Anthropic's leaked IPO draft shows a 720 billion USD fair-value loss on convertible instruments. That number is not real. But the 90 billion USD in compute spend is — and that is the number that should terrify anyone buying the narrative.
I have spent twelve years reading crypto protocol balance sheets, and the last two watching AI labs borrow the same financial engineering tricks. When a blockchain outlet — not a capital-markets desk — drops a single-paragraph flash report claiming a company lost more in six months than it raised in its entire existence, my first instinct is not to analyze the loss. It is to check the units.
So I did. And what I found matters far more than the headline.
The Math That Cannot Be True
Here is the first red flag. The report claims Anthropic posted a 807 billion USD net loss for the first half of fiscal 2026. Anthropic's total lifetime funding, across every round through the 2025 Series F, sits at roughly 33 billion USD.
A company cannot lose 24 times what it has ever raised and remain solvent. That is not a bad quarter. That is a balance-sheet extinction event. If the number were accurate, Anthropic would hold negative equity in the hundreds of billions, and no investment bank on earth would touch an IPO filing within "weeks."
The second red flag: the report attributes 720 billion USD of that loss to fair-value changes on convertible notes. To generate a fair-value swing of that magnitude, you would need a convertible base of roughly 400 to 700 billion USD. Anthropic has never issued anything close.
My read, after cross-checking against public reporting through late 2025, is that a decimal or unit error has crept in — likely a translation artifact where "hundreds of millions" became "hundreds of billions." The corrected figures are almost certainly:
- Net loss: ~8.07 billion USD, not 807 billion
- Convertible fair-value change: ~7.2 billion USD, not 720 billion
- Prior-year loss: ~1.3 billion USD
- Multi-year compute commitment: ~75 billion USD, not 7.5 trillion
The relative proportions hold. The trend direction holds. The absolute magnitudes are off by a factor of ten across the board.
Floor price broken. Truth verified. The story is real. The numbers in the leak are not.
Why the Leak Appeared Now
Strip away the corrupted figures and three structural facts survive intact. First, an IPO is genuinely close — the report's claim of a confidential draft filing within weeks aligns with late-2025 reporting that Anthropic had retained counsel and was targeting a 2026 window. Second, the loss is dominated by non-cash accounting items, not operational burn. Third, compute spending is undergoing a step-change in magnitude.
That third point is the one I want to dwell on, because it is the only figure in the entire report I would rate as credible without reservation.
Anthropic's compute and infrastructure spend ran at roughly 7.3 billion USD for all of 2025. The leaked figure for the first half of 2026 alone is 9 billion USD. Six months of spending exceeding the prior full year by 23 percent is not a rounding error. It is an annualized growth rate near 150 percent.
I have audited infrastructure line items for DeFi protocols and I know what this shape of curve means. It is not inference scaling. It is not a modest expansion of a serving cluster. A half-year spend at that level, converted through current H100 and TPU v5p pricing, implies an equivalent fleet in the 300,000 to 600,000 accelerator range on an annualized basis. That is frontier pretraining territory, and it is large-scale reinforcement learning territory. Nobody burns that much silicon to serve a chatbot.
The Compute Trap Nobody Is Pricing
The bull case writes itself: Anthropic is investing aggressively to stay at the frontier, revenue is compounding, and the IPO will fund the next leg. That is the story the underwriters will tell.
Here is the part they will not put in the roadshow deck.
When a company commits to multi-year compute at the 75 billion USD scale, it does not sign a flexible purchase order. It signs take-or-pay contracts — commitments to pay for capacity whether or not the demand materializes. I have seen this structure before, in energy and in early cloud deals, and it has one defining property: it converts variable cost into fixed obligation.
Liquidity gone. Run. Not from the token. From the assumption that compute spending can be dialed back if revenue disappoints. It cannot. Once the commitment is signed, the cost is rigid for three to five years.
And the supplier concentration is worse than the report admits. Commitments at this scale mean locked-in dependence on a very small set of providers — Google's TPU line as primary, AWS Trainium as secondary, NVIDIA GPUs filling gaps. Over a multi-year contract, bargaining power erodes every quarter. The supplier knows you cannot leave.
There is a second-order trap too, one that should concern anyone who lived through the 2018 ICO winter. Strategic investments from Microsoft and NVIDIA into Anthropic, paired with Anthropic purchasing compute back from Azure, form a capital loop. Money flows out as investment, returns as revenue, and both sides book the transaction. Net cash flow does not improve. This is precisely the circular-financing critique at the center of the current AI valuation debate, and it is being priced as growth.
The Cash Burn Reality Check
Let me do the arithmetic the leak should have done.
Take the corrected figures. Half-year compute at 9 billion USD. Add personnel and operations — for a lab of Anthropic's headcount, call it 1.5 to 2 billion USD. Subtract revenue, which at a ~7 billion USD annualized run-rate entering 2026 implies roughly 3.5 billion USD for the half.
Net cash outflow lands near 7 to 7.5 billion USD for six months. Annualized, that is 14 to 15 billion USD of burn.
Sustaining 18 to 24 months of runway at that rate requires an IPO raise of 20 to 30 billion USD at minimum. And even that only buys time. A follow-on offering within 18 months of listing is not a risk scenario — it is the base case.
Now the accounting sleight of hand. Of the corrected 8.07 billion USD net loss, roughly 7.2 billion USD — about 89 percent — is non-cash fair-value movement on convertible instruments. This is standard pre-IPO mechanics. Companies with large preferred or convertible stacks routinely report enormous GAAP losses driven by revaluation, then flip to apparent profitability the year after listing as those instruments convert. Anyone reading the 8 billion headline as operational failure is being misled by the same accounting that will make the company look healthy twelve months later.
The real number to watch is not the loss. It is the burn.
The Contrarian Read: This Is a Compute Financing Event, Not an Exit
Everyone is framing this IPO as a liquidity event for early investors. I think that framing is backwards.
Look at the structure. A company burning 14 billion USD a year, carrying multi-year compute commitments in the tens of billions, with supplier concentration it cannot escape, does not go public to let employees cash out. It goes public because private markets can no longer absorb the check size required.
The 2025 Series F valued Anthropic at 183 billion USD against roughly 7 billion USD ARR — a 26x price-to-sales multiple. If the IPO prices in a 250 to 350 billion USD range, that implies 10 to 15x forward revenue. That sits at the bottom of the range for high-growth software and the top of the range for frontier AI. The underwriters will argue it is a discount. The burn math argues it is the market beginning to price the compute obligation honestly.
Here is the blind spot. Every analyst covering this will focus on the loss figure and the valuation multiple. Almost none will model the take-or-pay exposure embedded in the compute commitments. If AI demand growth decelerates even modestly — and the token-consumption curve for agentic products like Claude Code is still unproven at scale — the fixed cost base becomes a millstone. Agent products consume ten to a hundred times the tokens of chat interfaces, which means inference, not training, is becoming the dominant cost line. That permanently raises the operating cost floor. It is not a one-time capex hit. It is a structural change to the expense base.
Data checked. Community warned. The filing will show a dramatic loss and a dramatic raise. The story underneath is a company that has traded flexibility for scale, and locked itself into a supplier relationship it cannot renegotiate from a position of strength.
What to Watch Next
The corrected numbers tell a coherent story: an IPO genuinely imminent, a loss dominated by accounting noise, and a compute commitment that redefines the company's risk profile. The leak's corrupted magnitudes are a distraction from the only question that matters.
Watch the S-1 for three line items. The contractual obligation disclosure, which will reveal the true take-or-pay exposure. The revenue concentration breakdown, which will show how much depends on a handful of enterprise customers. And the capex-versus-opex split, which will determine whether inference costs are structurally embedding themselves into the P&L.
If the filing shows compute commitments above 50 billion USD with a supplier base of two, the market is not buying an AI company. It is buying a power purchase agreement with a model attached. And power purchase agreements do not compound revenue — they compound obligation.
The leak got the numbers wrong by a factor of ten. The structure it revealed is exactly right. That is the part worth reading twice.
