Deep Dive

In early June 2026, four of the fiercest rivals in artificial intelligence — the heads of OpenAI, Anthropic, Google DeepMind, and Microsoft's AI division — signed the same letter to Congress. So did the companies that manufacture synthetic DNA. Their shared ask: make it mandatory to screen every commercial DNA order for dangerous sequences, to keep AI from helping someone build a bioweapon.

It sounds like the responsible thing. But there's a strange detail buried underneath. Eight months earlier, Microsoft's own chief scientist had published a study showing that AI can redesign known toxins to slip past that exact screening. The fix the labs are lobbying for is one their own technology already beats.

So we did the boring thing. We read the letter, the bill it points to, the studies on both sides, and followed the one question that cuts through it: who actually pays?

The answer is the whole story. The rule binds one industry — the DNA-synthesis manufacturers — and asks nothing of the AI labs that named themselves the new risk. No limit on their models. No release gate. Two of the DNA companies the rule would regulate signed the letter themselves, because a mandate locks in the screening they already do and raises the cost on every cheaper rival. The regulated industry asked to be regulated.

And the fix may not even work. The screening matches orders against a list of known dangerous sequences — but AI can generate tens of thousands of new variants the list has never seen, and benchtop DNA printers are starting to remove the ordering step entirely.

None of that means the threat is fake. The danger is real and climbing; screening DNA is genuinely good policy; and there's an honest case that getting rivals to agree on shared, enforceable rules is the only way out of a race to the bottom. Both readings can be true at once — the danger real, and the fix convenient.

This is the political economy the headlines flattened into "AI labs warn Congress." Deep Dive goes beneath the headline: who the rule actually binds, why the people asking for oversight chose a rule that costs them nothing, and the oldest pattern in politics — when someone powerful asks to be regulated, check whose name is on the bill.

RELATED EPISODES

Claude Mythos: The AI That Breaks Everything — the model whose disclosure triggered the same regulate-me move this letter repeats, one rung wider

How Anthropic Actually Makes Money — why hobbling your own model just hands the lead to a rival, and why the labs reach for supply-chain rules instead

CHAPTERS

00:00 The contradiction in the letter

00:30 Four rivals, one ask

01:08 Is the threat even real?

02:44 The one wall that mattered

04:52 Follow the cost, not the noise

07:09 Why the regulated asked to be regulated

08:13 How DNA screening actually works

10:14 Where the screen leaks — twice

12:12 Cynical, sincere, or both?

14:31 The pattern — whose name is on the bill

15:12 Three predictions

SOURCES

The June 2026 open letter to Congress (OpenAI, Anthropic, Google DeepMind, Microsoft AI + the DNA-synthesis manufacturers)

Wittmann, Horvitz et al., "Strengthening nucleic acid biosecurity screening against generative protein design tools," Science (Oct 2025)

The Virology Capabilities Test — frontier models vs. expert virologists

RAND 2024 red-team study on LLM operational uplift for bioweapons

Anthropic's ASL-3 uplift trial (novice acquisition-plan study)

S. 3741, the Biosecurity Modernization and Innovation Act (Cotton / Klobuchar)

SecureDNA — cryptographic DNA-order screening (Esvelt, Yao)

This spring, one fight became the biggest story on the internet. The version everyone shared was clean: an evil company robbed an 83-year-old man of his life's work — a 780-set Star Wars LEGO collection — and a heroic YouTuber named Reckless Ben exposed them. Hundreds of millions of views. A villain, a hero, a crowdfund in the hundreds of thousands. Everyone agreed on what happened.

So we did the boring thing. We read both lawsuits, the franchise contract, and the live court docket — and watched what five different lawyers concluded about it.

Almost nobody got it right. Including some of the lawyers.

The old man has a real grievance: when corporate seized the store over the local owners' unpaid fees, the consignment paperwork — and the company's own franchise contract — says the collection was never theirs to take. But an obscure rule buried in commercial law hands consigned goods to the store's creditor anyway, so he can be legally right and still never see a cent. The famous $200,000 figure? It traces back to the store's own marketing post, not an appraisal. And his claim is stuck in a "civil dead zone": too big for small-claims court, too small for any lawyer to touch.

Going viral was the only escape hatch left — and the data on that move is brutal. Most online firestorms change nothing; companies fold to stop the press, not because sales drop, and a niche LEGO chain is the hardest target there is. It "worked" here only because the crowd paid. Not the company. Every other time a system like this broke, someone with a duty was forced to pay. This time, nobody was.

Then it gets stranger. The champion who got this case in front of millions may be the reason it falls apart. The deeper you read, the more the hero's own conduct buries a case that was winnable — and the more the cartoon villain turns into something colder and, in some ways, worse. The police who raided Ben's rental hunting for stolen LEGO (and found none) probably broke the Fourth Amendment — and will almost certainly keep their immunity anyway. Even the people the mob hammered hardest may turn out to be victims too.

Four parties. Real harm spread across all of them. No single, simple villain — which is exactly what lets the side with the deepest pockets wait everyone else out. There's one live trial left, on June 17. And whatever the verdict, the collection is already gone.

This is the slow, tangled version sitting in the court file nobody clicks on — the one the viral story flattened. Deep Dive goes beneath the headline on the stories that matter: the institutions, the fine print, and who actually pays when a system breaks.

RELATED EPISODES

How the Statute of Limitations Killed Musk v. Altman — a strong claim that never reached the merits because an obscure rule got there first

The AI Layoff Gap: What CEOs Tell Investors vs. What They Tell the State — one story for the public, the opposite on the record

CHAPTERS

00:00 The raid that found nothing

00:20 A life's work, on consignment

01:43 The internet finds its hero

02:23 Almost everyone is wrong

03:43 The legal trap with his name on it

05:28 The $200k myth and the dead zone

08:24 Does going viral actually work?

10:46 The hero's problem

14:54 The police, the warrant, the bodycam

17:08 The company, and who's really a victim

20:49 Why digging in is the expensive move

23:08 June 17 — and what won't change

SOURCES

Both lawsuits + the live docket (Gormans v. corporate; corporate's racketeering suit vs. Ben)

The Bricks & Minifigs franchise/consignment contract

Leaked police bodycam (the department's own release) + Ben's synced clips

Five lawyers' analyses — LegalEagle on title; a civil-rights attorney on the warrant (Franks, 1978)

404 Media ("the most disastrous brand response one could imagine")

Bricks & Minifigs' June 2026 statement (restitution + drop-suit offer)

On the first day of GitHub Copilot's new billing, a tech journalist checked his account and found Copilot projecting him a $180 bill — and he got off easy. Other developers watched a $10 plan balloon into thousands. Same subscription, same price. The meter did it.

Microsoft launched seven of its own AI models the same week, and the headlines all said Microsoft and OpenAI were breaking up. That read is mostly wrong. The real story is an accounting problem that breaks the whole business model of selling AI.

Normal software is almost pure profit — one more user costs almost nothing. AI breaks that: every answer burns real compute, so the more someone uses it, the more it costs the company running it. The Wall Street Journal reported in 2023 that Copilot was losing ~$20 a month per user — up to $80 on power users — on a $10 plan.

Then the smoking gun, 24 hours apart. June 1st: Copilot goes metered, the free fallback removed. June 2nd: Microsoft drops its own model into the picker — and its own rate card prices GPT-5.5 at $30 per million words of output versus $4.50 for Microsoft's model. About 6.7 times cheaper, a price Microsoft set for itself, so routing a heavy user to it collapses the cost on that user.

A startup already lived this arc. Cursor resold Anthropic's model at a reported negative 30% gross margin — a $200 plan against ~$5,000 of compute — then shipped its own model and turned positive. Same law that pushed Apple off Intel onto the M1.

But this isn't a divorce. Microsoft's first move against the trap wasn't a model — it was the meter, which moved the loss onto whoever pays the bill. And it kept OpenAI licensed through 2032. The honest counter holds too: $4.50 only proves what Microsoft charges itself, and the "AI margins are turning the corner" story is mostly compute margin — all-in it was still about 33%. So MAI is a bargaining chip, not an exit.

The next time a price jumps on a tool you depend on, ask whether it's really greed, or just the trap.

Deep Dive goes beneath the headline — AI, finance, security, and the systems running underneath.

RELATED EPISODES

How Anthropic Actually Makes Money — the same per-token margin math, from the lab''s side.

Will AI Replace Software Engineers? What the Data Actually Says — Cursor''s rise; here, its compute bill.

The Real Cost of AI: Who's Actually Paying for the AI Build-Out — there, your power bill; here, your Copilot bill.

The AI Chip War: Why Everyone's Watching the Wrong Fight — the same build-your-own logic, one layer down.

CHAPTERS

00:00 Quick financial disclaimer

00:08 The day-one bill shock

01:05 Not a breakup — an accounting problem

01:28 Why AI breaks the SaaS model

03:12 June 1st: Copilot becomes a meter

04:10 June 2nd: Microsoft ships its own model

05:14 Cursor already lived this trap

06:24 The law: a big buyer builds its own

08:08 Why it's still not a divorce

09:59 The honest ledger, and the calls

SOURCES

Copilot lost ~$20/user/mo on a $10 plan — AI has no SaaS economies of scale (WSJ 2023)

Inference price falling ~50x/yr while usage explodes; ~24x token growth by 2030 (Epoch AI; Goldman)

June 1 Copilot goes metered (free fallback removed); June 2 MAI-Code-1-Flash ships into the picker (GitHub)

Microsoft's rate card — GPT-5.5 $30 vs MAI $4.50 per 1M output tokens (~6.7x cheaper); 4.7M paid subscribers (GitHub)

Dev bills jumped $29→~$750 and $50→~$3,000 (TechCrunch)

Cursor ran ~-30% margin reselling Anthropic ($200 plan ≈ $5,000 compute), then shipped its own model (Newcomer)

OpenAI licensed through 2032, $38B revenue-share cap; GitHub VP says Copilot >$100M ARR; all-in margin ≈ 33% (Microsoft; The Information)

———

This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.

Everyone says China will win the AI race on cheap power — twice America's electricity, six times the buildout, data centers stood up in six months. That framing is half-wrong. China has the watts. Its AI data centers are sitting empty.

Up to 80% of the AI capacity China built sits idle. Flagships run at 20 to 30% against a 60% government target. Rents for a top-end GPU server fell by more than half in a year — a glut, not a shortage.

One number ends the watts argument: AI is barely 1.7% of China's electricity. Under two percent of the grid can't be the constraint. Chips are. Price out a frontier data center and electricity is about 7% of the bill, chips and servers about 60% — and in chips, America leads roughly 99 to 1, from the same Stanford report the bulls quote for power.

So why is this still a contest? Watts and wafers aren't two races — they're substitutes. China can't freely buy Nvidia's efficient chips, so it runs Huawei silicon that matches them at 4.1 times the power. Cheap power isn't China's prize; it's its workaround. In three provinces, Beijing cut the AI power price to about half the coastal rate — but only if you drop Nvidia for domestic chips.

Now turn the mirror. America has the chips; what it's running out of is power — and unlike China's, this bill has a name on it. John Steinbach in Manassas, Virginia opened a $281 January bill, up from about a hundred. In a December auction, data centers drove $6.5 billion of the mid-Atlantic's costs — $6.2 billion of it for centers not yet built. The bottleneck isn't generation; it's a grid queue measured in years, through transformers and steel America barely makes.

The honest version: the capacity gap is real (China added 429 gigawatts in 2024 to America's 51), Chinese models closed to near-parity on a fraction of the compute, and a captive-solar bet could still bring America's wall down. A live race, not a settled one — and it goes to whoever tears their wall down first.

Deep Dive goes beneath the headline on the stories that matter — AI, finance, security, and the systems running underneath.

RELATED EPISODES

The AI Chip War: Why Everyone's Watching the Wrong Fight — the bottleneck slid logic to memory; here it lands on electricity.

NVIDIA Just Forecast $91 Billion Without China — the same embargo, from the company whose silicon China can''t buy.

Why Wall Street Is Betting Billions on Nuclear Power for AI — the US power wall, priced into the megacaps.

How Microsoft Is Restarting Three Mile Island for AI — the household-bill story, a year later and steeper.

CHAPTERS

00:00 Cold open — the empty data centers paradox

01:17 Start with the paradox: the buildings are empty

02:13 AI is under 2% of the grid — so what binds?

03:22 The real ceiling: chips, not watts (99 to 1)

05:35 The twist: cheap power as an embargo workaround

06:24 The subsidy weapon: half-price power for domestic chips

07:20 Turn the mirror: America hits the power wall

09:20 The real bottleneck: the grid and who it runs through

10:33 The strong version of the other side

12:59 The verdict and a dated prediction

SOURCES

China''s new AI compute up to ~80% idle; flagships 20-30% vs a 60% target; AI ≈ 1.7% of China''s electricity (MIT Technology Review; ASPI Strategist)

Energy is ~7% of a frontier data center''s cost vs ~60% for chips/servers (Epoch AI)

The compute gap is ~99:1, from the same Stanford AI Index the bull side cites for power (Stanford HAI AI Index 2025; CSIS)

Huawei''s CloudMatrix matches Nvidia at 4.1x the power; three provinces cut AI power to ~half coastal rates only for facilities dropping Nvidia for domestic chips (SemiAnalysis; FT)

Manassas bill hit $281 (from ~$100); PJM''s Dec 2025 auction = $6.5B from data centers, $6.2B for centers not yet built (Consumer Reports; Utility Dive)

China added 429 GW of power in 2024 vs the US''s 51 (OpenAI/EIA; RMI; Fortune)

Three companies make the memory that every AI chip on Earth depends on. They're sold out years in advance, their margins beat NVIDIA's, and in May 2026 all three bought a stake in their own biggest customer. The supplier became the owner.

This episode goes inside one component: HBM, the high-bandwidth memory bolted next to every AI accelerator — and the hardest chip on Earth to make. Two questions the headlines skip: why can only three firms make it, and what happens when the supplier owns the customer.

The answer to the first is arithmetic. An HBM chip is twelve to sixteen wafer-thin memory dies stacked vertically, and the package only works if every die works — so yield multiplies instead of averaging. At a 95% per-die yield, a sixteen-high stack collapses to about 44%. A new entrant doesn't need to be good; it needs to be near-perfect — and NVIDIA reportedly relaxed its own HBM4 spec to fit what the three best memory makers on Earth can actually yield.

Three firms making essentially all of it, sold out years forward, is an oligopoly — and SK hynix posted a 72% operating margin in early 2026, an all-time high that outshines the company buying its chips. Memory is now 63% of an AI chip's component cost, and roughly 90% of all HBM is made in one country: South Korea.

The move that anchors it all: on May 28, 2026, Micron, Samsung, and SK hynix all joined Anthropic's $65 billion round at a $965 billion valuation — the first time the entire memory oligopoly took equity in a single AI lab.

Which leaves the real question: a durable moat, or the top of a violent cycle? Memory has always been brutally boom-bust, the bears cluster on a 2027 inflection, and all three crossed a trillion dollars in market value within 21 days of each other. The honest read holds both — a real moat, with a real expiration date.

Deep Dive goes beneath the headline on the stories that matter — AI, finance, security, and the systems running underneath.

RELATED EPISODES

The AI Chip War: Why Everyone's Watching the Wrong Fight — named memory as the binding AI bottleneck; this goes inside it.

NVIDIA Just Forecast $91 Billion Without China — memory eating the AI chip's bill of materials, now 63%.

How LLM Inference Actually Works — the memory wall, and the wafer-scale SRAM bet that routes around it.

Why Wall Street Is Betting Billions on Nuclear Power for AI — the same 'AI needs X' thesis priced into trillion-dollar firms.

CHAPTERS

00:00 Three firms, sold out, and they bought their customer

01:04 Why only three: the yield math

05:14 Why China can't just become the fourth

06:21 An oligopoly that's sold out for years

08:11 Margins that outshine NVIDIA

10:18 Memory is 63% of the chip — who pays for it

13:21 The Korea chokehold

15:59 The supplier becomes the owner

17:58 A durable moat, or the top of a cycle?

20:34 Can anyone route around HBM?

SOURCES

HBM stack-yield is multiplicative — at 95% per die, a 16-high stack ≈ 44%; one bad die kills it (nomadsemi; SemiEngineering)

Three firms ≈ 97% of HBM, ~90% in Korea; NVIDIA reportedly relaxed its HBM4 spec to fit yields (AI Frontiers; TrendForce)

Samsung HBM3E ~18 months late; dismantled its HBM team in 2019 (Tom's Hardware; SamMobile)

SK hynix Q1 2026: 72% operating margin, sold out ~3 years; Micron 75% gross margin (CNBC; Yahoo Finance)

HBM = 63% of AI-chip component cost; DDR5 kit ~$80–100 → $350–600; Dec 2024 BIS HBM rule (Epoch AI; BIS/CSIS)

Anthropic $65B round at $965B — all three memory makers join, May 28 2026; Intel customer-equity counter-precedent (TechCrunch; Bloomberg)

———

This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.

On May 27, 2026, Robinhood did something no major U.S. broker had done: it handed autonomous AI agents the keys to live stock trading. Paste one link into ChatGPT or Claude, fund an account, and the AI can buy and sell for you — and you can switch off the part where it asks first. It's rolling out to all 27.4 million customers.

Then there's the fine print. Robinhood's terms say it "does not control, supervise, monitor, recommend, or audit" the agents and is "not responsible for losses." Every take so far stops there: you're on your own, be careful.

But that's not the real story. The real story is whether that sentence even holds. A brokerage may not be able to waive its duty to supervise — FINRA Rule 3110 — and that rule was written for a human, not a chatbot you brought from outside. By law, you can't quietly delete a securities rule; Robinhood's own customer agreement even concedes its liability limits don't touch your FINRA rights. The disclaimer is a bet, not a shield — and it gets tested the first time an agent blows up a real account.

We trace how it actually works (and why prompt injection makes it dangerous), then the part nobody's connecting: every previous time automated finance broke, someone with a legal duty paid. Knight Capital's bug cost the firm about $440 million in 45 minutes — and the firm ate it. The robo-advisors stayed fiduciaries; when Schwab's quietly skimmed, it paid about $187 million back to clients. Robinhood itself paid the largest fine in FINRA history in 2021 — the same year it froze GameStop trading and its CEO testified before Congress. Every time, a party with a duty paid. This time, that duty's been engineered out.

The cold-open twist: two weeks before launch, Robinhood's own chief legal officer — a former SEC commissioner — warned at FINRA's conference that third-party AI giving financial advice with no oversight is dangerous, and pitched a safer "walled garden, not scraping Reddit." Then his company shipped the open opposite.

We give the strongest case that this is fine (it's self-directed trading; the SEC just blessed "tool, not advice"), plus three dated predictions. The bigger question under all of it: can a regulated middleman hand your money to an autonomous AI — and walk away from the result? Whatever breaks here writes the rules for agentic commerce everywhere.

Deep Dive goes beneath the headline on the stories that matter — AI, finance, security, and the systems running underneath.

RELATED EPISODES

When AI Agents Go to Court — who's liable when the actor is an AI

How AI Agents Actually Work — the engine now trading your account

Why the Fed Summoned Wall Street's CEOs Over an AI Model — AI meets regulated finance

Will AI Replace Software Engineers? What the Data Actually Says — same hand-the-keys question, next domain

CHAPTERS

00:00 Disclaimer — none of this is investment advice

00:08 His own lawyer warned against it

00:43 Robinhood hands AI the keys to 27M accounts

01:45 How it actually works (and prompt injection)

04:31 "We're not responsible" — is that even allowed?

06:26 Who paid, every time automation broke before

09:01 Why Robinhood did it: the money

10:34 The case that it's fine

12:34 Three predictions

13:18 The verdict

SOURCES

Robinhood agentic-trading launch + terms (TechCrunch); Q1 PFOF $623M 10-Q

FINRA Rule 3110 + 2026 Oversight Report; Gallagher remarks, FINRA conf (May 2026)

SEC tool-not-advice staff letter (Apr 2026) — does NOT address AI agents

Precedents: Knight Capital; Flash Crash; Schwab robo Fair Fund; Robinhood FINRA

California AB 316; Air Canada ruling (Moffatt)

———

This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.

Sixteen senior developers used the best AI coding tools — Cursor, Copilot, Claude — on real tasks at real companies. They finished 19% slower than when they worked without AI. And they were sure the AI had sped them up — by about 20%. The researchers ran the experiment again in February 2026 to settle it, and the experiment itself collapsed: developers refused to do the tasks without the AI. The counterfactual is gone. The experiment to measure if AI helps engineers has lost its control group.

Meanwhile, the money is undeniably real. Cursor crossed $2 billion ARR. GitHub Copilot has 4.7 million paying seats. Claude Code reportedly hit $2.5 billion in its first year. Someone is paying, at scale, for something we cannot reliably measure. The wins on bounded tasks are real — Stripe runs 1,300 AI-written merges per week, Rakuten agents tackle problems inside 12.5-million-line codebases. But two-year telemetry across 22,000 developers shows the cost: code churn up 800%, bugs per developer up over 50%, deployments per week DOWN 11.7%. More code, generated faster, shipping slower.

The cut is landing entirely on the first rung. Entry-level engineering postings fell 60% from 2022 to 2024. Programmer employment for ages 22 to 25 is down 20%. India's four biggest IT firms added 3,910 net employees over a year — firms that used to hire 10,000 in a single quarter. Amazon's CEO said work that used to take 40 engineers now takes six. Stripe's leadership worries out loud what entry-level looks like in ten years. We're automating the junior work that makes a senior, with no plan for how to make seniors without those years.

The optimists have history. Compilers, offshoring, spreadsheets — each was supposed to end jobs, and didn't (there are four times as many accountants now). But three things are new: speed (toy to threat in three years), scope (it replaces ladders, not tasks), and the strangest — we can no longer measure what we're trading.

The verdict: no, AI is not replacing software engineers the way the headlines mean. But yes, it's already replacing the on-ramp. The danger was never the robot that codes — it's the missing rung on the ladder. We're sawing it off while flying blind.

One bold prediction: before the end of 2026, a big-name company that quietly stopped hiring junior engineers reverses course. The trigger won't be quality — engineering pipelines fail slower than that. It'll be one CFO deciding the math looks wrong, before the damage shows. Sixty percent confidence, not ninety.

RELATED EPISODES

How AI Agents Actually Work — same bounded-vs-fuzzy split

The AI Layoff Gap — macro layoff narrative; here it's the engineering cut

When AI Agents Go to Court — liability when an AI PR breaks production

The Loop Closed in the Sandbox — AI doing AI's own work, shipped to every engineer

CHAPTERS

00:00 19% slower (and they couldn't tell)

00:34 A productivity gain nobody can measure

01:10 The money is real

02:12 The METR study, and why the experiment broke

05:15 Maybe we're measuring the wrong thing

05:46 Why the code is getting worse

08:13 The benchmark scandal

08:55 The cut lands on juniors

12:21 The pipeline time bomb

13:50 What history says — and what's different

15:44 Denmark, and the verdict

16:33 The bold prediction

SOURCES

METR (Jul 2025 + Feb 2026): 19% slower with AI; experiment broke when devs refused to work without it

NBER (Feb 2026): 9-in-10 firms report no measurable AI impact

Goldman Sachs (Mar 2026): AI ~zero to US GDP; ~30% gain in narrow uses including software

Faros AI (22,000 devs, 2 yr): churn +800%, bugs +50%, deploys -11.7%

SWE-bench Verified ~80% → held-out SWE-bench Pro ~46% (contamination)

BLS / Stanford / NY Fed: entry postings -60%, ages 22-25 -20%, CS-grad underemployment >40%

Cursor ~$2B ARR, Copilot 4.7M seats, Claude Code ~$2.5B ARR (single-source)

Org cases: Shopify, Amazon 40→6 (Jassy 2025), Stripe 1,300/wk, Klarna rehire, India big-4 +3,910

Jevons history; Denmark NBER null; Anthropic CEO forecast

In March 2026, Waymo crossed half a million paid rides a week with nobody in the driver's seat — tenfold growth in two years. Two months later it pulled its cars from four cities for driving into floods. Both are true, and the gap between them is the whole story: there's no longer one self-driving car industry. There are two, and they share a name and almost nothing else.

One quietly became a real business. Waymo has driven 170 million miles with no one in the seat, and the reinsurer Swiss Re found it has 88% fewer property-damage and 92% fewer injury claims than humans. The real tell isn't the safety number, though — it's that the scorecard changed, from how often a safety driver grabs the wheel to paid rides per week, compounding. You can't fake half a million paid driverless rides. But the same month it crossed over, it hit the wall: everywhere it's cleared to drive adds up to about 1,400 square miles — the size of Rhode Island — with no service in a single snowy city.

Tesla makes the opposite bet — cameras only, drive anywhere — but it hasn't shipped: a thirty-car Austin pilot, most with a human aboard, and two crashes caused by the teleoperator, not the AI. A third player may be biggest of all: China's Apollo Go hit the same quarter-million-rides-a-week line the same month, on a car that costs a quarter as much. And does any of it make money? Two numbers circulate for a Waymo ride — $1.40 and $43 a mile — both correct: the next trip's cost, versus the whole company divided by its miles, with Google eating about $40 of every one.

The real bottleneck isn't red tape — it's transparency: US crash-reporting was relaxed exactly as the fleets scale. A jury hit Tesla with a $243 million Autopilot verdict — the first crack in the carmaker's wall, though it blamed the driver for two-thirds. So how close are we? Both answers are true: closer than the skeptics think, since a real paid driverless service exists and is growing — and further than the believers think, since it only works inside a sunny, mapped box. The robotaxi is no longer a promise. It's a product. It's just a product that, as of this month, still pulls over when it rains.

RELATED EPISODES

How AI Agents Actually Work — the same 'autonomy we want but can't fully trust' tension; a robotaxi is that tension at 40 mph (nobody in the seat, 70 people on call to advise)

The Humanoid Robot Race — the 'read the production filings, not the press releases' move; the robotaxi version is the paid-ride curve + the NHTSA crash filings

The AI Layoff Gap — there the narrative ran ahead of the data; self-driving is the mirror image (the leader's data finally caught up, the laggard's narrative still leads by a decade)

CHAPTERS

00:00 Two industries that share a name

01:08 Waymo — the one that crossed over

02:16 The safety data even the skeptic trusts

03:21 The wall — where it still breaks

06:14 Tesla's opposite bet

08:59 China's bet — Apollo Go

09:35 Does any of this make money?

12:03 The rules — and the crack the car falls through

13:28 Who pays when it crashes

14:12 The verdict — and the bets

SOURCES

Waymo — 170.7M rider-only miles, ~500K paid rides/week (March 2026, ~10x in two years)

Swiss Re reinsurance study: ~88% fewer property, ~92% fewer injury claims vs humans (+ peer-reviewed analysis)

Waymo May 2026: flood pauses (San Antonio, Atlanta), freeway-service pause, NHTSA recall

Tesla robotaxi — NHTSA crash disclosures (Austin pilot; two teleoperator-caused crashes)

Baidu Apollo Go — ~250K driverless rides/week in China (Oct 2025); ~1/4 of Waymo's vehicle cost

Waymo economics — Morgan Stanley ~$1.40/mi marginal vs ~$43/mi fully-loaded; Alphabet Other Bets -$2.1B on $411M (Q1 2026)

Cruise (~$9B, GM) + Argo (~$7B, Ford/VW) shutdowns — the AV capital graveyard

Benavides v. Tesla — $243M Autopilot verdict (2026); NHTSA crash-reporting relaxation; Goldman insurance projection

There's a free tool you can install with a single command. Point it at a downloaded AI model — Llama, Gemma, Qwen — and in about thirty minutes, on a used gaming card, it permanently removes the model's ability to refuse. It's called Heretic, it hit number one on GitHub, and it works because of a quiet discovery: a model's refusal isn't woven all through its mind — it's a single direction in the math, and the safety often runs only a few tokens deep.

But the tool isn't the real story. Heretic works on Llama, Gemma, Qwen, and Mistral because those labs shipped the cheap, removable kind of safety — a refusal layer that peels right off. Durable safety provably exists. One major lab filtered the dangerous knowledge out of its model before training, and a research team that wove safety in during pretraining watched it survive ten thousand attempts to strip it, where the bolt-on kind collapses in a few hundred. Most labs simply didn't pay for it.

We take the honest counter-case seriously. The "safe" models over-refuse — one benchmark caught the most cautious model rejecting ninety-nine percent of perfectly harmless questions — and most demand is ordinary: privacy, fiction, research. Has a stripped model actually caused real-world harm? Almost none on record; the scary names like WormGPT were a different method entirely. But that empty column is its own kind of warning — abliteration runs on the attacker's own machine, with no call home and nothing to log.

Then there's the law. The European Union built the world's most aggressive AI law, deciding danger by a single number: how much computing power went into training them. The small and mid-sized models Heretic targets sit below that line, so no one is ever required to safety-test them — and the person who strips the safety spends a few cents of compute, far too little to ever become the legally responsible owner. Enforcement goes live August 2nd, 2026, against a gap a one-command tool walks straight through.

The scandal, if there is one, is quieter than the headlines: the safety on most open models was built to be the kind that comes off, and the law written to catch that is watching the wrong number. The question was never whether open AI can be made safe. It's whether anyone selling it decides to.

RELATED EPISODES

Claude Mythos: The AI That Breaks Everything — predicted the open-source gating endgame ('it becomes an arms race'); this episode is the empirical confirmation

The Brand Survives the Arrests (ShinyHunters) — a removable control plus an industrialized exploit pipeline, the same security shape

The Mandate That Couldn't Be Met (Palo Alto CVE) — when the rule polices the wrong thing; the regulatory-gap parallel

CHAPTERS

00:00 The one-command tool — and how abliteration works

05:11 Who wants this — and the over-refusal trap

05:57 Does stripping the safety keep the model smart?

08:12 Why a release can't be recalled — and has it caused harm?

09:34 What the labs already know — and gpt-oss

12:42 The law that measures the wrong number

15:37 What a truly safe open model would take

16:37 The verdict

SOURCES

Heretic — open-source abliteration tool (#1 trending on GitHub, Nov 2025); creator comments via the Financial Times

Arditi et al., 'Refusal in Language Models Is Mediated by a Single Direction' (NeurIPS 2024)

OpenAI gpt-oss — open-weight release with CBRN-filtered pretraining + adversarial misuse evaluation (Aug 2025)

EleutherAI & UK AI Security Institute — 'Deep Ignorance': data-filtered durable safety that survived ~10,000 fine-tuning steps

Badllama — safety removed from a Llama model in ~1 minute on a single GPU, for pennies

EU AI Act — GPAI compute threshold (10^25 FLOP), open-source exemption, and penalty enforcement live August 2, 2026

OR-Bench (Cui et al.) — over-refusal benchmark; the most cautious model rejected ~99% of benign prompts

US NTIA (2024) — found insufficient evidence to restrict open model weights

On May 20th, 2026, SpaceX filed to become the largest public offering in history. Bankers are guiding toward roughly $1.75 trillion — about 94 times sales — and the share price on the cover of the filing is left blank. Here's the part that should stop you: last year SpaceX lost $4.9 billion, it carries a $41.3 billion accumulated deficit, and only one of its three businesses actually makes money.

We read the actual S-1. Starlink is the engine — $11.4 billion in revenue, $4.4 billion in operating income. The rocket business runs a small operating loss (after $3 billion a year on Starship), and the xAI division lost $6.4 billion on $3.2 billion of revenue. In 2024, SpaceX turned its first profit — $791 million. Then it bought Elon Musk's AI company, xAI, and once the results combine, that profit becomes a $4.9 billion loss. The rocket company is even filing as a software company — and 93% of the $28.5 trillion market it points to is the AI division losing the most.

Two things make the price work. First, SpaceX's largest disclosed AI customer is Anthropic, paying $1.25 billion a month to rent computing power — on a contract either side can cancel with 90 days' notice. Second, in the weeks before the filing, the index rules quietly changed: Nasdaq cut its waiting period, and the S&P opened a proposal to waive its profitability requirement for exactly these companies. If adopted, it would force trillions in index-fund money to buy SpaceX regardless of the numbers.

Then we value it honestly, in both directions. A research firm's sum-of-the-parts and Aswath Damodaran's independent model both land near $1.2 trillion — leaving a half-trillion-dollar gap to the asking price with no financial justification, only a story. The bull case is real too: Starlink is close to a monopoly, grew revenue about 50% last year at ~40% operating margins, and could scale its $4.4 billion profit toward $18 billion. But the base rate is sobering — about 9 of 10 comparable growth IPOs underperformed the market in their first year. SpaceX is an extraordinary company. The question is whether $1.75 trillion is an extraordinary price.

This episode is analysis and opinion, not investment advice. Do your own research.

RELATED EPISODES

How Anthropic Actually Makes Money — this S-1 is the document that revealed the Anthropic-SpaceX compute lease; the owner-vs-renter depreciation flip

How the Statute of Limitations Killed Musk v. Altman — covered the Feb 2026 xAI->SpaceX merger + the 3-way IPO governance comparison

The Last Independent: Cerebras — the revenue-vs-valuation IPO pattern this echoes at 60x the scale

CHAPTERS

00:00 A $1.75 trillion IPO that loses money

01:02 Three businesses, only one makes money

02:26 What $1.75 trillion is actually paying for

04:08 How the profit became a $4.9 billion loss

05:50 The Anthropic compute deal

06:56 The index rules engineered to force the bid

09:42 What it's worth: the $1.2 trillion floor

11:00 The bull case for Starlink

13:25 Governance and who's actually buying

14:18 What history says happens next

15:26 The verdict and two predictions

SOURCES

SpaceX Form S-1 (May 20, 2026; CIK 0001181412) — SEC EDGAR; Reuters, Axios, TechCrunch

Independent valuations ~$1.2T: Aswath Damodaran (NYU Stern) + research-firm sum-of-the-parts

Nasdaq-100 fast-entry change (May 1, 2026); S&P DJI profitability-waiver consultation (closes May 28, 2026)

SpaceX–Anthropic compute deal — S-1 ($1.25B/month, 90-day mutual termination)

Tesla S&P 500 inclusion (Dec 2020) — forced-buying precedent; IPO first-year base rates (Rivian, Meta)

———

This episode is for educational and informational purposes only and does not constitute financial, investment, or trading advice, nor a recommendation to buy or sell any security. Deep Dive is not a registered investment adviser. All investing involves risk. Consult a licensed financial professional before making investment decisions.