Anthropic's $80.7 Billion First-Half Loss Is Mostly a Mark on Amazon's Own Paper. The $518 Billion It Owes Is the Number That Bites

Anthropic's first-half 2026 net loss of $80.7 billion is not what the headline implies. About $72 billion of it comes from revaluing convertible notes and similar financial instruments, the very paper that financed the compute buildout, leaving roughly $8.7 billion of loss to be paid for in cash.

The charge has a mirror image, and it sits on a public balance sheet. Amazon marked its Anthropic convertible notes at $42.2 billion on March 31 and $97.9 billion on June 30, moving the position up $55.7 billion in a single quarter, alongside $92.5 billion of Anthropic nonvoting preferred stock. The total was $190.4 billion at June 30. What Anthropic books as a loss, Amazon books as an asset.

There is a twist that changes how both sides should be read. Amazon did not book most of the mark-up as earnings. Accumulated other comprehensive income rose $41.4 billion in the June quarter, and total shareholders' equity climbed $109.7 billion against $62.6 billion of quarterly net income. The gain that offsets Anthropic's write-down lands in Amazon's book value rather than its earnings, which is exactly why the loss looks violent at Anthropic and quiet at the shareholder.

Set the accounting aside and the prospectus is still a hard document, because this company's balance sheet is a payment schedule. Anthropic has committed to at least $518 billion of cloud, compute and infrastructure obligations over about a decade, spread across six counterparties. Roughly 80% of that is non-cancelable or payable whether or not Anthropic uses a token of the capacity.

Spread across ten years, $518 billion is about $52 billion a year. Anthropic ended 2025 with $20.28 billion of cash and short-term investments, and it burned roughly $8.7 billion in the first half of 2026. The gap between one year of payments and the balance sheet is why the company is going public at all: a $100 billion raise at a $2 trillion mark funds roughly two years of the take-or-pay schedule.

The first claim on Anthropic's cash is upstream, and it is the one place the loss is unambiguously good news. NVIDIA reported $96.2 billion of revenue in its fiscal second quarter, up 106% year over year at a 75.0% gross margin, and $303.0 billion across the trailing twelve months. Broadcom's $161.2 billion of Anthropic equipment leases are worth about 1.8 times Broadcom's own $89.1 billion of trailing revenue. Neither company is funding Anthropic's losses; they collect the cash that the losses consume.

The second claim is where the loss turns into a credit event. CoreWeave, which signed a multi-year Anthropic agreement in April 2026, ended the second quarter with $51.6 billion of total debt, up from $35.1 billion three months earlier, against $5.0 billion of equity and $45.2 billion of net debt on $3.9 billion of trailing EBITDA. Over the trailing year it generated $6.9 billion of operating cash flow, spent $20.6 billion on capital expenditure, and closed the $13.7 billion gap with $16.8 billion of net new debt. Oracle, which is not a neocloud but has borrowed heavily to build data centers, raised $43 billion of debt in fiscal 2026 alone.

The operating leverage is real, and it is the best news in the filing. Compute and infrastructure spend fell from 209% of revenue in the first half of 2025 to about 160% for full-year 2025 and roughly 55% in the first half of 2026, cutting compute intensity by two-thirds inside twelve months. Anthropic was cash-flow positive in the third quarter and told investors it was profitable on an adjusted basis, the second straight quarter of adjusted profit after the first in the second.

The problem is the shape of the growth, not its level. Revenue rose 145% sequentially in the second quarter and 45% in the third. Delivering the $190 billion to $200 billion of 2028 revenue the company has shown investors means carrying a run rate of roughly $67 billion after the third quarter to about $190 billion within six quarters, which works out to sustaining roughly 19% sequential growth every quarter for a year and a half. A $2 trillion valuation is about 31 times the late-July run rate, roughly 435 times 2025 revenue, and 10 to 10.5 times the company's own 2028 projection.

Anthropic's revenue is not collected from end users the way a software company's is. About 47% of 2025 sales, some $2.2 billion, ran through Amazon's and Google's cloud marketplaces, and around one-fifth of first-half 2026 revenue came from just two customers. The same companies are Anthropic's largest landlords, its largest shareholders, the channels through which Claude is sold, and the holders of the paper that produced the $72 billion charge. The concentration is the reason the loss and the gain can be the same event.

In the next quarter only two things matter: the November listing and the first public print. If the IPO prices above $2 trillion, the convertible notes convert at a mark that validates the $72 billion charge and the funding gap closes in one step. If it prices below, Anthropic faces the same $52 billion-a-year schedule with less capital behind it, and the neocloud lenders, not the model, become the variable.

Over one to three years the milestones are the 2028 revenue number, the 2033 and 2036 payment cliffs, and the credit spreads that fund the capacity in between. The specific danger is not demand. It is that Anthropic has locked in a decade of capacity at 2026 prices while inference prices keep falling, and roughly 80% of that book cannot be cancelled. If the 2028 revenue arrives at a gross margin below the roughly 92% API gross margin that SemiAnalysis has estimated, the take-or-pay commitments turn from a growth asset into a fixed cost with no exit.

None of this makes Anthropic a bad business. Revenue went from $1.1 billion in the first half of 2025 to $16.3 billion in the first half of 2026, compute intensity fell by two-thirds, and the third quarter was cash-flow positive. It makes the document a different kind of document than the headline implies. The $80.7 billion loss is a price tag on the IPO rather than a measure of the business, the $8.7 billion is the real cost of running the model, and the $518 billion payment schedule is the actual risk being priced.