Anthropic's Safety Lead Puts A >10% Extinction Probability On The Product Going Into Its $2T IPO

On September 9, 2026, Anthropic pretraining researcher Jacob Coxon, 27, posted on X that he had resigned after three years split between Anthropic and OpenAI, accusing both of \"racing straight to self-improving superintelligence and gambling with our lives.\" Within hours, Anthropic's alignment science lead Evan Hubinger endorsed the post directly: \"Jacob is correct here — we really do earnestly believe AI could kill all humans!\" and added that he personally puts the probability above 10% within the next decade. Coxon's exit is the first internal defection Anthropic has weathered since Mrinank Sharma's February 9 resignation letter warned the \"world is in peril,\" and the first to land on the eve of a public S-1.

Anthropic told its roadshow audience to underwrite a $30 trillion total addressable market built on the premise that frontier AI companies with rigorous safety programs will capture the lion's share of cognitive work. Hubinger's tweet does not contradict that pitch — it confirms the underlying premise is load-bearing and that even the people running the safety program rate the downside as a double-digit-percent probability. The same day, Reuters reported the public S-1 has slipped to late September, with the roadshow now opening mid-October and a listing window of late October that lands just days before the November U.S. midterms.

Anthropic's compute stack concentrates dollar flow through three listed suppliers, and the Coxon-Hubinger episode changes the risk distribution for each in a different way. Amazon is the largest single equity holder at $8B invested plus an additional up to $20B committed in April 2026 alongside 5 gigawatts of Trainium capacity and a $100B AWS spend commitment. Alphabet followed in late April 2026 with $10B of fresh capital plus up to $30B more, alongside access to roughly one million next-generation TPUs. Broadcom co-builds the TPU silicon via a three-way deal disclosed April 6, 2026 that delivers about 3.5 gigawatts of next-generation TPU capacity to Anthropic starting in 2027.

On the GPU side, NVIDIA remains the residual compute supplier — Anthropic trains on a multi-cloud mix but its compute footprint is overwhelmingly TPUs and Trainium. For downstream channel exposure, Microsoft is the relevant mirror — its OpenAI partnership sits opposite Anthropic's stack, so any rerating of the safety narrative for one lab is mechanically the inverse for the other. Oracle competes for the same enterprise AI-infrastructure dollars via OCI and the Stargate build-out, and CoreWeave sits in the marginal GPU-cloud bucket whose multiples will re-rate with any change in the AI safety discount applied to the whole complex.

The backdrop into which Anthropic's IPO is now landing is not neutral. Gallup's May 2026 survey showed 71% of Americans oppose nearby data-center construction, with 48% strongly opposed; Quinnipiac's March 2026 poll found 55% of Americans believe AI will do more harm than good and only 21% trust the results they get from AI platforms. A Brookings tally through the first quarter of 2026 counted 75 data-center projects blocked or delayed by local opposition — $130B of planned construction paused. Anthropic's own S-1 is expected to flag AI backlash as a material risk factor, per legal-firm analysis of peer filings.

Short-term, the catalysts are calendar-driven: the public S-1 lands after Labor Day (late September); the roadshow opens mid-October; the listing window closes at the end of October, two days before the November 3 midterms. Coxon's resignation and Hubinger's quote will appear as exhibit risk in pre-marketing one-on-ones with the four bookrunners — Morgan Stanley, Goldman Sachs, JPMorgan and Citi — and at least one syndicate desk will want a written safety-risk addendum before the roadshow begins. If the S-1 hits with a Hubinger-style quote already embedded in the risk factors, the discount is partially priced in; if it lands silent, expect a same-day step-down in Amazon, Alphabet and Broadcom on the news cycle alone.

Long-term, the structural question is whether 'safety as moat' survives as a fundable thesis. If Anthropic successfully IPOs inside the $2T band, the safety premium is institutionalised — every subsequent frontier-AI valuation will price it in, and Alphabet and Broadcom capture the second-order TPU build-out that follows. If the listing prices below $1.5T, or pushes past November, the safety narrative cracks and the capex case for the multi-gigawatt GPU and TPU build-out re-rates down. That re-rating is what CoreWeave is most exposed to and what Microsoft's OpenAI stack most benefits from.