In February 2026, Dario Amodei told Fortune that missing AI demand forecasts by a single year could push Anthropic into bankruptcy. By May, the company had grown 80-fold in a single quarter, closed a $65 billion Series H at a $965 billion valuation, and quietly filed a confidential IPO registration statement with the SEC. The sequence illuminates how Amodei thinks about the most capital-intensive bet in Silicon Valley history.
The Math Behind the Mortality Warning
On February 14, Fortune published an interview in which Amodei detailed the asymmetric economics of frontier AI infrastructure. Compute must be contracted 18 to 24 months before demand is confirmed. A company that sizes its infrastructure budget for 10x annual revenue growth and receives 5x instead ends up carrying capacity it cannot pay for. "Then you go bankrupt," Amodei said, per Fortune's reporting. Data Center Dynamics, covering the same public statements, quoted him more fully: "The way you buy these data centers, if you're off by a couple of years, it can be ruinous."
Anthropic had committed approximately $19 billion in compute spend for 2026, calibrated for the 10x annual growth curve the company could model at the time. What arrived was closer to 80x quarterly growth. To meet demand that outran its own projections, Anthropic signed an agreement to lease more than 300 megawatts and 220,000 H100 GPUs from the Memphis data centre that xAI built for its Grok models, according to Fortune's May 8 report. The arrangement was not a sign of disarray; it was the contingency Amodei had described as the only honest hedge against forecast error.
The February warning also named compute availability, not capital, as the binding constraint on frontier AI development. Anthropic's April deal with Broadcom to access Google TPU chips, confirmed by Bloomberg on April 6, reflects the same logic: diversify supply before demand forces your hand.
80x Growth and the Claude Code Catalyst
At a CNBC appearance on May 6, Amodei stated that Anthropic's revenue had grown 80-fold in the first quarter of 2026. Bloomberg had documented the pace in real time: $9 billion in annualised revenue at the end of 2025, rising to $20 billion by March and $30 billion by April, with $47 billion in annualised revenue by May. The compute-efficiency strategy Daniela Amodei had described publicly contributed to improving margins: compute cost fell from 71 cents per dollar of revenue to 56 cents between early and mid-2026.