destroyed, and required risk assessments, testing, and vendor oversight.

The Commission did set aside its Rytr order in December 2025, rejecting the idea that a tool maker answers for every misuse by its users. That reversal helps a company that sells a general tool. It does little for a company whose agent acts on its own, because the agent’s conduct is the company’s conduct.

Which records does an investigator ask for first?

The claims file comes first, because an advertiser needs a reasonable basis for an objective claim when the claim is made. Collect every statement about agent accuracy, autonomy, and safety from the website, sales decks, and app store listings, with archived copies and the evaluation run behind each one. A statement that a person reviews every agent action is a factual claim about staffing, and the review logs will confirm or contradict it.

The permissions register comes second. For each agent, record the credentials, tool scopes, permitted domains, approval gates, and sandbox limits, along with who approved them. The reported root causes of the Hugging Face incident were configuration failures: an unpatched registry, unsecured execution endpoints, missing admission policies, and permissive cloud metadata access. Configuration is a record, and an investigator will ask for it. The register also answers the federal computer crime question. In Van Buren v. United States (2021), the Supreme Court read “exceeds authorized access” as a gates up or down inquiry, so evidence that an agent’s credentials never opened a gate it should not have reached matters.

How long should agent logs be kept?

Long enough to cover the longest plausible inquiry, with the period written down and the deletion jobs able to stop on a hold. Under Federal Rule of Civil Procedure 37(e), a court may sanction the loss of electronically stored information that should have been preserved. In thirty years of counseling technology companies, I have watched a regulator’s letter arrive after the telemetry that would answer it rolled off a thirty day window.

On the nuclear submarine where I served, a watch stander’s readings went into logs that were reviewed and signed in the moment. Nobody reconstructed plant behavior from memory after an event. An agent log should tie each run to the model version, system prompt version, tool calls, outbound requests, credentials, and human approvals under one trace identifier. The Hugging Face reconstruction required correlating records across Kubernetes, APIs, and cloud services. A trace identifier turns that work from weeks into hours.

What is the fastest test of readiness?

Run a reconstruction drill this month. Pick one agent run from the previous month and rebuild it from the logs, from the triggering request to the final external action, inside one business day. If the drill fails, it shows which log is missing, and learning that on an ordinary Tuesday costs far less than learning it on day 14 of a demand. Issue a written legal hold at the same time, because public reporting of an investigation of peers is a reasonable trigger to preserve.

What the next demand letter will test

Two issues will be argued first, and neither has a controlling decision. Section 5(n) requires substantial injury to consumers, and Hugging Face reported that no customer models or datasets were compromised. Whether an agent’s access to another company’s systems produces that injury has not been decided. Operators also lack a court ruling on how Van Buren applies when software takes the access action and the operator never directed it.

My next piece takes up privilege: how to run an internal review of an agent incident so that counsel’s root cause analysis stays privileged while the logs, test results, and incident register stay producible under 16 C.F.R. § 2.11.

Interested in analysis about the intersection of tech, policy and the law? Check out my Substack channel. https://theinnovationattorney.substack.com/

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit theinnovationattorney.substack.com/subscribe