Stratechery · Ben Thompson Anthropic and Alignment Mon, Mar 2

Thompson's deep-dive on the showdown between Anthropic and the Department of War over AI safeguards — and what it means for the future of AI governance.

Key Takeaways

  • The federal government moved to designate Anthropic a supply-chain risk, an unprecedented label never before applied to an American company, after Anthropic refused to remove safeguards around mass domestic surveillance and fully autonomous weapons.
  • Anthropic's public statement from Dario Amodei drew two hard lines: no mass domestic surveillance, no fully autonomous weapons — positions it says were never part of prior DoW contracts.
  • OpenAI swooped in immediately, securing a classified DoD contract while Anthropic talks collapsed.
  • Thompson's core argument: if AI is as powerful as its advocates claim (nuclear-weapon-equivalent), then governments will inevitably seek to control it — or destroy whoever refuses. Anthropic's position is "fundamentally misaligned with reality."
  • The binary: either Anthropic accepts a subservient role to the U.S. government, or the government moves to destroy or remove Anthropic's leadership.
  • Palmer Luckey argued the core question is democratic accountability — should military AI be regulated by elected leaders or corporate executives?
  • Thompson's remedy: new laws and accountable oversight structures, not unelected executives making unilateral calls about powerful national-security capabilities.
"International law is ultimately a function of power; might makes right... Anthropic talks a lot about alignment; this insistence on controlling the U.S. military, however, is fundamentally misaligned with reality."
Read on Stratechery →
Stratechery · Ben Thompson Anthropic's Skyrocketing Revenue, A Contract Compromise?, Nvidia Earnings Wed, Mar 4

Anthropic Revenue

  • Anthropic has surpassed a $19B annualized revenue run rate, up from $9B at end of 2025 — more than doubling in a matter of months.
  • Growth driven primarily by Claude Code adoption and strong enterprise subscription uptake; Anthropic now leads OpenAI in API spend on the Ramp platform.
  • Anthropic is valued at $380B and appears on track to surpass OpenAI's $20B annualized revenue mark.
  • Enterprise AI subscriptions are more durable than consumer ones — enterprises value productivity and will pay; consumers won't (and the consumer model eventually becomes advertising).

A Contract Compromise?

  • Amodei signaled at the Morgan Stanley TMT conference that Anthropic wants to de-escalate: "We will try our very best to do that."
  • Analysis from Under Secretary of State Jeremy Lewin: Anthropic's contract failed because surveillance restrictions were defined too vaguely, giving Anthropic operational veto power. OpenAI succeeded by using specific, legalistic definitions — giving DoW clarity and confidence.
  • Thompson's hope: Anthropic uses the OpenAI contract as a template to de-escalate and re-engage with the Pentagon.

Nvidia Earnings

  • Nvidia reported 94% profit growth and $68.1B in Q4 revenue (+73% YoY), beating estimates. Stock still fell 5.5% as investors wanted stronger forward guidance.
  • Three AI inflection points driving token demand: (1) LLMs answering questions, (2) reasoning models generating tokens before answering, (3) agents running reasoning models autonomously — a squared increase in addressable compute market.
  • Jensen Huang: "Inference equals revenues now for our customers because agents are generating so many tokens." Compute = revenues in the agentic era.
  • Bear case (software threatened by AI) and bull case (more software created with AI as the limit) both require more tokens — and therefore more Nvidia chips.
Read on Stratechery →
Stratechery · Ben Thompson Technological Scale and Government Control · Paramount Outbids Netflix for Warner Bros. Tue, Mar 3

Technological Scale and Government Control

  • Nuclear weapons were government-funded from day one (Manhattan Project, ~$36B inflation-adjusted). AI is different: private companies built it for markets vastly larger than government, which means government wasn't the funder — but it will still demand control.
  • Bob Noyce's insight with chips: the civilian market would dwarf the government market, and volume-driven scale would produce far better technology. AI has this dynamic "on steroids."
  • The uncomfortable reality: property rights and private decision-making are ultimately enforced by someone — and if that someone is the government, it can choose to override them. Thompson isn't endorsing this, just describing it.
  • Key implication: tech companies building AI of nuclear-equivalent power should not expect to retain veto rights over government use, regardless of whether that's normatively good.

Paramount Outbids Netflix for Warner Bros.

  • Paramount Skydance won the bidding war for Warner Bros. Discovery at $31/share, after Netflix declined to match, calling it a "nice to have, not must have."
  • Netflix stock jumped on news it lost the deal — suggesting the market (and perhaps Netflix internally) saw the acquisition as risky or overpriced.
  • Paramount is promising $6B in synergies by 2029 — a horizontal deal (both are in the same businesses: streaming, cable, studios, theatrical).
  • More antitrust scrutiny expected than Netflix's proposed deal, which was more vertical. Thompson thinks it should still be approved — consolidation is the natural response in a secularly declining industry.
  • Netflix's saving grace: it may eventually be able to acquire both Paramount and Warner Bros. simultaneously as debt loads force the issue, avoiding separate integration costs.
Read on Stratechery →
Emerge · Liam Gill OpenAI on Defence Contract Wed, Mar 4

Market overview + the OpenAI/Anthropic Pentagon fallout from Liam Gill's tech and markets newsletter.

Dow 48,501 +0.91%
S&P 500 6,816 -0.42%
Nasdaq 22,516 -3.12%
Russell 2000 2,608 +5.09%
Bitcoin $68,512 -22.67% YTD
  • In a tense all-hands, Sam Altman told employees OpenAI will not "weigh in" on specific military operations — decisions like the Iran strikes or the Maduro capture are the government's sole purview.
  • Altman acknowledged the "sloppy" timing of OpenAI's Pentagon announcement (which came hours before a major joint U.S.-Israel bombardment of Iran), but defended the deal as the right move.
  • OpenAI will build its own "safety stack" but ultimately defers to Defense Secretary Hegseth on operational use cases.
  • Silicon Valley rift widening: OpenAI and Elon Musk's xAI signal willingness to work within "any lawful use" mandates; Anthropic is reportedly preparing to challenge its supply-chain blacklist in court.
  • As the U.S.-led Iran operation enters day 3, the mission's stated objectives have shifted five times, drawing bipartisan criticism over a lack of unified exit strategy.
  • Commerce Secretary Howard Lutnick voluntarily agreed to testify before the House Oversight Committee following scrutiny over newly surfaced ties to Jeffrey Epstein.
Read at Emerge →
Vector · Grant Janich Meta's New AI Engine Room Wed, Mar 4

Meta's Applied AI Engineering Org

  • Meta is standing up a new Applied AI Engineering organization led by Reality Labs VP Maher Saba, reporting to CTO Andrew Bosworth.
  • Radical structure: up to 50 individual contributors per manager — an unusually flat ratio designed for execution speed over organizational overhead.
  • Core mission: build the data infrastructure and evaluation pipelines ("the data engine") that turn capable models into dominant ones, partnering directly with the Superintelligence Lab led by former Scale AI CEO Alexandr Wang.
  • Meta's bet: the gap between a good foundation model and a market-leading one increasingly lives in data quality, evaluation rigor, and iteration speed — not just parameter count.
  • New models code-named Avocado and Mango are in development; Zuckerberg promised shipping "in the coming months."
  • Broader trend: AI-native workflows (coding assistants, automated testing) are compressing org charts and reducing the need for traditional middle management across Silicon Valley.

OpenAI Building Its Own GitHub

  • OpenAI is internally developing its own code repository — a direct alternative to GitHub (owned by Microsoft, OpenAI's biggest backer).
  • The project was prompted by repeated GitHub outages that disrupted engineering workflows; staffers have reportedly discussed making it available externally, which would put OpenAI in direct competition with Microsoft's $7.5B developer platform.

Other News in Brief

  • Anthropic approaches $20B annualized revenue, even as Pentagon feud becomes a marketing tailwind.
  • UK funds £40M frontier AI research lab — a small but symbolic bet on homegrown capability.
  • Alibaba's top Qwen model leader steps down, injecting uncertainty into China's most visible open-model effort.
  • Guild.ai raises $44M Series A for AI agent governance and auditability across stacks.
  • Ethan Agarwal, a 40-year-old tech entrepreneur, announces congressional run in California's 17th district.
Read at Vector →
Above Avalon · Neil Cybart Day 2 of Apple's Spring Launch — MacBook, iPad Air, iPhone 17e Wed, Mar 4

New Products Announced

  • MacBook Air (M5) — $1,099 starting price ($100 higher than M4, reflecting doubled base storage from 256GB to 512GB). The 512GB option is actually $100 cheaper than last year.
  • MacBook Pro (M5 Pro and M5 Max) — Price floor $200 higher, also reflecting doubled starting storage. Unlike Air, storage pricing for equivalent configs is unchanged.
  • M5 Pro / M5 Max chips — New "Fusion Architecture" combines two dies into a single SoC. M5 Pro claims up to 30% faster multithreaded CPU vs M4 Pro. Up to 128GB unified memory on M5 Max.
  • Studio Display ($1,599) updated; 27" Studio Display XDR ($3,299) is new; the $4,999 32" Pro Display XDR is being discontinued.

iPad Air (M4)

  • iPad Air (M4) brings unified memory to 12GB — 50% more than M3 — with 120GB/s bandwidth. Pricing unchanged.
  • M-series gap between iPad Pro and iPad Air returns to one generation (M4 vs. M5).
  • Apple is heavily marketing Apple Intelligence as a key iPad Air selling point, giving the device AI "futureproofing" for several years.
  • Of this week's announcements, the iPad Air (M4) will sell in the highest volume.

More on iPhone 17e

  • At $599, the iPhone 17e is in "no-man's-land" — not low enough to meaningfully shift the prepaid market or emerging markets, where price points need to be $300 or under.
  • With iPhone 17 seeing renewed upgrader momentum, a $599 entry model feels a bit out of place — it may cannibalize very little.
  • The real low-end iPhone question remains unanswered: can Apple bring a premium experience to the $349-$399 range to grow share in China (~25%), India, Brazil, and southern Europe?
Read at Above Avalon →
Tomorrow's Briefing · Arielle Phillips X to Creators: Stop Tricking People With Fake War Videos — Or No Money for You Wed, Mar 4
  • X announced that creators who post AI-generated conflict videos without disclosing they're fake will lose access to the Creator Revenue Sharing Program for 90 days.
  • Repeat offenders get permanently removed from the program.
  • Rationale from X product head Nikita Bier: during an actual war, people need real information — not AI-generated "Call of Duty: Misinformation Edition."
  • Enforcement will use a combination of AI detection tools and Community Notes.
  • Notable caveat: the policy covers only war/conflict content. AI fakery about politics, products, celebrities, or crypto is still allowed.
  • Critics note that X's revenue-sharing model (which pays for engagement and requires a paid subscription to join) has long incentivized clickbait and dramatic content — real or not.
Read at Tomorrow's Briefing →
Stratechery · Ben Thompson This Was an Xbox — Week of Feb 23 Fri, Feb 27

Weekly roundup from the Stratechery bundle. Highlights:

Xbox: From Owning the Living Room to Ceding Hardware

  • After Phil Spencer's exit, Thompson toured Xbox's 25-year history — a strategy misaligned for at least 15 years, where Microsoft held onto Xbox as "the sole piece of evidence the company could be cool."
  • The new Xbox head is not a gamer. Thompson's read: Microsoft is finally doing what it should have done a decade ago — exit the first-party console hardware business.

Space Chip Fabs — Technically Possible, Completely Impractical

  • In 2016, Jeff Bezos floated building chip factories in space. With chip constraints urgent again, companies are exploring it — but an Asianometry video breaks down why LEO chip fabs face insurmountable challenges: cooling, radiation, constant maintenance, and reimagining the entire chip stack.

Other Content

  • Greatest of All Talk had Rachel Nichols on — from covering Michael Jordan as an intern to the collapse of the Washington Post sports section, with a side of Wemby and Spurs playoff hopes.
  • Stratechery video: Thin Is In (Apple hardware design trends).
Read on Stratechery →
Stratechery · Ben Thompson An Interview with Bill Gurley — Runnin' Down a Dream Thu, Feb 26

Ben Thompson interviews long-time (now retired) venture capitalist Bill Gurley about his new book Runnin' Down a Dream — on building a career you love, the Uber saga, and the current state of venture capital.

  • Gurley's book draws on his decades in VC, including his time on the Uber board during one of the most dramatic corporate governance sagas in Silicon Valley history.
  • The conversation covers how to identify and pursue work you're genuinely passionate about — the "runnin' down a dream" thesis for career building.
  • Gurley and Thompson discuss how the modern VC landscape has changed, including the role of large late-stage rounds, founder-friendly terms, and whether traditional VC economics still hold.
Read on Stratechery →