Deep Dive

Deep Dive

Deep Dive is long-form research on AI, tech, and the global economy. Single host, weekly episodes, 25-35 minutes each. The story behind every headline — built from primary sources and original analysis. Recent topics: • AI deanonymization research • Data center infrastructure economics • Strait of Hormuz geopolitics • Agentic AI security • Frontier model behaviors Find Deep Dive across platforms: 📺 YouTube · @DeepDiveAIShow 📱 TikTok · @notdeepdiveai 📷 Instagram · @notdeepdive 🔗 All links · linktr.ee/notdeepdive Tap follow for new episodes.

  1. How Close Are We to Self-Driving Cars in 2026?

    7 HR AGO

    How Close Are We to Self-Driving Cars in 2026?

    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

    16 min
  2. How Anyone Can Strip the Safety Out of an Open-Source AI Model

    7 HR AGO

    How Anyone Can Strip the Safety Out of an Open-Source AI Model

    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

    18 min
  3. The SpaceX IPO: Why the Biggest IPO in History Loses Money

    1 DAY AGO

    The SpaceX IPO: Why the Biggest IPO in History Loses Money

    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 (filed May 20, 2026; CIK 0001181412) — primary source via SEC EDGAR; reporting from Reuters, Axios, TechCrunch Aswath Damodaran (NYU Stern) — independent SpaceX valuation, ~$1.22 trillion Sum-of-the-parts analysis by a research firm — ~$1.25 trillion across seven segments Nasdaq-100 methodology change (effective May 1, 2026) — fast-entry eligibility S&P Dow Jones Indices — profitability-waiver consultation (comment period closes May 28, 2026) SpaceX–Anthropic compute agreement — disclosed in the S-1 ($1.25B/month, 90-day mutual termination) Tesla S&P 500 inclusion (December 2020) — forced-buying precedent IPO first-year base rates — Rivian and Facebook/Meta comparables

    17 min
  4. Why Wall Street Is Betting Billions on Nuclear Power for AI

    2 DAYS AGO

    Why Wall Street Is Betting Billions on Nuclear Power for AI

    In the spring of 2026, a company filed to go public at a $1.66 billion valuation — with no revenue, a "going concern" warning from its own auditors, and a reactor that exists only as an eight-inch test well. It was the second time it had tried to go public. And in the same six weeks, two other energy companies raised about $3 billion between them, all selling Wall Street the same one-sentence pitch: AI needs power. This episode reads that IPO wave as a capital-markets bet, not an energy explainer. Put the three companies — Fervo (geothermal), X-energy (small nuclear reactors), and Deep Fission (a deep-borehole reactor) — on a single risk ladder, and a pattern appears: the market priced engineering risk and time-to-power before it priced revenue. Fervo, the one with actual revenue, booked about $140,000 of it last year — against a $10 billion valuation. Even the banks sorted the deals by risk before the rest of us did: the blue-chip names took Fervo and X-energy; Deep Fission got the second-tier shops. Then the parts the headlines skip. Where the money actually goes (the durable margin sits upstream, in enriched fuel and a cleared spot on the grid — not the celebrated reactor startups). Whether the demand is even real (it's growing fast, but gas, not nuclear, is the actual near-term bridge, and the grid operators are already trimming their forecasts). And the question a federal ruling decides by the end of June: when a data center forces an expensive grid upgrade, who pays for it — the data center, or you? We close on the history that should worry you, the genuine bull case, and an honest update to a prior call that nuclear-for-AI was "mostly marketing" — early, not wrong, with a carve-out geothermal earned on a real cost curve. The through-line: the market is buying the sentence — AI needs power — not yet the megawatts, because the megawatts mostly aren't here. Plus two dated predictions. RELATED EPISODES The Real Cost of AI: Who's Actually Paying for the Build-Out — the ratepayer/cost-incidence side this episode updates (and the 'nuclear is mostly marketing' call it revisits) How Microsoft Is Restarting Three Mile Island — why restarts, not new builds, add electrons The Last Independent: Cerebras — the revenue-vs-valuation IPO pattern this wave echoes The AI Chip War — the bottleneck progression (logic → packaging → memory → power) this is the next chapter of CHAPTERS **This episode is analysis and opinion, not investment advice. Do your own research.** 00:00 Disclaimer — analysis, not investment advice 00:08 Cold open — a $1.66B reactor that doesn't exist yet 01:15 Three things to keep in front of you 01:36 Why power became the bottleneck 02:59 Three companies, one bet — the IPO wave 04:19 The risk ladder: producing vs. pre-revenue 06:40 Where the money actually goes (fuel + the grid) 08:03 Is the AI-power demand even real? 10:00 Who actually pays (your bill, and Phoenix) 11:22 The history that should worry you 14:15 The verdict, and two predictions 15:11 Where I land — the story vs. the megawatts SOURCES Deep Fission S-1 (filed May 20, 2026) — SEC EDGAR (CIK 0001918102); via Bloomberg, Axios, World Nuclear News, TechCrunch X-energy IPO pricing (April 23, 2026) — IPOScoop, X-energy launch PR Fervo Energy IPO pricing (May 12, 2026) — Fervo PR, Bloomberg, Fortune LBNL / DOE 2024 data-center electricity report (4.4% 2023 → 6.7-12% by 2028) IEA Electricity 2025 (gas/coal vs nuclear share of new data-center demand) FERC Docket RM26-4 — large-load interconnection cost allocation (rules ~end June 2026); CSIS, Holland & Knight NuScale CFPP cancellation (Nov 2023); Vogtle 3&4 cost overrun; V.C. Summer abandonment Fervo de-risking: Stanford Geothermal Workshop (Feb 2024) — 35% learning rate, $9.4M→$4.8M/well Net Zero Insights / PowerMag — 2025 nuclear-fission equity (~$1.3B, record)

    16 min
  5. How Anthropic Actually Makes Money

    3 DAYS AGO

    How Anthropic Actually Makes Money

    Anthropic — the company behind Claude — just posted its first-ever operating profit: a projected $559 million on $10.9 billion of revenue in one quarter, from internal projections reviewed by the Wall Street Journal. It's a real milestone. It's also softer than the headline, for a reason almost nobody is explaining. Every month, Anthropic writes a check for more than a billion dollars to a data-center operation now controlled by Elon Musk — about $15 billion a year, roughly 80% of that supplier's entire revenue, on a contract either side can cancel with 90 days' notice. And that bill was discounted during a spring 2026 ramp — the exact quarter the profit appears. Here's why it's structural, not a one-off. All of tech is arguing about whether hyperscalers like Microsoft and Oracle hide the true cost of their chips by stretching depreciation schedules — Michael Burry argues they're understating it by about $176 billion (his own model). Anthropic has the opposite problem: it rents its chips instead of owning them, so it can't smooth that cost at all. The lease is fixed. That's why one discounted quarter flips it to profit and the next can flip it back. You can't smooth a rent check. We walk down the AI "margin ladder" — chipmaker around 75%, cloud 50–55%, model labs 50–60% gross but operating-negative, apps near 25% — to show why a middle-of-the-ladder company is being valued like it sits at the top. Its last confirmed valuation was $380 billion; reports of $900 billion-plus, and an implied ~$1 trillion on thinly-traded private shares, are headlines, not clearing prices. Then we weigh the bull and the bear, because both are right. Bull: the cost to serve a dollar of revenue fell from 71 to 56 cents in a single quarter, and one analysis has inference margins jumping from 38% to over 70%. Bear: the contracts funding the build-out are only 90 days deep, and Bain estimates AI needs about $2 trillion in annual revenue by 2030 and faces an ~$800 billion shortfall. The bridge between them is one question — do you own your compute, or rent it? RELATED EPISODES Musk v. Altman — the courtroom rival is now Anthropic's compute landlord NVIDIA Just Forecast $91 Billion Without China — NVIDIA's ~75% margin is the top rung of the ladder this episode climbs down The Real Cost of AI: Who's Actually Paying for the Build-Out — the 2026 unit-economics update The Last Independent: Cerebras — the 'your supplier is also your competitor' structure CHAPTERS **This episode is analysis and opinion, not investment advice. Do your own research.** 00:00 Disclaimer — analysis, not investment advice 00:08 Cold open — the Musk check 02:07 The bill: ~$15B/yr to Musk, and the 90-day exit 04:15 The revenue, and the first profit 05:35 Why the profit is just one quarter (the discount) 06:08 The depreciation fight: Burry vs. the audited anchor 07:37 The inversion — you can't smooth a rent check 08:38 The margin ladder: who keeps the money 09:45 The valuation: $380B confirmed vs. $1T thin air 10:20 Bull vs. bear, and the three break conditions 13:50 The verdict: own your compute, or rent it 14:34 Two things to watch, and the close SOURCES SpaceX S-1 (May 2026) — via Reuters, Axios, TechCrunch Anthropic — Series G announcement (Feb 12, 2026) Wall Street Journal — Q2 projections, via PYMNTS and CNBC Amazon Q4 2024 earnings release / 10-K (Feb 2025) Michael Burry / 'Cassandra Unchained' (Nov 2025) Tanay Jaipuria, citing Bessemer (Sep 2025) SemiAnalysis — 'AI Value Capture' (May 2026) Bain — 6th Annual Global Technology Report (Sep 2025)

    16 min
  6. NVIDIA Just Forecast $91 Billion Without China

    4 DAYS AGO

    NVIDIA Just Forecast $91 Billion Without China

    On May 20th, 2026, NVIDIA reported its biggest quarter ever — $81.6 billion in revenue, up 85% year over year — and guided to $91 billion for the next quarter while explicitly assuming zero data-center compute revenue from China. A year earlier, that same quarter carried $4.6 billion of China sales. NVIDIA zeroed China out and grew anyway. The same export-control regime that cut NVIDIA off from China did not cut China off from the world's developers. On OpenRouter — the platform developers use to route to whichever model wins — Chinese open-weight models reached about 45% of all tokens, up from roughly 1.2% eighteen months earlier. DeepSeek lined up its first outside capital (a reported ~$45B valuation, though the actual raise is ~$300M). The catch: those models are still trained on NVIDIA chips — only inference has tilted to domestic silicon. NVIDIA's $50.3 billion of operating cash in a single quarter settles whether its own demand is real. The bubble question migrated downstream to its buyers: OpenAI has committed on the order of $600 billion of compute through 2030 against roughly $20 billion of revenue — and the $100 billion NVIDIA-into-OpenAI deal that anchored the circular-financing story never actually closed (absent from NVIDIA's own 10-Q, and Jensen Huang denied the figure). The through-line: export controls bet that mutual dependence was a chokepoint. The first hard quarterly data says it held in neither direction. America's AI machine proved it doesn't need China as a customer; China proved its developer economy doesn't need NVIDIA as a supplier. The decoupling already happened — and both economies are still standing. RELATED EPISODES The AI Chip War: Why the Bottleneck Keeps Moving — the export-control bet this print is the verdict on The Real Cost of AI: Who's Actually Paying for the Build-Out — the circular financing and bubble math Platform Engineering at AI-Native Companies — the three predictions that just landed The Last Independent: Cerebras — the non-NVIDIA inference precedent CHAPTERS **This episode is analysis and opinion, not investment advice. Do your own research.** 00:00 Disclaimer — analysis, not investment advice 00:08 Cold open — $91B without China, and the mirror 01:21 The export-control bet 01:53 The print: $91B guided, $0 from China 03:14 The cash-flow proof — and why the stock fell anyway 04:38 China's side: refusing the H200, building its own 05:28 How China captured the developers (OpenRouter) 07:08 DeepSeek's first raise — hype vs. real 08:09 The mechanism: memory and the architecture cadence 09:19 The memory squeeze hits the phone market 11:34 The bubble question — and the $100B deal that never closed 14:18 The verdict: the decoupling already happened 15:08 Predictions, and the close SOURCES NVIDIA Q1 FY2027 press release + 10-Q (SEC EDGAR, filed May 20, 2026) — revenue, guidance, China-zero outlook, cash flow Reuters / Tom's Hardware — H200 China customs block (Jan 14, 2026) + Lutnick zero-shipped confirmation (Apr 22, 2026) OpenRouter 'State of AI' 100T-token study + April 2026 rankings — Chinese-model token share The Information / Bloomberg — DeepSeek first external round (valuation vs raise; investor list single-source) TrendForce (Dec 26, 2025) + IDC + Counterpoint — HBM/DRAM wafer allocation and smartphone-market squeeze NVIDIA 10-Q + Spyglass — the $100B OpenAI deal absent/downsized; Huang denial Yahoo/CNBC + PYMNTS + a16z OpenRouter — OpenAI ~$600B compute vs ~$20B ARR; developer token growth

    17 min
  7. How the Statute of Limitations Killed Musk v. Altman

    6 DAYS AGO

    How the Statute of Limitations Killed Musk v. Altman

    On May 18, 2026, a nine-person advisory federal jury in Oakland took under two hours to dismiss every claim in Musk v. Altman — and dismissed them on the calendar, not the merits. Judge Yvonne Gonzalez Rogers immediately adopted the verdict under FRCP 52, saying she was "prepared to dismiss on the spot." The three surviving claims — breach of charitable trust against Altman/Brockman/OpenAI (3-year SOL, cutoff Aug 5, 2021), unjust enrichment (2-year SOL, cutoff Aug 5, 2022), and aiding-and-abetting against Microsoft (3-year SOL, cutoff Nov 14, 2021) — all failed the discovery-rule test on Musk's own September 24, 2020 tweet: "OpenAI is essentially captured by Microsoft." Microsoft's counsel called the suit "more than a year too late." But the substantive question — was OpenAI's nonprofit-to-for-profit conversion a real breach of charitable trust? — was never reached. The merits weren't adjudicated. They were preempted. Two state AGs had already settled the question administratively. On October 28, 2025, California's Bonta and Delaware's Jennings issued non-objection statements ratifying the 26% nonprofit / 27% Microsoft / 47% employees-investors PBC split. The conversion question is now ratified by two AGs but never tested in court. The trial evidence is permanently public and none of it reached the jury: Sutskever's 52-page dossier alleging Altman's "consistent pattern of lying," Brockman's 2017 diary calling the conversion "wrong to steal the nonprofit," Murati's deposition that Altman was "not always" candid, Microsoft's 24-hour $25B absorption plan, the 745-of-770 employee petition, the jackass trophy. And the structural irony nobody else has named: Musk's own xAI dropped its Nevada PBC status on May 9, 2024 — three months before he filed the lawsuit on the theory that abandoning a nonprofit AI mission was illegitimate. xAI was then absorbed into SpaceX February 2, 2026 at a $1.25T combined valuation, with the S-1 surfacing May 20, 2026 — two days after the verdict. Three trillion dollars of pre-IPO governance lands in 18 months: OpenAI ~$1T (Q4 2026/2027), SpaceX-with-xAI $1.75T (June 2026), Anthropic $380B+ (October 2026). The next AI lab to attempt similar conversion inherits administrative ratification without judicial validation. RELATED EPISODES The Mythos Bifurcation — the OpenAI/Anthropic governance split Claude Mythos — the AI governance arc this verdict partially resolves The Cerebras IPO — the 20x oversubscribed precedent for OpenAI's path How Microsoft Is Restarting Three Mile Island for AI — the Microsoft side of OpenAI's $250B Azure commitment CHAPTERS 00:00 Cold open — the verdict, under two hours, May 18, 2026 01:20 How the statute of limitations actually works 02:30 Three claims, three cutoffs, the September 2020 tweet 03:17 The advisory jury and Gonzalez Rogers's bench finding 04:29 The appeal, and why it is a steep climb 06:30 Trial evidence that didn't matter — Sutskever's 52-page dossier 07:44 The November 2023 firing reconstructed 08:21 Microsoft's 24-hour absorption plan 09:21 Brockman's diary, Murati's deposition, the jackass trophy 12:22 Two AGs ratified what no court has — Bonta + Jennings, Oct 28, 2025 15:02 The Friar-vs-Altman IPO timing fight 17:56 The xAI structural irony — Musk did the exact thing he sued over 21:04 Anthropic's Long-Term Benefit Trust 22:27 Three trillion dollars of pre-IPO governance in 18 months 23:19 Five dated predictions 25:16 Closing — what the verdict didn't decide SOURCES Court coverage — Musk v. Altman (N.D. Cal.): TechCrunch, NPR, MIT Tech Review, CBS California AG Bonta + Delaware AG Jennings — non-objection statements (Oct 28, 2025) OpenAI + Microsoft official recapitalization disclosures (Oct 28, 2025) Fortune — OpenAI cash burn + compute commitments (Nov 12, 2025) Bloomberg — OpenAI $110B at $730B (Feb 27, 2026) Anthropic Long-Term Benefit Trust documentation (Corp Gov Harvard) CNBC — xAI dropped Nevada PBC status (Aug 25, 2025); SpaceX S-1 (May 2026)

    27 min
  8. How Microsoft Is Restarting Three Mile Island for AI

    19 MAY

    How Microsoft Is Restarting Three Mile Island for AI

    On September 20, 2024, Constellation Energy announced the largest single-buyer corporate power purchase agreement ever signed. Microsoft buys 20 years of output — every megawatt-hour — from a restarted Three Mile Island Unit 1, renamed the Crane Clean Energy Center, targeting 2027 (originally 2028). Analysts model the price at $98 to $115 per MWh against $50/MWh PJM wholesale. The $1.6B restart works only because three subsidies stack: the IRA Section 45U PTC at up to $15/MWh, the 45Y clean energy credit, and a $1B DOE Loan Programs Office loan that closed November 18, 2025. But the plant being restarted is Unit 1. Unit 2 — the sister reactor — partially melted down on March 28, 1979. The cause was a Pilot-Operated Relief Valve that stuck open for 2 hours 22 minutes while a control-room indicator light, wired to the valve's command solenoid rather than its actual position, told operators it had closed. Eighteen months earlier at Davis-Besse — same Babcock & Wilcox PWR family — the same valve stuck open. Shift supervisor Mike Derivan diagnosed it in 20 minutes. The plant was at 9% power. B&W engineers Joseph Kelly (November 1, 1977) and Bert Dunn (February 9, 1978) wrote memos warning that "core uncovery and possible fuel damage would have resulted" at full power. Managers Don Hallman and Bruce Karrasch dismissed the warnings. Met-Ed was never notified. The Kemeny Commission found B&W's PORV had failed 11 prior times, 9 stuck open. The 2026 restart audits whether 47 years of post-1979 safety architecture — Kemeny, NUREG-0660, NUREG-0737, INPO, 10 CFR 50.155 — stuck. Up to 7,500 hours of NRC pre-restart inspection. Three license amendment requests pending hearing. Intervention petitions due April 27, 2026. A FERC waiver for 760 MW of capacity rights from retired Eddystone units, decision needed by June 1. PJM's June 30 base capacity auction as the operational hinge. And the grid math doesn't care. PJM's load forecast: 30 GW of data center growth between 2025 and 2030. The entire U.S. nuclear restart pipeline is 2.25 GW. Capacity auction prices hit the price cap in December 2025, falling 6,625 MW short of reliability targets for the first time in PJM history. RELATED EPISODES The Real Cost of AI — the Virginia residential bill correction ($11.24/mo from Jan 2026); the transformer cycle The AI Chip War — PJM capacity auction history; the Stargate loop solvency math The Cerebras IPO — capital-flow vs power-flow framing on AI infrastructure The Strait of Hormuz — format reference for regulatory-case-file layering CHAPTERS 00:00 Cold open — Davis-Besse 1977: the valve sticks open at 9% power 01:09 The 1977 lesson took 18 months and a meltdown to stick 01:41 How the PORV failed: indicator light wired to the wrong thing 02:26 The Kelly memo, the Dunn memo, and what B&W management did 03:21 March 28, 1979 — 2h 22m, 32,000 gallons, 11 prior failures 04:15 Kemeny: 'the fundamental problems are people-related problems' 05:50 Unit 1 vs Unit 2 — the critical distinction, and 2019 retirement 06:49 The Microsoft–Constellation deal and the three stacked subsidies 09:33 Why a $1 billion federal loan that wasn't strictly needed 09:41 The FERC behind-the-meter rejection and the virtual-PPA structure 11:26 Grid math: 30 GW demand vs 2.25 GW restart pipeline 14:13 Why not SMRs? The NuScale cancellation and the LCOE comparison 15:59 The improvised regulatory path: no formal 'un-decommissioning' rule 19:17 Spring 2026 hinge dates: April 27, June 1, June 30 21:34 Closing — 47 years, one indicator light, and the price of certainty SOURCES Constellation Energy — Crane Clean Energy Center announcement (Sept 20, 2024) DOE Loan Programs Office — $1B loan close (Nov 18, 2025) NRC info-finder: Crane Clean Energy Center docket Kemeny Commission report (October 1979) UCS Dave Lochbaum — PORV indicator-light wiring + Davis-Besse 1977 link Bloomberg Businessweek — Bennett & Wade (May 7, 2026) Utility Dive — Microsoft-Constellation PPA + FERC waiver

    24 min
See All (42)

About

Deep Dive is long-form research on AI, tech, and the global economy. Single host, weekly episodes, 25-35 minutes each. The story behind every headline — built from primary sources and original analysis. Recent topics: • AI deanonymization research • Data center infrastructure economics • Strait of Hormuz geopolitics • Agentic AI security • Frontier model behaviors Find Deep Dive across platforms: 📺 YouTube · @DeepDiveAIShow 📱 TikTok · @notdeepdiveai 📷 Instagram · @notdeepdive 🔗 All links · linktr.ee/notdeepdive Tap follow for new episodes.

Information

  • Creator
    Deep Dive
  • Years Active
    2k
  • Episodes
    42
  • Rating
    Clean
  • Copyright
    © Deep Dive
  • Show Website
    Deep Dive

You Might Also Like