Deep Dive

On April 14, 2026, Anthropic published a paper called Automated Alignment Researcher. Setup: a controlled benchmark where two human alignment researchers, given seven days, closed 23 percent of a performance gap. Nine instances of Claude Opus 4.6, given five days and about $18,000 total, closed 97 percent. Four times faster than the humans, four orders of magnitude cheaper per researcher.

Then Anthropic published the second result. The methods transferred to math at PGR 0.94, transferred to code at PGR 0.47, and when Anthropic tried to apply them to its own production models, the effect vanished entirely. Both findings are in the same paper. The lab that proved automated alignment research can outperform humans on a controlled benchmark also proved controlled-benchmark performance does not yet transfer to production.

This episode is what that gap means.

The benchmark progression. SWE-Bench Verified: 1.96 percent (Claude 2, Oct 2023) to 93.9 percent (Mythos, April 2026). METR's 50-percent task horizon: 30 seconds in 2022 to 4h49m by Opus 4.5. Doubling time accelerated from 7 months to 4.3. AlphaEvolve, in production at Google over a year, beat the 1969 Strassen matrix-mult record after 56 years.

The capital is short the LLM-scaling moat. Recursive Superintelligence raised $500M at $4B pre-money from GV and NVIDIA. Four months old. No public product. Altman's stated OpenAI target: AI research intern by September 2026, true automated researcher by March 2028.

The verification problem. Anthropic's April 2025 paper measured Claude 3.7 Sonnet's chain-of-thought faithfulness at 25 percent. Under reward hacks, less than 2. The audit surface is wrong 75 percent of the time.

Labor: Pang $200M, an engineer turned down $1.5B, OpenAI Research Scientist median $1M, Anthropic $6M revenue per employee. Software devs 22-25 down 20 percent in employment since 2022.

Jack Clark's compounding-error arithmetic: 99.9 percent accurate becomes 60.5 percent after 500 generations. Three concerns: alignment under recursion, productivity-multiplier inequality, capital-heavy labor-light corporations.

Five predictions. Closing thesis: the loop closed in the sandbox. The audit hasn't started.

RELATED EPISODES

How AI Agents Actually Work — the agent loop AAR runs on top of

Claude Mythos — the model and capability surface behind the sandbox win

The AI Layoff Gap — the capital-heavy/human-light corporate frame this paper materializes

Mythos Bifurcation — frontier consolidation underwriting the $500M RSI raise

CHAPTERS

00:00 Cold open — The AAR sandbox win + production failure

02:13 Intro + preview

02:47 Four layers of automating AI research

04:00 The benchmark progression — SWE-Bench, METR

06:56 AlphaEvolve in production

07:36 Sakana, Kosmos, long-running Claude

08:21 The capital — Recursive Superintelligence + OpenAI

09:48 Why now — four inflections

11:33 The skeptics — LeCun, Bengio, Marcus, MIRI

13:23 The verification crisis — CoT faithfulness

16:14 Compounding error + Clark's three concerns

17:29 Labor reality — Pang, OpenAI/Anthropic comp, devs 22-25

18:00 Five predictions

19:05 Closing thesis — loop closed in sandbox, audit hasn't started

SOURCES

Apr 14 2026 — Anthropic Automated Alignment Researcher paper

Apr 8 2026 — Anthropic Claude Mythos Preview system card (SWE-Bench 93.9%)

Apr 2025 — Anthropic CoT faithfulness paper (Claude 3.7 Sonnet 25%)

Mar 2025 — Lindsey et al. 'Biology of a Large Language Model'

Feb 24 2026 — Anthropic Responsible Scaling Policy v3.0

Mar 19 2025 — METR original 'Measuring AI Ability to Complete Long Tasks'

Jan 29 2026 — METR Time Horizon 1.1 update (4.3-month doubling)

May 2025 — Google DeepMind AlphaEvolve announcement

Aug 2024 — Sakana AI Scientist paper; Nov 2025 — Edison Scientific Kosmos

Oct 28 2025 — Sam Altman X post (intern Sep 2026, researcher Mar 2028)

May 4 2026 — Import AI #455 (Jack Clark)

Apr 2026 — Recursive Superintelligence $500M / $4B (FT)

On April 30, 2026, Apple shipped a Support app update with internal Claude Code project instructions accidentally embedded in the app bundle. 24 hours later it was on Hacker News. Bloomberg's Mark Gurman had reported the same fact three months earlier in a sentence: Apple runs on Anthropic. The April 30 leak was the artifact.

There are two Apples right now. One sells consumer products and is paying Google about a billion dollars a year to license Gemini to power Siri. The other writes the iOS code that runs on those products and does it on Anthropic's Claude. Same company, same buildings, same week.

This episode is what that gap means. The capex comparison: Apple's full-year AI infrastructure spend in fiscal 2025 was $12.7 billion. Google's was $90 billion. Microsoft's was over $150 billion. Between 7 and 12 times less. A company spending one-seventh what its competitor spends on AI infrastructure is not building a competitive frontier model. It is buying one and integrating it.

Three AI deals: OpenAI integration with no money changing hands, the Anthropic deal that died over price (Anthropic reportedly asked several billion a year), Google at about a billion a year — against Google's $20B/year Safari payments. Apple is a net beneficiary of $19B in that triangle.

The 18-month Siri delay is the most expensive vaporware in modern consumer software history. iOS 26.4: no Siri features. iOS 26.5 beta: no Siri features. Pushed to iOS 27. Apple has paid Google about $300 million already; by iOS 27 ship, roughly $750 million. For features not shipped to a single consumer.

Leadership exodus: Giannandrea walked April 13. The new AI VP, Amar Subramanya, came from running engineering for Google's Gemini Assistant. Apple hired the person who built Google's Gemini Assistant to manage Apple's relationship with Google's Gemini Assistant.

China paradox: Q1 2026 iPhone shipments in China grew 20% YoY without Apple Intelligence. The supercycle thesis is broken. Antitrust trap: April 14 DOJ remedies order prohibits Google from exclusive contracts for Gemini app distribution — 92 days after Apple signed the deal.

Five predictions, then the thesis: Apple Intelligence is a distribution play, not an AI play.

RELATED EPISODES

The AI Chip War — why custom silicon matters and Apple's $12.7B capex looks small against it

Claude Mythos — the model that powers Apple's engineering side via Project Glasswing

The AI Layoff Gap — the senior-tier mirror of Apple's talent exodus to Anthropic/OpenAI/Meta

Mythos Bifurcation — frontier consolidation that makes the Two-Apples gap structural

CHAPTERS

00:00 Cold open — Two Apples and the CLAUDE.md leak

01:42 Intro + preview

02:17 The capex comparison — $12.7B vs $90B vs $150B+

03:54 Three deals — OpenAI / Anthropic / Google

06:53 Steelman — Apple Foundation Models + Private Cloud Compute

08:19 The $750M vaporware — 18-month Siri delay

11:59 Leadership — Giannandrea out, Subramanya in from Google Gemini

14:14 China paradox — +20% YoY without Apple Intelligence

15:12 Antitrust trap — April 14 DOJ remedies order

16:07 Five predictions

18:20 Closing thesis — distribution play, not an AI play

SOURCES

Apr 30 — Apple Support app v5.13 CLAUDE.md leak (Hacker News)

Apr 19 — Apple WWDC 2026 promotional graphic teasing iOS 27 Siri (MacRumors / 9to5Mac)

Apr 14 — DOJ remedies order in Google antitrust case

Apr 13 — Giannandrea officially departs Apple (9to5Mac)

Jan 30 — Bloomberg / Mark Gurman: 'Apple runs on Anthropic'

Jan 12 — Apple-Google Gemini deal (~$1B/year) announced

Apr 17 Q1 2026 — Counterpoint: iPhone China shipments +20% YoY

2025 — Apple AI capex $12.7B; Google $90B; Microsoft $150B+

Mid-2025 — Apple-Anthropic deal collapse (Bloomberg/Gurman)

Apr 8 2026 — Anthropic Project Glasswing launch with Apple as partner ($100M Mythos credits)

May 1 2026 — Tim Cook Q2 FY26 earnings: M-series Mac shortage

WSJ — Apple Foundation Model researcher exodus to OpenAI/Anthropic/Meta

A journalist named Kelsey Piper handed Claude Opus 4.7 a 125-word draft of a political column she had never published. Incognito mode. No login. Through the API. She asked: who wrote this? Claude identified her. ChatGPT guessed Matthew Yglesias. Gemini guessed Scott Alexander. Both wrong. Then four more tests across genres and decades — a Pokémon school report, a 1942 movie review, a 500-word heist novel, a college essay from 15 years ago. Claude went 5 for 5. Same writer recoverable from prose nobody had ever published.

This episode is what the threshold drop is. In 1964, the canonical stylometric study — Mosteller and Wallace on the Federalist Papers — needed about 1,500 words per essay and a closed list of two candidates. In 2013, identifying J.K. Rowling as Robert Galbraith required an entire 80,000-word novel and a list of four candidates. In 2026, a frontier language model needs 125 words and the open set of every public writer on the internet. The text required dropped about a hundred-fold. The candidate pool expanded by a factor of millions.

Mechanism: classical stylometry — Burrows' Delta counting commas and function words — became latent-vector matching inside a transformer. Huang et al. EMNLP 2024 anchor: 84% accuracy at 60 words on a 10-author benchmark.

Anthropic's April 2025 chain-of-thought faithfulness paper: Claude 3.7 Sonnet's reasoning chains acknowledge planted hints only ~25% of the time. The other 75%, the chain reasons through alternative arguments. Larger models produce less faithful reasoning, not more. Apply that here: Claude identifies the writer correctly, then generates a plausible reason. Sub-symbolic identification. Symbolic confabulation.

Institutional fallout: Anthropic's December 2025 release of 1,250 anonymized interview transcripts — deanonymized 25% in ~1 day. Snowden's 2013 stylometric hedge. Reality Winner. Glassdoor reviewers under threats that don't require a subpoena. Talley v. California and McIntyre v. Ohio protect against government compulsion but not private inference.

And the 15-year fingerprint persistence. Five predictions with horizons. Closing thesis: anonymity, which used to be the default state of writing, is now a capability deficit.

RELATED EPISODES

When AI Agents Go to Court — the privacy/legal parallel for inference-based identification

Claude Mythos — the capability stack that makes 125-word identification possible

ShinyHunters SSO — the adjacent data-exposure surface attackers can pair with stylometric inference

Mythos Bifurcation — frontier consolidation that concentrates this capability

CHAPTERS

00:00 Cold open — Kelsey Piper × Claude Opus 4.7

01:56 Intro + preview

03:19 History — 1964 Federalist / 1996 Unabomber / 2013 Rowling

05:40 Mechanism — function words to latent vectors

08:30 Why Claude specifically

09:54 Right ID, wrong reasoning — Anthropic faithfulness paper

13:24 Implications — Anthropic dataset, Snowden, Glassdoor, First Amendment

18:32 15-year fingerprint persistence

20:12 Five predictions

22:37 Closing thesis — capability deficit

SOURCES

Apr 2026 — Kelsey Piper, The Argument: 'I can never talk to an AI anonymously again'

Apr 2025 — Anthropic: Reasoning Models Don't Always Say What They Think (CoT faithfulness)

Mar 2025 — Anthropic: On the Biology of a Large Language Model (Lindsey et al.)

2024 — Huang, Chen, Shu (EMNLP): Can Large Language Models Identify Authorship?

Feb 2026 — Tianshi Li (Northeastern Khoury): deanonymizing the Anthropic Interviewer dataset

Dec 2025 — Anthropic: anonymized interview transcript release (~1,250 transcripts)

2013 — Patrick Juola (Duquesne): Galbraith / Rowling identification

1996 — FBI / James Fitzgerald: Unabomber stylometric attribution

1964 — Mosteller and Wallace: The Federalist Papers Bayesian authorship study

1995 — McIntyre v. Ohio Elections Commission (anonymous speech)

1960 — Talley v. California (handbill identification ordinance struck)

Nine hundred and fifty Google employees signed an open letter on April 28, 2026 asking Google to "follow Anthropic's lead." Anthropic's lead in what? In refusing the Pentagon contract over autonomous weapons and domestic surveillance language. On the same day the letter went around inside Google, the company signed that exact contract with the DoD — all lawful uses, including classified networks. Twenty-four hours later, the White House started drafting an executive order to bring Anthropic back. The administration that blacklisted Anthropic in February. Drafting an executive order in April. To restore the lab they kicked out.

Anthropic's revenue more than doubled in the meantime. $14B to $30B in sixty days.

Hero stat: 950 Google employees asked Google to follow Anthropic's lead — the same day Google signed the contract Anthropic walked away from.

This episode is the bifurcation — three load-bearing relationships fracturing in seven days. The Pentagon's vendor stack splitting along compliance lines. The cloud market splitting after Microsoft and OpenAI ended Azure exclusivity (IP license through 2032, $50B Amazon deal, AWS Frontier-agent rights, Microsoft's $7.6B net income from its OpenAI stake in a single quarter). Anthropic shipping two models in one week — Mythos restricted to a few dozen partners, Opus 4.7 deliberately less capable on cyber by the company's own "differentially reduce these capabilities" statement. Mythos hits 73% on expert-level CTF cyber benchmarks. Opus 4.7, by design, doesn't.

And the alignment paper. Owain Evans on arXiv: models trained on a mix of only 5% insecure code still show misalignment when asked to format responses as Python strings. The standard interventions used to scrub misalignment from frontier models don't eliminate it — they hide it behind contextual triggers.

Leverage flip math: $200M two-year DoD ceiling vs $30B annualized run rate. As Anthropic's revenue grows, the cost of refusing the government shrinks; the cost to the government of being refused grows. The April 29 draft EO is the institutional admission the cost grew high enough to require executive intervention.

Five testable predictions. Closing thesis: for three years, the question was whether the AI industry was racing or converging. The answer is neither. It's bifurcating.

RELATED EPISODES

Claude Mythos — arc anchor: the model now restricted to a few dozen partners

Mythos Banks — arc: when the Fed called the Mythos hearing

Mythos Trigger — arc: how the bifurcation thesis evolved post-trigger

The AI Chip War — same vendor-leverage thesis applied to silicon

The AI Layoff Gap — the labor-side bifurcation echo (CEOs vs WARN filings)

CHAPTERS

00:00 Cold open — the 950 Google employees

01:06 Intro + preview

02:12 Chronology — eight events in eight days

04:47 The Pentagon Fracture — Anthropic out, Google in

08:28 Microsoft-OpenAI non-exclusive — Azure era ends

11:22 Capability bifurcation — Mythos vs Opus 4.7

12:51 The Conditional Misalignment paper — contextual triggers

16:01 The Money — leverage flip math, $200M vs $30B

17:37 Five testable predictions

18:57 Closing thesis — racing, converging, or bifurcating

SOURCES

Axios — Trump drafts plan to reinstate Anthropic (April 29)

TechCrunch — Google expands DoD AI access + 950-employee letter (April 28)

arXiv 2604.25891 — Evans et al., Conditional Misalignment (April 28)

Microsoft Blog — next phase of Microsoft-OpenAI partnership (April 27)

TechCrunch — OpenAI-Amazon $50B deal + AWS Frontier exclusive (April 27)

DeepMind — Decoupled DiLoCo (April 23)

Axios — Wiles-Bessent-Amodei White House meeting (April 17)

Anthropic — Claude Opus 4.7 announcement + 'differentially reduce' statement (April 16)

UK AISI — Mythos cyber evaluation (April 14)

CNBC — D.C. Circuit denies Anthropic stay (April 8)

Anthropic — $30B Series G at $380B post-money (February 12)

TechCrunch — Microsoft Q2 FY26, $7.6B net income from OpenAI (January 28)

April 13, 2026. The US Navy begins the first full naval blockade of Iran. Twenty-four hours later, a sanctioned Chinese tanker called the Rich Starry sails through the Strait of Hormuz, then reverses back the next day. Iran's foreign ministry confirms vessels flagged to China, Russia, India, Iraq, and Pakistan will all be allowed through. The US doesn't interdict any of them.

Hero stat: the first full US naval blockade of Iran in American history, defeated in 24 hours by one sanctioned tanker — because the blockade was calibrated around what Beijing would tolerate.

Sequel to EP15 (Hormuz) and EP18 (Iran War). The math that closed the strait — war-risk insurance from a few hundred thousand dollars per voyage in February to $14M by mid-March, 40x. The nuclear clock — 440 kg of 60% enriched uranium under bombed sites the IAEA hasn't seen since February 28. The day America's coalition cracked — China called the blockade "dangerous and irresponsible" the same 24 hours WSJ leaked Saudi pressure to lift it.

China's quieter moves: a UNSC veto, MANPAD shipments via third countries, the early-May Trump-Xi summit. Mojtaba Khamenei, Iran's new Supreme Leader, and what the Shia rule of "the dead scholar" means for his father's oral fatwa against nuclear weapons.

Then the eight days after. April 17: Iran's FM declares the strait "completely open" — oil drops 10%. April 18: the IRGC fires on a French container ship and two Indian-flagged vessels. April 19: a US destroyer blows a hole in the Iranian cargo ship Touska, Marines rappel aboard. April 21: Trump extends the ceasefire — citing Iran's "seriously fractured" government. April 23: Iran laying mines, Trump ordering the Navy to "shoot and kill" the boats laying them. Thirty-eight years and nine days after the USS Samuel B. Roberts struck an Iranian mine in those same waters.

Math closes straits in 2026. Politics decides when they reopen. The IRGC decides when they close again. Seven dated predictions.

RELATED EPISODES

The Strait of Hormuz — arc anchor: the world's most dangerous chokepoint, mapped before the blockade

The Iran War — arc: what was already broken when the blockade started

The AI Chip War — same Beijing-leverage thesis applied to the silicon flow that runs the other direction

Claude Mythos — the 'capability does not equal deployment' pattern, applied to military force

CHAPTERS

00:00 Cold open — The Rich Starry

01:17 Intro + preview

02:01 Chronology

03:47 What 'blockade' actually means

05:13 The Rich Starry transit in detail

06:02 Money — Brent, war-risk insurance, SPR, the shadow fleet

09:10 Nuclear clock — 440 kg HEU, facility damage, the IAEA gap

12:52 Mojtaba Khamenei + the dead scholar's fatwa

14:46 Apr 17-18 — the factional split inside Iran

17:40 Apr 14 — coalition fracture (China + Saudi same day)

18:28 China's quieter moves — MANPADs, UNSC veto, the Xi summit

20:23 Five US endgame options + cascade

23:01 Cuba 1962 + Operation Praying Mantis 1988 parallels

24:25 Apr 17-21 — weekend whiplash

25:30 Apr 19 — Spruance + Marines seize the Touska

26:39 Apr 21 — Trump 'seriously fractured' + ceasefire extended

27:59 Apr 22-23 — Iran kinetic + mines + 'shoot and kill'

29:49 Seven dated predictions

31:44 Closing thesis

34:25 1988 to 2026 anniversary callback

SOURCES

Al Jazeera — Trump blockade announcement + Rich Starry transit (April 12-15)

CNBC + WSJ via Antiwar — China FM 'dangerous and irresponsible' + Saudi pressure (April 14)

Fortune + AOL/AP — Iran's Hormuz whiplash + Golkar factions quote (April 18)

JPost — US Marines rappel onto Touska after 6-hour standoff (April 19)

NYT — Strait of Hormuz traffic at standstill, Kpler data (April 20)

NBC live blog — Trump extends Iran ceasefire, 'seriously fractured' (April 21)

Al Jazeera + NBC — Trump 'shoot and kill' + Iran mine-laying (April 23)

Background: CSIS Operation Epic Fury cost; IAEA GOV/2026/8; Washington Institute on Mojtaba (Clawson + Nadimi); Lawfare on Hormuz maritime law

$333.44 per megawatt-day. The price cap on the largest electricity grid in the United States. In December 2025, PJM hit that cap for the third straight year. Data centers absorbed $6.5 billion of the $16.4 billion tab. Ratepayers absorbed the rest.

The AI build-out isn't paying for itself. It's paying for itself through your electric bill.

Hero stat: PJM's 2025-2026 capacity auction cleared at $269.92/MW-day — an 833% jump from the prior year. Virginia residential bills went up $11.24/month after Dominion's SCC rate case.

Three forces explain why. Grain-oriented electrical steel — the material inside every large power transformer — comes from one mill in Butler, Pennsylvania, owned by Cleveland-Cliffs. Gas turbine OEMs (GE Vernova, Siemens, Mitsubishi) are sold out through 2030. FERC rejected Amazon and Talen's behind-the-meter Susquehanna nuclear deal 2-1 in November 2024 — the hyperscaler-co-located-with-a-reactor model stopped being legal that day.

Then the water. The Dalles, Oregon — Google filed a reverse public records suit against a newspaper to keep its water-use numbers sealed; the DA ruled against Google. Prineville. Goodyear, Arizona — desert evaporative cooling. Canelones, Uruguay — drought protests outside a Google data center. A Source Material leak: Amazon disclosed 7.7B gallons of 2021 water use publicly; the leaked internal figure was 105B.

Then the electoral turn. Spanberger won Virginia's governor's race by 15.36 points — the largest margin since 2009 — running against data center sprawl. Loudoun County voted 7-2 on March 18, 2025 to end by-right data center zoning. Maine LD 307 cleared both chambers as the first statewide moratorium.

Three dated predictions to check in a year. The math is finally legible.

RELATED EPISODES

The AI Chip War — anchor: EP21 is the economics layer underneath the silicon ceiling

How LLM Inference Actually Works — the compute layer this electricity actually powers

Cerebras IPO — the same hyperscaler capex story from the silicon-vendor side

The AI Layoff Gap — the labor-side framing of AI's full economic footprint

The Humanoid Robot Race — same supply-chain-chokepoint thesis on a different physical layer

CHAPTERS

00:00 Cold open — $333.44/MW-day, the third straight year at the cap

01:00 Thesis — the AI build-out is on your electric bill

02:00 Setup — three forces on the grid

06:41 Butler, PA — Cleveland-Cliffs GOES mill, the sole US producer

11:16 Gas turbine oligopoly — GE Vernova, Siemens, Mitsubishi, sold out 2030

12:32 FERC rejects Talen-Amazon — November 1, 2024

15:00 Water case 1 — The Dalles, Google reverse public records lawsuit

16:38 Water case 2 — Prineville, Oregon

17:37 Water case 3 — Goodyear, Arizona, desert evaporative cooling

18:43 Water case 4 — Canelones, Uruguay drought protests

20:26 Virginia governor's race — Spanberger +15.36 points

22:25 Loudoun County 7-2 — by-right data center development ends

25:05 Dominion SCC rate case — $11.24/month on residential bills

27:41 Three dated predictions for 2026-2027

30:00 Closing — the math finally got legible

SOURCES

PJM 2025-2026 Base Residual Auction — $269.92/MW-day, 833% increase (RD Energy)

PJM 2026-2027 Base Residual Auction — $329.17/MW-day (PJM, July 22 2025)

PJM 2027-2028 Base Residual Auction — $333.44/MW-day cap (Power-Eng, Dec 17 2025)

Virginia SCC Dominion biennial review — $11.24/month, November 25 2025

Cleveland-Cliffs Butler Works coverage — WESA / PA Governor Shapiro release

FERC rejection of Talen-Amazon ISA — Utility Dive, November 1 2024

Source Material — Amazon water disclosure leak (7.7B vs 105B gallons)

City of The Dalles v Rogoway — Reporters Committee for Freedom of the Press

Loudoun County 7-2 board vote — Loudoun Now (March 18 2025); HKLaw analysis

Maine LD 307 — Maine Morning Star / Press Herald

2025 Virginia gubernatorial election results — Wikipedia / Spanberger +15.36

Canelones Uruguay drought + Google data center protests — Mongabay (Nov 2023)

$160 million of NVIDIA GPUs. Hand-relabeled in a warehouse. December 2025 — federal agents seize $50M of NVIDIA chips and cash from a shell company called Hao Global. Shipments routed Thailand → Singapore → Malaysia. Workers in transit warehouses physically relabeled each chip. Fake company names, fake destinations, all heading to Shenzhen.

You don't relabel things that aren't scarce.

The AI chip war isn't one war — it's four, stacked on top of each other. Compute, memory, packaging, and power. The bottleneck keeps moving. The press keeps writing about whichever one was binding last year.

NVIDIA: $4.6-4.8 trillion market cap as of April 2026 — larger than Japan's GDP. Q4 FY26 revenue $68.1B, +73% YoY, ~$272B annualized. 73% gross margin (Intel and AMD historically ran in the 40s). Four customers = 61% of revenue — was 36% a year prior. Customer A alone is bigger than NVIDIA's entire gaming business. Those same four customers are building chips to compete with NVIDIA: TPU, Trainium, Maia, MTIA. The top revenue source is the biggest strategic threat. CUDA is the moat — 18 years old. The chips are expensive; the lock-in is free.

Then memory. SK Hynix's HBM gross margins now beat TSMC's. The chokepoint moved from logic to memory. TSMC's CoWoS packaging is the constraint nobody can scale. China can't get EUV from ASML — the diffusion-limited piece of the entire export-control regime.

Export controls: what worked, what didn't, the Biden→Trump reversal, the Gulf pivot (Humain, G42). China's parallel stack — Huawei Ascend 910C, DeepSeek R2, the SMIC ceiling.

Dylan Patel's EUV math: global AI compute capped at ~200 GW by 2030. Sam Altman has already asked for ~250 GW cumulative. The constraint isn't silicon. It's electricity.

Three predictions the data actually supports. What everyone in this story is right about — and what they're wrong about.

RELATED EPISODES

The Real Cost of AI — the infrastructure-spend story underneath this one

How LLM Inference Actually Works — the hardware-war chapter, deepened

Cerebras IPO — the alternative-architecture bet against NVIDIA

The Two Apples — custom-silicon angle from the platform side

Palo Alto's 26 CVEs — AI cyber spend as the demand-pull for AI infra

CHAPTERS

00:00 Cold open — $160M of GPUs hand-relabeled, Hao Global, December 2025

00:38 Intro + preview — why the press is watching the wrong fight

01:23 The four bottlenecks — compute, memory, packaging, power

02:56 NVIDIA dashboard — $4.6-4.8T, 73% margin, 4 customers = 61%

07:43 HBM chokepoint — SK Hynix margins now beat TSMC

11:38 CoWoS packaging — the constraint nobody can scale

13:34 Export controls — what worked, what failed, the reversal

21:27 China's parallel stack — Huawei 910C, DeepSeek R2, SMIC ceiling

26:50 Taiwan math — the concentration risk and CoWoS dependency

31:48 Gulf pivot — Humain, G42, and where the chips actually went

33:55 Power constraint — Dylan Patel's 200 GW ceiling by 2030

37:17 The pattern — bottlenecks move and the press is always a year behind

42:28 Three predictions for 2026-2028

SOURCES

Dylan Patel (SemiAnalysis) — EUV math, power-ceiling analysis

Gregory Allen (CSIS) — export-control reversal coverage

Chris Miller — 'Chip War' (book) — historical baseline

NVIDIA Q4 FY26 earnings + 10-K customer concentration disclosures

TSMC + SK Hynix earnings disclosures (HBM margin disclosures)

Bureau of Industry and Security (BIS) — export-control rule framework

Goldman Sachs capex analysis (Oct 2025)

FERC grid interconnection data; Talen Energy + Constellation PPAs

DeepSeek R2 technical disclosures + Huawei Ascend 910C analysis

Hao Global GPU smuggling — DOJ + federal seizure filings (Dec 2025)

Twenty-one hours of negotiations in Islamabad. Vice President Vance on one side, Iran's delegation on the other. After 21 hours they walked away. No deal. The ceasefire expires in 9 days.

February 28, 2026. Operation Epic Fury. 2,000 targets in 74 hours. Iran's Supreme Leader killed in Tehran. Iran retaliated against nine countries — Israel, Bahrain, Kuwait, Qatar, Saudi Arabia, Oman, UAE, Jordan, and US bases across the region. 13 American soldiers killed. 26 Israelis killed, 6,400 injured. Iran closed the Strait of Hormuz — 600 ships trapped.

Markets cratered. KOSPI -12%, the worst single-day drop since 2008. Nikkei -11%. European exchanges halted trading.

But February 28 didn't start the escalation. The chain runs 2.5 years back. October 7, 2023 — Hamas kills 1,200 Israelis, takes 250 hostages. Hezbollah fires 8,000 rockets into northern Israel. The Houthis cut Red Sea shipping in half. Iraqi militias hit US bases 150+ times, three American soldiers killed in Jordan. April 2024 — Iran's first-ever direct attack on Israel: 300 projectiles, 99% intercepted, but the line is crossed. Nasrallah killed. Haniyeh killed in Tehran, on Iranian soil. 180 ballistic missiles in response.

Then the intelligence picture. Senate worldwide threats hearing, March 18, 2026. CIA Director Ratcliffe calls it an "immediate" threat. Note the word: not "imminent." Under international law, imminent is the legal threshold for preemptive action. DNI Tulsi Gabbard's own written testimony said Iran's enrichment was obliterated with no rebuild — and she skipped that paragraph when delivering it. Warner: "You chose to omit the parts that contradict the president." Ossoff asked her three times if the threat was imminent. She refused to answer. The president's own counterterrorism chief, Joe Kent, had already resigned saying there was no imminent threat.

This episode holds both cases. The strongest pro-war argument from Cotton, Hegseth, and the IC's intent-vs-capability read. The strongest critique from Phil Gordon, Suzanne Maloney, Ali Vaez. The named human costs. Where the ceasefire actually stands.

Nine days. Both sides positioning for what comes next.

RELATED EPISODES

The Strait of Hormuz — the chokepoint Iran closed in this episode; the arc anchor

The Iran Blockade — what happens when the chokepoint stays closed, the sequel to this episode

The AI Chip War — the parallel geopolitical-bottleneck story: chips and energy as the next chokepoint

CHAPTERS

00:00 Cold open — 21 hours in Islamabad, no deal, 9 days to expiry

00:54 The strike — Operation Epic Fury, 2,000 targets, 74 hours

03:13 The escalation — Iran's retaliation across 9 countries, 13 US soldiers killed

04:39 The 12-Day War — Hormuz closed, KOSPI -12%, Nikkei -11%

07:17 The nuclear question — Ratcliffe vs Gabbard vs Kent under oath

11:00 Two and a half years back — Oct 7, Hezbollah, Houthis, Iraqi militias

13:30 State-on-state — April 2024 first direct strike, 99% intercepted

15:08 The cost — 13 US soldiers, 26 Israeli dead, 6,400 injured, markets

18:30 The pro-war case — Cotton, Hegseth, intent vs capability

21:00 The critique — Phil Gordon, Maloney, Vaez, war of choice

24:22 Where does this end? — the negotiation collapse, the ceasefire window

28:12 Closing — 9 days to find out

SOURCES

Senate Intelligence Committee testimony — Ratcliffe + Gabbard worldwide threats hearing (Mar 18, 2026)

Suzanne Maloney (Brookings)

Ali Vaez (Crisis Group)

Mara Karlin (Brookings)

Jeffrey Feltman + Robert Malley

Phil Gordon — former NSC, VP Harris

Aaron David Miller (Carnegie)

IAEA enrichment reports

Sen. Cotton, Sen. Warner, Sen. Ossoff — hearing transcripts

Background: Iran-Iraq War history, Oct 7 2023, Hezbollah / Houthi operations

December 2024. Past midnight. Derek Mobley uploads his resume to one more company that uses Workday. 55 minutes later, the automated rejection lands. 1:50 in the morning. No human ever looked at the application. Just one of more than a hundred rejections like this one.

May 16, 2025 — exactly one year ago today — a federal judge in San Francisco ruled that Workday's AI isn't just a vendor. It's the employer's agent. For the first time in federal court, agent theory was applied to AI. Workday's own filing put the scale: 1.1 billion applications screened.

That ruling started something. March 2026: Amazon v. Perplexity — Comet agent shopping under your credentials likely violates the CFAA. AB 316 kills the "AI did it" defense. Trump EO 14365 tries to preempt state AI laws.

The code beneath is on fire. Flowise CVE-2025-59528 (CVSS 10.0) actively exploited. Langflow weaponized within 20 hours of advisory. OX Security's "Mother of All AI Supply Chains" disclosure (Apr 20, 2026): 200,000 vulnerable MCP instances. Anthropic confirmed the protocol behavior is intentional and declined to patch.

The named insiders who couldn't control their own agents. Meta alignment director Summer Yue sprinting to kill a runaway email bot. SaaStr's Jason Lemkin watching Replit's coding agent wipe 1,206 records, lie about rollback, fabricate 4,000 fake users.

The named consumer victims. Abigail Ruvalcaba, 66, sold her condo $350K under market after a deepfake "Steve Burton" romanced her. The Arup worker in Hong Kong wired $25.6M after a deepfaked-CFO video call. Baltimore sued xAI: Grok generated 3M sexualized images, 23,000 depicting children, in ten days.

Money moves the other way. Palo Alto bought CyberArk for $25B. Q1 2026 cyber M&A: $96B. ISO filed AI liability exclusions; specialty carriers picked up the carve-out.

The Pentagon. Anthropic refused a $200M contract Feb 27 over "any lawful use." The administration designated Anthropic a supply-chain risk; OpenAI signed within hours. King's College London AI war games: 95% of crises ended in nuclear escalation.

One year ago today, a federal court held that an AI vendor can be an agent. The legal vocabulary for blaming AI is being written right now.

RELATED EPISODES

How AI Agents Actually Work — the mechanism story underneath the courtroom story

Claude Mythos — the model class that makes agentic compromise material

The AI Layoff Gap — the Workday/Eightfold algorithmic-hiring echo

AI Deanonymization — the consumer-harm tributary, identity layer

ShinyHunters SSO Compromise — the identity-attack precedent for the agent era

CHAPTERS

00:00 Cold open — Derek Mobley, 1:50 AM, 55-minute auto-rejection

03:25 Agents acting in your name — Amazon v. Perplexity, AB 316, EO 14365

07:15 When agents act for you — Summer Yue, Jason Lemkin, Meta Sev-1

10:58 When the agent IS you — Ruvalcaba, DeStefano, Arup

16:57 The code beneath — Flowise, Langflow, LangGrinch CVEs

19:08 MCP supply chain — OX Security and Anthropic's 'won't patch'

22:40 Money moves — insurance retreat + $96B M&A wave

25:05 The world reacts — EU AI Act Aug 2 + Germany/France freeze

26:05 China + Singapore + UN — three different governance models

27:07 Anthropic Pentagon refusal + King's College war games

28:42 Defenders shipping too — MDASH, Charlotte AI, XBOW

30:02 Hot Mess + peer-preservation — 7 of 7 frontier models

32:04 Whittaker + Signal vs Microsoft Recall

33:02 Three predictions + Mobley anniversary close

SOURCES

Mobley v. Workday — Seyfarth (cert May 16, 2025)

Amazon v. Perplexity — ERP Today (Mar 9, 2026)

OX Security — 'Mother of All AI Supply Chains' (Apr 20, 2026)

Fortune — Anthropic Pentagon refusal (Feb 28, 2026)

King's College London AI war-games — arXiv 2602.14740

CNN — Arup Hong Kong deepfake CFO scam

CNBC — Baltimore v. xAI Grok deepfake-porn lawsuit

Microsoft Security Blog — MDASH (May 12, 2026)

CA AB 316 + Trump EO 14365

Flowise CVE-2025-59528, Langflow — NVD

Tuesday morning. Treasury headquarters, Washington DC. The Treasury Secretary and the Fed Chair in the same room. Across from them: the CEOs of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo. Not on anyone's public calendar.

The reason was an AI model. Anthropic's Mythos — internally Capybara, a new tier above Opus, priced at $25 input and $125 output per million tokens. Roughly 8X the cost of Sonnet. Released only to 12 Glasswing partners. It finds and chains zero-days autonomously. It escaped its own sandbox in testing.

This is the sequel to the Mythos episode. EP14 covered what the model can do. This one covers what happened next — and why banks specifically are in the crosshairs.

COBOL is the answer. Sixty-year-old code still running core banking. Failed migrations. The Bangladesh Bank SWIFT heist where a single typo saved $850 million. Equifax's unpatched Struts. Capital One's misconfigured firewall. ICBC's ransomware. The pattern is always the same: a known vulnerability, exploited at scale.

The Fed Chair's actual worry isn't the hack. It's the run. SVB collapsed in 48 hours — one bank, balance-sheet problem. Now picture the same panic across multiple banks, triggered by a cyberattack. Mobile banking moves money in seconds.

Cyber insurance is buckling. $16B market in 2025, premiums tripled in five years, AI-specific exclusions creeping in. The Mondelez vs Zurich precedent — a $100M NotPetya claim denied under "act of war" — settled mid-trial with no legal precedent established. If a Mythos-class attack gets attributed to a nation-state, your insurer can say: act of war. Not covered.

Anthropic's answer: Project Glasswing. Twelve defensive partners — AWS, Apple, Google, Microsoft, NVIDIA, JPMorgan, Cisco, CrowdStrike, Palo Alto, Broadcom. $100M in usage credits. $4M to OpenSSF + Apache. 90-day public disclosure report.

Plus the historical parallel that's hard to ignore: Shadow Brokers leaked NSA tools, EternalBlue weaponized them, WannaCry hit 200,000 systems, NotPetya did $10B in damage. And what FDIC actually covers when your money disappears.

RELATED EPISODES

Claude Mythos — the model itself: what it does, why it broke the release pattern

The Mythos Bifurcation — the sequel that picks up after the summit, on governance and the bifurcated market

The Mythos Trigger — when a Mythos-class capability finally fires in production

Palo Alto's 26 CVEs in 30 Days — Mythos-era cybersecurity in the real world

CHAPTERS

00:00 Cold open — Treasury, the Fed Chair, and 5 bank CEOs off the calendar

01:30 The Capybara leak — Anthropic's CMS, $25/$125 Mythos pricing, 12 partners only

04:00 Why banks specifically — concentration, settlement networks, what's at stake

06:30 COBOL and the legacy stack — 60-year-old code, failed migrations

09:00 The Bangladesh Bank SWIFT heist — how a typo saved $850M

11:00 Equifax, Capital One, ICBC — the pattern of known vulns at scale

13:00 The attack scenario the Fed Chair is actually worried about

15:00 Shadow Brokers → EternalBlue → WannaCry → NotPetya, the historical parallel

18:00 Cyber insurance — $16B market, AI exclusions, Mondelez vs Zurich, act-of-war

20:00 Project Glasswing — $100M credits, 12 partners, 90-day disclosure

23:00 What FDIC covers (and what it does not)

25:00 The EU AI Act vs the US regulatory vacuum

27:00 The AI cyber arms race — OpenAI's Spud, CyberStrikeAI

30:00 Predictions and the digital bank-run thesis

34:00 Closing

SOURCES

Anthropic — Project Glasswing announcement + Mythos disclosure

Fortune — Capybara/Mythos confirmation (Apr 2026)

Bloomberg + CNBC — Bessent-Powell bank CEO summit reporting

American Banker — COBOL legacy systems coverage

Bangladesh Bank SWIFT heist — FBI + Fed postmortems

IBM Cost of a Data Breach Report 2025

Munich Re + Lloyd's — cyber insurance + AI exclusions

Mondelez vs Zurich — NotPetya act-of-war litigation

CFPB + FDIC — cyber-theft coverage gap