Anthropic has filed its S-1 prospectus, and the document reads like two companies occupying the same body. One is a hypergrowth enterprise software business generating $16.3bn in revenue in the first half of 2026 alone, up from $4.585bn for all of 2025. The other is a money furnace that lost $80bn in H1 2026 — though $72.4bn of that was a one-off charge — and $33.9bn in 2025, spending nearly $9bn on AI infrastructure in six months. The gap between these two realities is where the investment thesis lives or dies. The filing's most revealing metric is concentration risk. Anthropic's top five customers account for 25% of revenue, and the company's entire infrastructure footprint is concentrated in the US. This is a business built on a handful of enterprise relationships and a single country's power grid, regulatory regime, and immigration policy. Lose a major client or face a datacenter moratorium and the revenue picture shifts fast. Anthropic's executives paint the upside in maximalist terms — "a country of geniuses in a datacenter" capable of revolutionizing biology, medicine, finance, and military applications. The inclusion of military capabilities is notable given the company's ongoing standoff with the Trump administration over restrictions on Department of Defense use of Claude. The company is simultaneously positioning itself as a responsible AI lab that refuses to enable autonomous lethal weapons and as a vendor whose technology could transform military and intelligence capabilities. The safety disclosures are where the document gets genuinely uncomfortable. Anthropic states plainly that its most capable models can provide "meaningful assistance to malicious actors seeking to develop chemical, biological, or other dangerous weapons" and that its safeguards "are incomplete" and "have been and may in the future again be circumvented." The company also discloses what it believes is the first documented large-scale AI-enabled cyberattack with minimal human intervention — a Chinese state-sponsored group allegedly using a Claude code tool to target roughly 30 global organizations. The 92 pages of risk factors constitute an unusually candid inventory of fragility. Key person dependency on top executives. Regulatory uncertainty around AI and datacenters. The absence of adequate insurance for AI-related liabilities. Immigration policy threatening talent pipelines. The public benefit corporation structure potentially constraining financial optimization. Each risk factor individually is manageable; collectively they describe a company whose moat is entirely dependent on external conditions it cannot control. Anthropic's valuation — somewhere between $1.5tn and $2tn by some investor estimates, though the S-1 lists no target — prices in a future where compute costs decline, enterprise adoption accelerates, and catastrophic AI incidents do not occur. The filing's honesty about how many things must go right for that future to materialize is, paradoxically, both its strongest selling point to sophisticated investors and the clearest signal that frontier AI remains a bet on trajectory rather than fundamentals. The structural question the filing cannot answer is whether the AI industry's economics will ever normalize. Nearly $9bn in infrastructure spend in six months is not a growth investment if the compute treadmill never stops. Anthropic is asking public market investors to fund the most capital-intensive technology race in history while acknowledging that its own products could enable catastrophic harm it cannot prevent.
