Sebastian Mallaby on OpenAI’s Cash Crunch, the AI Bubble Debate, and the China AI Race
Ed Elson interviews Sebastian Mallaby — venture historian, Hassabis biographer, and Council on Foreign Relations economist — on his standing bet that OpenAI runs out of money within 18 months. The argument threads a needle most AI commentary blurs: a specific company in trouble sits inside a sector that is not. Mallaby then turns to the government-stake gambit he calls ‘cheating capitalism’, and to why he thinks the United States can no longer beat China on AI and must negotiate with it instead.
Key ideas
- An OpenAI bubble, not an AI bubble. Mallaby splits the question the market keeps conflating. OpenAI he rates a coin-flip — a 50% chance that by next summer it cannot go public, cannot raise enough privately, and has to sell itself at a discount to a cash-rich acquirer. The wider sector — the compute, the semiconductors, the data centres — he judges real, because the underlying technical progress since ChatGPT has been genuine and fast. The bear case is a management story about one company, not a demand story about the field.
- The burn is structural, and the fundraising is theatre. OpenAI earned ~$13bn against ~$34bn of spending — a ~$21bn operating loss — while serving 900 million consumers, only ~5% of whom pay, many in India, Brazil and Indonesia where you cannot charge much. Its headline $122bn raise was, on inspection, roughly two-thirds future promises conditional on a successful IPO or paid in compute. The theatrical number, Mallaby argues, is a ‘head fake’ to manufacture the appearance of momentum.
- The government stake is an end-run around capitalism. OpenAI’s floated proposal to hand Washington a ~5% stake is, on Mallaby’s read, Altman recruiting ‘the investment banker to whom you can’t say no’. The template is Intel, up ~400% since the government took 10% last year — not on merit but because Commerce phoned other firms and steered contracts its way. Picking a national-security winner in chips he grants; extending it to one of several US foundation-model labs he calls unjustified, and warns ‘welcome to China’ if the real aim is propping up the stock market.
- The US can’t beat China on AI, so it must talk to it. After eight days touring Chinese labs, Mallaby found engineers who do discuss safety and excel at applications, models (JIPU) close to the anthropic frontier, and Chinese share of developer traffic on OpenRouter rising from under a third in late 2025 to ~60% by mid-2026. Chip export controls have not held them back — partly because a Malaysia-sized loophole lets Chinese builders rent offshore Nvidia compute. He reaches for a Cold War frame: compete hard, but negotiate a non-proliferation deal on open-weight models, as the US and USSR did on nuclear arms.
- Safety can’t be self-imposed inside a competitive lab. The lesson of his DeepMind book: Demis Hassabis wanted an independent ethics board and a single unified lab, and both hopes ‘crashed and burned’ — Google would not cede a fiduciary veto, and humanity proved too tribal for one lab. If a company cannot bind itself, only governments can enforce restraint on all players at once — which, for open-weight models, means a deal that includes Beijing.
Content
The runway thesis, twelve months on
Mallaby first wrote ‘This is what convinced me OpenAI will run out of money’ in January, betting the company had 18 months. He holds the bet. The original case was a burn rate he calls ‘crazy’ against an internally projected $660bn five-year spend — unraisable even for a fundraiser he credits as ‘a magician’. Since then he grants genuinely good news: OpenAI has cut Stargate data-centre commitments, killed the loss-making Sora video model, and imposed some order on chaotic management — but only ‘half successfully’. It remains squeezed between Anthropic, better at the enterprise applications customers actually pay for (coding, cyber-security, agents), and Google’s Gemini, which monetises retail users through search advertising now doing record revenue.
The IPO delay to 2027 he reads as fear of scrutiny — the WeWork parallel, where an audited prospectus met public daylight and the offering collapsed. OpenAI is boxed: it needs the IPO because private markets cannot fund the burn, yet attempting it risks a WeWork failure, after which ‘it’s really cooked’. A down-round is the other exit and just as painful — marked at $852bn post-money but trading far lower in secondaries, an honest re-rating to $600bn would gut employee equity, trigger departures, and convulse the momentum machine.
Cheating capitalism
The government-stake idea draws Mallaby’s sharpest language. He accepts Elson’s framing that it is ‘cheating’ the game — an end-run around the level playing field that is meant to be the wellspring of capitalist efficiency. A CFR colleague’s count puts 30 US companies under announced or actual federal equity stakes since January 2025. The Intel precedent looks like vindication only because the government manufactured Intel’s contracts; strip that out and it is state winner-picking. He allows the chip case a national-security logic — Taiwan concentration, invasion risk — but denies OpenAI any equivalent claim: it is one of several US labs, strategically replaceable. The systemic-to-the-stock-market defence he dismisses as a Fed-put writ political: once Washington picks winners, ‘you don’t have a market anymore’.
The China turn and the Cold War frame
The interview’s back half is Mallaby’s China reversal. A 2022 supporter of chip export controls — he wrote the Washington Post essay making the case — he now doubts they work, because Chinese models are not far behind and the controls leak. The deeper worry is proliferation: a near-frontier Chinese model released open-weight could hand any criminal the tools to hack the financial system, with no off-switch. Distillation — querying a strong US model to generate training data, a contract violation but not a federal crime, and one Musk’s xAI has also used — keeps Chinese models cheap and close. His prescription is the Cold War’s dual track: mutually-assured-destruction-style competition alongside a non-proliferation regime, which for AI requires negotiating with Beijing that neither side dismisses as dovish. He is ‘cheered’ that Trump’s executive order — nominally voluntary, in practice Commerce ordering Altman to seek per-customer sign-off — shows Washington will regulate when a model (the anthropic ‘mythos’ capability) frightens it enough. The gap is that no one yet wants to talk to China.
Related
- Sebastian Mallaby — guest; venture historian and Hassabis biographer
- Ed Elson — host
- Scott Galloway — Prof G Markets co-host (not in this episode)
- OpenAI's Leaked Financials, SpaceX's Valuation, and the AI Bubble — the Zitron financials Mallaby builds on
- Gary Marcus on AI's Overstated Intelligence, LLM Economics, and the Limits of Scaling — adjacent LLM-economics scepticism
- Howard Marks on the AI Bubble, Irrational Exuberance, and Investing Under Radical Uncertainty — the bubble question from an investor’s seat
- Token Economics — the AI unit-cost problem underlying the losses
- Narrative Valuation — valuations resting on a founder’s story rather than fundamentals