AI to ROI

Stripe agreed to pay $7.5 billion for OpenRouter, roughly six times the $1.3 billion valuation the company raised at just 85 days earlier. Ray Rike and Peter Buchanan break down why a payments company bought the plumbing that routes AI model traffic, what OpenRouter's 5.5 percent pass through economics actually look like, and why model routing has become the control point for enterprise token spend. The bigger story is not the headline multiple. It is that CFOs have run out of patience on AI cost, and the routing layer is where token maximization turns into token cost optimization.

• The deal math. A three year old company with fewer than 100 employees and 8 million developers on the platform went from a $1.3 billion round in May to a $7.5 billion acquisition in mid August. Annualized revenue was roughly $140 million in July, up 3x since April, then up another 15 percent to $160 million within weeks. Reported split: about $1.5 billion to the founders, $6 billion to investors.

• The business model. No subscriptions, no site licenses, no annual contracts. OpenRouter takes 5.5 percent of prepaid credit value and passes 94.5 percent through to the model provider with no markup. Bring your own key customers pay a 5 percent overage fee above thresholds of roughly $25,000 per month on standard plans and $200,000 per month on enterprise.

• The contrarian bet that paid off. Founded in early 2023 when consensus said a handful of frontier labs would win outright, OpenRouter bet no single model would win and that developers would need a neutral layer in the middle. With roughly 70 percent of token traffic now flowing to open weight models and 55 trillion tokens per week crossing the platform, that bet looks prescient. The routing data is the moat, not the software.

• Why OpenAI subsidized a gateway it does not own. OpenAI cut prices on three models by half on OpenRouter to win developer share on neutral ground. Distribution is the scarce commodity right now, and at a 50 percent discount the effective customer acquisition cost approaches zero. The discount is almost certainly temporary. Anthropic, with an IPO closer in, did not match it.

• A market with no agreed definition. Gartner sizes the pure play AI gateway market at roughly $250 million in 2025 growing to about $2.9 billion by 2031, a 51.6 percent CAGR. IDC frames it far more broadly as AI cost governance and orchestration at roughly $11.2 billion, potentially subsumed into a $54.8 billion API management market. The category is crowded fast: Portkey, Requesty, Kong, Eden AI, plus gateways shipped this year by Ramp, Snowflake, Databricks and Cursor, and a rumored Meta entry.

• What boards and CFOs should watch. The open question is whether Stripe keeps OpenRouter operating as a standalone neutral entity for the next two to three years or bundles routing, metering, billing and payments into a single lock in system too early. Ray's guidance to enterprise buyers: resist the instinct to negotiate AWS style multi million dollar model commitments. Models are evolving too quickly to lock in on today's price performance. Preserve the flexibility to switch, and evaluate every routing provider on one question. Is it an honest broker?

Read the full analysis in the August 25th AI to ROI Big Story newsletter at ai2roi.substack.com, published every Tuesday alongside our weekly summary and analysis of the top 10 stories in AI.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Ray Rike sits down with Sundeep Goel, co-founder and CEO of Mavvrik, for a deep dive into the second annual AI cost governance research report from Maverick and Benchmarkit. With 30+ years in enterprise software, Sundeep brings a practitioner's view of why AI spend is proving far harder to forecast and govern than cloud spend ever was, and what finance and technology leaders need to do about it.

• Forecast accuracy is getting worse, not better. Only 11% of companies can forecast AI spend within plus or minus 10% in 2026, down from 15% the prior year, even as aggregate AI spend has climbed sharply.

• Unexpected AI costs are already changing business decisions. 49% of companies with AI-powered digital products have had to modify pricing, and 25% have stalled or shelved an AI project due to cost overruns.

• Traditional per-seat and percentage-of-spend pricing models are breaking down. Variable, high AI cost of goods sold is pushing companies toward consumption-based and outcome-based pricing.

• Cost tracking and cost allocation are two different problems. 98% of companies say they track AI costs at an aggregate level, but only about 5% can allocate that spend down to the customer, application, or agent level where it actually drives decisions.

• Cost governance is the precursor, cost optimization is the destination. Drawing a parallel to the 30%+ waste long documented in cloud spend, Sundeep argues a similar magnitude of waste exists in AI spend today, and that dynamic model selection, prompt caching, and rate optimization are the next frontier.

• ROI has a solvable half and a harder half. Cost is measurable today; value remains subjective and requires a company-specific framework to quantify.

Ray and Sundeep close with three rapid-fire questions on who should own AI ROI measurement, the key variables for driving it, and advice for early-career professionals navigating an AI-disrupted job market.

If you're finding value in the insights our guests share, subscribe to the AI to ROI podcast on your favorite platform and leave a five-star rating. Have a guest suggestion? Reach out to Ray Rike on LinkedIn.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

For 24 months, the market narrative held that Microsoft was losing the AI race to the lab it had funded. Then came the July 30th earnings call, and the stock added roughly $450 billion in market capitalization the next day, the largest single-day gain by any public company on any exchange in history.

In this week's Big Story, Ray Rike and Peter Buchanan work through why the reassessment happened and what it says about where enterprise AI value is actually accruing.

The thesis is not that Microsoft builds the best model. It does not. The thesis is that Microsoft has figured out how to monetize the gap between model quality and market value at the exact moment the novelty premium on frontier AI is wearing off and enterprise buyers are starting to ask price-performance questions instead of capability questions.

What the episode covers:

• The numbers behind the trade. Azure grew 43% and crossed $100 billion in annual revenue for the first time. The AI-specific business, bundling AI consumption with Copilot and GitHub, hit a $37 billion annual run rate, up 123% year over year. Total quarterly revenue reached roughly $90 billion, up 18%, with net income up 31% to nearly $36 billion.

• The concentration question inside that number. Roughly $24 billion of the $37 billion AI run rate is hosted by OpenAI, about 7% of the company's total revenue. Peter frames the two ways to read it, as systemic customer concentration risk or as evidence that hundreds of thousands of workloads now run through Azure, and explains why he leans toward the second.

• The toll booth position. Microsoft is the only cloud provider that hosts the OpenAI, Anthropic, and Mistral models alongside its own MAI family on the same platform. As model orchestration becomes a real buying criterion, that means an Azure customer never has to leave the platform to route traffic across model families, and gets one contract and one bill for all of it.

• The equity stakes that make Microsoft indifferent to who wins. A 27% position in OpenAI now valued around $220 to $240 billion, plus an Anthropic stake that produced a $3.2 billion unrealized gain last quarter and added 33% to earnings per share.

• Seven years of OpenAI partnership history and how it changed. The original $1 billion investment in July 2019 in exchange for Azure exclusivity, the $13 billion follow-on in 2023 with exclusive commercial API rights, the mid-2025 strain when OpenAI signed separately with Oracle for compute, and the public benefit corporation restructuring that converted profit sharing rights into equity while Microsoft gave up hosting exclusivity, retaining a first mover window on new releases and non-exclusive licensing rights running through 2032.

• The MAI model family and Satya Nadella's frontier diffusion strategy. Seven models announced at Build in June with Microsoft owning the IP. MAI Thinking One is a mixture-of-experts reasoning model with about a trillion parameters, only 35 billion of which are active at any time, making it cheap to run at scale. MAI Code One Flash, a 5 billion parameter coding model, became the default in GitHub Copilot. Image, voice, transcription, and cybersecurity models fill out the set.

• What that routing actually saves. Microsoft reported an 84% reduction in GPU costs for image generation in PowerPoint and an 89% reduction in GPU costs for voice processing in the Dynamics 365 contact center by keeping high-volume, repetitive traffic on models it controls end-to-end rather than sending it to a frontier lab.

• Two production examples. Dragon Copilot, the clinical assistant built into Dragon Medical One, and DAX Copilot that drafts clinical notes from a patient visit, and Project Perception, an agentic cybersecurity platform running continuous vulnerability scanning and patching through coordinated agents rather than generating alerts for human triage. Both are high-volume, low-novelty work, which is exactly the profile these models were built for.

• Where the benchmark story does not match the marketing. On SWE Bench Pro, MAI Thinking scores around 53% against roughly 69% for the current Anthropic flagship on independent leaderboards. Microsoft's own AI chief has acknowledged a lead measured in months rather than weeks. Ray's point for buyers: MAI benchmark figures come from Microsoft's internal technical reports, and should be treated as vendor claims until independently validated.

• The distribution advantage that does not show up in any benchmark table. 70% of the Fortune 500 using Microsoft 365 Copilot, more than half a million companies in the Microsoft AI Cloud Partner Program, and 30 million paying Copilot seats, up from 20 million the prior quarter, plus more than 20 million developers and over 140,000 organizations on GitHub Copilot.

• The sovereignty play in Europe. A multi-billion dollar Mistral partnership funding European data center build out in exchange for prioritized capacity and distribution rights, which sidesteps the capital intensity of building EU capacity from scratch and answers the sovereignty question directly in a region where OpenAI and Anthropic are less established.

• Frontier Company, and the forward-deployed engineer story revisited. A $2.5 billion subsidiary staffed largely from existing engineers rather than new hires, with more than 6,000 forward-deployed engineers implementing inside client environments. Ray connects it back to the FDE economics the show covered previously, and Peter explains why pulling the large consultancies into the tent matters more than the headcount itself.

What CFOs and GTM leaders should take away:

• Route by workload, not by vendor loyalty. Frontier quality for the problems that need it, cheaper controlled models for Excel formulas, transcription, and support tickets. The difference in margin between those two decisions is the whole story.

• Treat self-reported model benchmarks as vendor claims. Until independent evaluation catches up, keep MAI models off your most technically demanding workloads.

• Don't skip the security diligence. Microsoft shipped real vulnerabilities this year, including a Copilot flaw that could have exposed customer files. The new Microsoft and NVIDIA open weights security association, which grew from 20 founding members to 125, is a good leading indicator but not a substitute for your own testing.

• Recognize what a vendor with equity in multiple labs is actually incented to do. When the provider profits regardless of which model you select, the steering pressure toward a specific model family drops.

• Weigh the balance sheet underneath the AI investment. Microsoft is funding all of this from a core business that grew 18% and saw net income up 31%, a cushion the pure-play labs do not have.

Ray's read: never bet against Microsoft, and the combination of a large profitable core business plus existing enterprise access is a genuine advantage. Peter's read: this is a price-performance play rather than a capability play, and whether it is a durable moat or a stopgap until the cost curve moves again remains an open question.

For the full analysis behind this week's big story, subscribe to the AI to ROI newsletter at ai2roi.substack.com.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

When Kimi K3 landed, the headlines said Chinese open-weight models had caught the American labs, and the AI trade sold off from chipmakers to the labs themselves. Nobody stopped to ask whether cheaper and more profitable mean the same thing.

In this week's AI to ROI Big Story, Ray Rike and Peter Buchanan run the actual business math on both sides of the fight and find that neither side has the balance sheet to fight a sustained price war.

The setup is stark. Anthropic is projecting its first-ever quarterly operating profit of roughly $559 million in Q2, with an annualized revenue run rate near $47 billion, up from $9 billion at the end of last year. That profit disappears the moment it tries to match open-weight pricing. Gross margin is running around 40%, roughly 10 points below the internal forecast, against more than $350 billion in data center commitments coming due over the next three to five years. OpenAI's picture is even thinner: roughly $30 billion in ARR, a projected $14 billion operating loss this year, data center commitments approaching $1 trillion, and an advertising business off to a slow start that needs to reach $100 billion by the end of the decade to close the gap.

What the episode covers:

• Why the Chinese open weight labs are not the subsidized price killers the coverage assumed, with Z.ai's gross margin falling from 41% to roughly 15%, DeepSeek near break-even on about $500 million of revenue and already back in market after a $7 billion round, and Moonshot and MiniMax raising at rising valuations rather than running toward profit
• The open weight versus open source distinction that changes the entire economic model, since every new customer requires more chips, power, and data center capacity, and Z.ai's own numbers show roughly 49 to 50% gross margin on customer hosted deployments versus about 19% when they host and serve via API
• The price war math itself: frontier models cost roughly $6 to $8 per million output tokens to serve, Anthropic's $25 per million on Opus produces about a 70% gross margin, and repricing down to the $4 to $6 range where Meta's Muse Spark sits flips that margin from positive 70% to negative 65%
• Why DeepSeek cut prices on a low-end model and then, two weeks later, told customers to prepare for substantial increases across the line, particularly on API access
• Google as the structural outlier, with 83% growth in its cloud and AI segment, Gemini embedded across fifteen products with more than a billion users each, and the Apple Siri deal extending reach toward two billion devices
• Where Kimi K3 actually fits, including the caveats nobody is pricing in: weights released only last week, no published large-scale production deployments, two to three times the token consumption on complex tasks, and infrastructure requirements around a 72 GPU rack that costs millions to install and millions a year to run
• The geopolitical wildcard, with Washington weighing sanctions or outright bans on Chinese open-weight models and distillation-related IP exposure still unresolved

The metric that resets the argument: cost per completed task

Price per token is the easiest unit to measure and the wrong one to buy on. Ray and Peter walk through a frontier lab evaluation that assumed a fully loaded remediation cost of $17 per failed attempt, roughly 10 minutes of a human operator's time. Claude Opus completed the task about 90% of the time at roughly $2.56. Meta's Muse Spark, priced at a quarter of Opus on tokens, succeeded 75% of the time and landed above $5.80 per completed task. The list price was 75% lower, and the delivered cost was more than double. A fifteen-point reliability gap did all the work.

The formula Ray offers turns a squishy quality debate into something a CFO can actually evaluate: cost per attempt, plus failure probability times fully loaded remediation cost, divided by success rate.

What CFOs and GTM leaders should take away:

• Build cost per completed task into vendor evaluation and make vendors compete on that number rather than on a token price list
• Segment AI workloads by what a failure actually costs, since a wrong answer in a regulated process or a customer support interaction carries a very different price than the model delta suggests, and standardizing on one model across both workload types to save on price is the common mistake
• Treat orchestration as a cost lever, not just the model choice, since Cursor's internal testing found coordination across multiple models delivered comparable code quality at a fraction of the cost of a single large model
• Run your own evaluations instead of trusting public leaderboards, since LMSYS Chatbot Arena measures human preference rather than task completion and says nothing about your workload
• Watch the switching trap, because a cheap model that attracts heavy traffic today can reprice two or three times higher in a quarter once the vendor needs margin
• Do not overreact to the open weight scare by standardizing on the cheapest option, since compute, power, and people costs are accelerating, and vendor durability still belongs in the evaluation

Ray's read: on real-world performance and total cost of ownership, the closed-weight labs still hold the advantage, their prices keep coming down, and they have every incentive to avoid a price war. Peter's read: open weight economics favor whoever hosts the model more than whoever built it, which makes the hyperscalers the quiet winners regardless of how this resolves.

For the full analysis behind this week's big story, subscribe to the AI to ROI newsletter at: ai2roi.substack.com

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

For three years, the AI infrastructure story has been about chips, power, and capital. In 2026, a fourth variable arrived that the hyperscalers were not prepared for: organized, well-funded, and increasingly successful local opposition to data center development.

Ray and Peter walk through the numbers behind the fight. Data Center Watch tracking shows at least $130 billion in US projects blocked or delayed in the first quarter of this year alone, with Carbon Direct putting the cumulative figure at $170 billion since 2024. Set against roughly $725 billion in projected 2026 capex across Alphabet, Amazon, Meta, and Microsoft, the question for operators is whether this is a genuine constraint or noise around the edges of an inevitable buildout.

The bigger shift is who carries the risk. With digital infrastructure funds raising $157.6 billion in 2025 according to PitchBook, and private capital taking positions like Blue Owl's 80 percent interest in Meta's $27 billion Hyperion campus, a permitting fight in a single county is now underwriting risk for pension funds and insurers nationwide. Delay no longer just moves a launch date. It extends the payback period and compresses return on invested capital.

In this episode:

• The scale of the buildout: 12 gigawatts of national compute capacity today, with the industry targeting 60 gigawatts by the end of the decade

• Who is writing the checks: OpenAI's $1.1 trillion in total infrastructure commitments, Anthropic's roughly $350 billion in compute commitments, and the rise of third-party capital as the load-bearing wall

• Why projects slip: 75 percent of capacity under construction already pre-leased, 1.6 percent vacancy, five-year grid interconnection backlogs, and 3-5 year transformer lead times

• What is driving residents into council meetings: $29.4 billion in added PJM customer costs, utility bills up as much as 267 percent in some markets, Google's water consumption up 34 percent year over year, and a Gallup finding that 71 percent of Americans oppose a data center in their own community, a higher share than opposes living near a nuclear plant

• How communities win: 833 active opposition groups across 49 states, more than 300 municipal bans and moratoriums since 2023, and Monterey Park's permanent ban approved with 88.34 percent of the vote

• Fifty different rulebooks: New York's statewide permitting pause, Ohio's 85 percent capacity payment requirement, Virginia's energy consumption tax, and West Virginia moving in the opposite direction

• The three playbooks that get projects built: Meta's community investment model in Louisiana, the legal route in Michigan, and Microsoft's quieter university and water reuse partnership in Wisconsin

The operator takeaway:

The buildout is getting slower and more negotiated, not stopped. For anyone modeling enterprise AI unit economics, the bottleneck has moved from chips to concrete, copper, and zoning boards. Cost models built on an assumption of falling compute prices need a second look this quarter.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

On this week's AI to ROI podcast, Ray Rike is joined by Stephan Bajaio, co-founder and CEO of VibeLogic and a former co-founder of Conductor, the enterprise SEO platform he helped build over 14 years, culminating in a WeWork acquisition, a management buyback, and a valuation approaching half a billion dollars. With 25 years spanning e-commerce, Yahoo, Time Inc., and a stint as a software CMO, Stephan brings an operator's view of what has actually changed in the buyer journey and what has not.

The conversation centers on a hard number. When Stephan asked the marketing team at a $3 billion company what share of their web traffic was being driven by LLMs, the estimates ranged from 20 percent to 50 percent. The measured answer was 1 percent, with a high conversion rate on that small base. That gap between perceived and measured contribution is the core problem for any executive being asked to fund an AEO, GEO, or AIO program this year.

Stephan introduces Web Presence Intelligence, a supply-and-demand framing designed for executive conversations rather than channel specialists. Demand is what goes into the search bar or the prompt. Supply is everything that comes back, including publishers, Reddit threads, affiliates, partners, and competitors. The strategic question becomes whether you are influencing where your buyer's opinion is formed, before you decide where to place the bets.

Ray pushes back directly, arguing that understanding where models source citations should now outrank owned web properties. Stephan holds his position, and the exchange gets to the heart of the allocation decision facing CMOs and CFOs.

In this episode:

• Web Presence Intelligence defined, and why supply and demand is the right executive language for a channel conversation

• The four P's of owned content: point, paragraph, page, path, and the audit finding that most sites never actually name the problem they solve in the customer's own words

• Why owned assets matter more when the interpretation layer keeps changing, illustrated by a professional who lost a decade of LinkedIn equity overnight

• Attribution reframed as a consequence of measurement rather than a measurement itself, and how to work backward from the sale to the second and third best proxies

• The baseline requirement, or what Stephan calls the before photo, and why AI deployed against a process you do not already understand produces outcomes you cannot judge

• Control groups in practice, including a market share test on terminology that showed one enterprise software company exactly what its brand guidelines were costing it in visibility

• Why the gold rush money went to the people selling pans, and where the actual near-term return sits: auditing your own sales calls, renewal conversations, and customer logs

• From funnel to hourglass, and why channel ownership models are breaking down as generalists get more capable with AI

• Augment, automate, or build net new: three different projects that require three different budgeting methods and three different attribution models

The operator takeaway: control what you can control, establish a baseline before you fund the initiative, and structure AI marketing investments so the return can be defended against assets you own rather than placements you do not own.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

AI spending continues to accelerate, but the ROI story has not kept pace. This week, Ray Rike and Peter Buchanan dig into the proposed fix that has the market talking: turning frontier AI labs into professional services firms through the forward-deployed engineer (FDE) model.

The role is not new. Palantir built the FDE function in 2005 to embed technical teams directly within customer environments, and by 2016, it had more FDEs than software engineers. What is new is the capital. Five ventures from the major AI labs and top three hyperscalers have committed more than $10 billion, betting that the bottleneck to enterprise AI value is not the model; it is getting that model wired into a customer's data, processes, workflows, and compliance requirements.

Ray and Peter connect this moment back to the ERP era, when SAP and Oracle needed four to five dollars of services for every dollar of software, and explain why the same people, process, and services reality is playing out again with agentic AI.

What the episode covers:

• Why OpenAI's $4 billion DeployCo, with a guaranteed 17.5% investor return, is the most aggressive and most financially puzzling bet of the five
• How Anthropic (Ode), AWS, Microsoft, and Google each took a different structural path, from joint ventures to capital-light partner ecosystem plays
• Why the MIT 95% pilot failure stat and McKinsey's finding that two-thirds of organizations have not started scaling AI make this a real problem, not a fringe one
• How incumbent consulting firms are playing every side at once to protect their AI practices
• Alex Karp's argument that the AI industry broke its own business model, and the irony of him making it

What CFOs and GTM leaders should take away:

• Ask any FDE partner for two or three production use cases with measurable outcomes before expanding scope
• Start narrow with one win, not four or five simultaneous projects
• Price for outcomes up front, before mid-project renegotiation
• Model the ongoing maintenance cost, since 20 to 40% of the initial investment often goes to keeping it running, and an outside FDE team at $300 to $800 an hour is an expensive long-term maintenance line item
• Weigh the neutrality trade-off honestly, since lab-backed ventures get you the deepest model roadmap access but also the deepest lock-in
• Lawyer up on data privacy and IP, because these engagements can feed your proprietary workflows back into the next model release

Ray's read: independent consulting firms are the likely long-term winners, with hyperscalers close behind. The AI labs are, in Peter's words, still leaving the Shire without their full posse ready to go.

Listen to the full breakdown, and for the deeper analysis behind each big story, subscribe to the AI to ROI newsletter at ai2roi.substack.com.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Twelve months ago, US frontier models controlled roughly 70 percent of AI traffic. Today, Chinese open-weight providers, led by DeepSeek, Z.ai, Moonshot, and Minimax, account for 45 to 61 percent of top-tier model traffic on OpenRouter, with DeepSeek alone processing more tokens than Google, Anthropic, or OpenAI individually.

In this Big Story edition, Ray Rike and Peter Buchanan unpack how this shift happened, how US labs and regulators are responding, and three scenarios for how the closed versus open weight competition plays out for enterprise AI buyers.

Key topics discussed:

• The pricing collapse driving enterprise migration. DeepSeek made a 75 percent price cut permanent in May, bringing its V4 Pro model to a fraction of a cent per million tokens versus $2.50 per million for GPT 5. Minimax delivers GPT 5.5 class coding performance at 5 to 10 percent of the cost. This is why Uber, Microsoft, and Walmart are now implementing formal usage governance on frontier models rather than treating cost control as temporary.

• The Mythos and Fable shutdown as a trust event. The 18-day suspension of Anthropic's top models over export control concerns spooked global enterprise buyers who realized mission-critical workloads could be cut off without warning. This single event accelerated the adoption of open-weight alternatives and pushed allied governments to invest in sovereign AI capacity.

• Distillation attacks and the IP leakage problem. Anthropic accused Alibaba's Qwen lab of running a large-scale adversarial distillation campaign, using tens of thousands of accounts and tens of millions of exchanges to extract agentic reasoning capability from Claude. This reframes the security conversation from model safety to unauthorized technology transfer, which is a distinct and arguably bigger risk for any enterprise relying on proprietary model capability as a moat.

• Cybersecurity parity is closing faster than expected. Multiple Asian labs, including Z.ai's GLM 5.2, Beijing based 360 Security, and Japan's Sakana AI, now claim benchmark performance approaching Anthropic's Mythos model on vulnerability detection and both offensive and defensive cyber tasks, often at significantly lower compute cost. This weakens the safety and capability gap argument that has justified restricting access to frontier models.

• Real deployments have moved from theory to production. Coinbase cut AI spend in half after migrating to Z.ai and Moonshot's Kimi models, even as token usage grew. Cursor shipped a coding tool built on Kimi, with a Grok-based version reportedly imminent. Andreessen Horowitz estimates 80 percent of its portfolio companies already use open-weight models in production AI products.

• Three scenarios for how this settles, and why the decision belongs at the board level. The hosts outline bifurcation (premium closed models for regulated use cases, open weight for commodity workloads), export control entrenchment (Washington treats the Fable ban as a template rather than a one-off), and capability convergence (the rationale for unilateral bans erodes as the performance gap closes). All three are already visible simultaneously, which means enterprise AI architecture decisions, including primary and backup model orchestration, are becoming strategic decisions that belong with the CEO and board, not just the technical team.

Why does this podcast episode matter for enterprise executives selecting models, especially for agentic AI deployments?

The vendor you choose today may not be the vendor you can use tomorrow, for reasons that have nothing to do with model quality. Regulatory risk, geopolitical exposure, and pricing volatility are now first-order variables in model selection, alongside capability and cost. Any agentic AI architecture built on a single model provider carries concentration risk that didn't exist a year ago. Building orchestration flexibility with primary and backup models, and understanding the true economics behind token pricing, is quickly becoming a board-level governance question rather than a procurement detail.

For the metrics and framework to instrument and report ROI on your AI investment, get the Big Book of AI Metrics at benchmarkit.ai under Media.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

On June 12th, the US Department of Commerce ordered Anthropic to suspend global access to Fable 5 and Mythos 5, its most powerful frontier models, with no advance warning and criminal penalties attached.

In this week's Big Story episode, Ray Rike and Peter Buchanan walk through the first use of export controls to take a deployed frontier model offline, why it backfired for security rather than strengthening it, and what a coherent federal AI regulatory framework would actually need to look like.

• The shutdown mechanics. Because Anthropic could not verify user nationality in real time, the directive knocked out access for every user globally, including Anthropic's own non US employees and more than 150 companies across 15 countries running critical infrastructure workloads.

• Three reaction tracks. Industry, the developer community, and allied governments each responded differently. OpenAI pushed back on talent restrictions while its legal team blocked coordination with Anthropic on antitrust grounds, and a coalition of 150 cybersecurity leaders published an open letter asking not for less regulation but for a transparent, science based process.

• The regulatory vacuum, in five forces. Ray and Peter unpack the drivers behind what they call a hair on fire crisis: capability jumps outpacing any statutory framework, a state level policy tsunami of over 1,500 AI bills across 45 states, data center community opposition, the unresolved Anthropic and DOD dispute, and a bipartisan Congressional letter questioning why comparable models were treated differently.

• A six part framework. The episode lays out proposed solutions modeled on existing regulatory precedent: global market access agreements similar to military sales processes, mandatory pre release testing run by NIST modeled on FAA certification, clear federal versus state jurisdictional lines modeled on pharmaceutical regulation, expanded child safety authority, FERC fast tracking for data center grid access, and coordinated environmental standards.

• What enterprise leaders should do now. Ray closes with practical guidance: review AI vendor contracts for shutdown protection since force majeure clauses were never written with export controls in mind, build fallback infrastructure for critical AI dependent workflows, delay rushing into brand new model releases, and automate tracking of the fast growing state regulatory landscape.

Full details are in the AI to ROI newsletter at ai2roi.substack.com. Subscribe, and consider reaching out to your representatives on this one.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.

Most companies treat AI as a layer they add on top of how work already gets done. Dan Zhang, Chief Business Officer and CFO at ClickUp, argues that it is exactly backward.

In this episode, Ray sits down with Dan to unpack the "100x Org," ClickUp's framework for rebuilding the business around AI rather than sprinkling tools and tokens on top of a human-driven workflow, and why that distinction determines whether an AI initiative shows up as activity or as income statement impact.

The conversation covers:

• Why ClickUp expanded the CFO's charter to own AI transformation end to end, after both a top-down mandate and a bottoms-up experimentation push failed to produce results that made it into production

• The "jobs to be done" framework Dan uses to separate primary work that actually moves the business from secondary work that just generates busy AI activity, and why most companies have a work redesign problem before they have an AI problem

• Dan's psychological test for AI ROI (would you pay for it with your own money) and why ARR per headcount is the right metric, but a lagging one that plays out over years, not weeks

• How ClickUp instrumented daily, not monthly, visibility into AI cost and token consumption, and why over half of enterprise companies in Ray's own research have blown their AI budget by 25% or more

• Why the real anxiety CFOs have about AI ROI isn't the return, it's confidence in cost management, and how a governance layer turns that anxiety into a guardrail instead of a monthly surprise

• Dan's rapid-fire advice on who should own AI ROI measurement, the two-to-three variables every CFO needs in place, and how early and mid-career professionals protect their relevance by owning the question, not just the answer

If your organization has AI activity but can't yet point to AI impact, or your CFO isn't sure whether AI spend is a competitive advantage or a leak, this episode gives you the operating model and the financial discipline to tell the difference.

See Privacy Policy at https://art19.com/privacy and California Privacy Notice at https://art19.com/privacy#do-not-sell-my-info.