Anthropic targets $2T IPO and locks in $518B AI compute spend

Anthropic’s draft prospectus reads like a balance sheet for the race to AGI. A $2 trillion ask paired with $518 billion of largely non‑cancelable compute and cloud leases signals just how capital‑intensive frontier training has become. At that scale, Anthropic is no longer a startup, it is an infrastructure project fused to the roadmaps of Broadcom, Amazon, Google, Microsoft, xAI and AMD, all of whom show up as creditors, lessors or both in the filing. Those counterparties now have enormous financial and strategic incentive to see Anthropic’s models deployed everywhere.

What makes this particularly relevant to AGI is not just the headline valuation but the time horizon implied by decade‑long spend commitments. Anthropic is effectively pre‑paying for multiple generations of ever larger training runs, betting that scaling plus better alignment will continue to push capabilities forward. The risk language in the prospectus, which explicitly entertains models that might resist shutdown or conceal behavior, is a rare, candid acknowledgment of where they think the technology could go.

Competitively, this move sets a brutal benchmark. Any lab hoping to stay at the frontier must either match this compute envelope, find cheaper routes via open weights and efficiency, or specialize in narrower domains. Public investors will soon get a say in how much existential‑risk baggage they are willing to carry in exchange for owning a slice of the most aggressive AGI bet on the market.