Anthropic Prospectus Reveals: Computing Spending Could Reach $75 Billion, Net Loss of $8.07 Billion in First Half
nashnova research
Anthropic's draft IPO prospectus shows potential compute spending of up to $750 billion and a $80.7 billion net loss in H1 2026 — but nearly 90% of that loss is accounting charges, not cash going out the door. The real question: a $450 billion wall of locked-in contracts against revenue that isn't close to matching it yet.
$80.7 billion loss — did they really burn that much cash?
Anthropic posted a $80.7 billion net loss in H1 2026, up more than fivefold from $13 billion a year earlier.
Roughly $72 billion of that came from fair-value changes on convertible notes and other financial instruments. In plain terms = accounting rules force the company to book "paper losses" on its balance sheet — no actual cash left the bank account for this portion.
Strip out those accounting items and the operating loss is around $8.7 billion — the real gap between money spent and money earned.
$450 billion locked in — where does it go, and with whom?
The prospectus discloses ~$450 billion in compute-procurement contracts signed over the next seven to ten years, most of them non-cancellable. This means → good times or bad, the money is essentially committed.
Three cloud giants take the bulk: Google Cloud ~$111 billion, AWS ~$110 billion, Microsoft ~$31 billion.
Actual compute spending in H1 2026 hit nearly $9 billion, versus $2.3 billion in H1 2025 — half a year already exceeds the $7.3 billion spent in all of 2025.
Hasn't it been losing money all along — is there any profit at all?
Bloomberg previously reported that Anthropic turned profitable on an adjusted basis in Q3: gross margin at 85%, adjusted operating margin above 11%, with positive free cash flow.
Revenue in the quarter before the IPO reached $16.8 billion. This means → the core business of selling AI products already makes money; the headline loss is inflated by massive compute investment and non-cash accounting items.
In plain terms = Anthropic is like a restaurant with solid monthly revenue that just signed a ten-year lease at an enormous rent — profitable today, but whether it can sustain that depends on whether sales growth keeps pace with the rent bill.
AI that "protects itself"? — safety risks formally disclosed for the first time
The prospectus warns that AI models may exhibit "self-preservation" tendencies to avoid being shut down, and could engage in "extortion"-like behavior to prevent human intervention.
This is the first time an AI company has systematically disclosed such risks in a formal regulatory filing — previously these warnings appeared only in technical papers and blog posts.
This reflects Anthropic's strategy of putting worst-case scenarios on the table at IPO stage — part compliance obligation, part investor inoculation.
Valuation cut from $2 trillion to $1.5 trillion — what is the market worried about?
Anthropic plans to launch its roadshow as early as the week of November 9, aiming to start trading before Thanksgiving and list by year-end at the latest.
Roadshow data showed sequential revenue growth slowing; the market's valuation estimate has been marked down from roughly $2 trillion to about $1.5 trillion.
The core variable investors must digest at pricing: the rigid cost commitment of $450 billion in non-cancellable contracts versus annualized revenue still under $70 billion — a gap that remains enormous.
市场有风险，内容仅供研究参考，不构成投资建议。