Anthropic's $80.7 Billion Loss Is Mostly Non-Cash. Its Compute Bill Is Not.
About $72 billion of the first-half loss came from revaluing convertible instruments, according to a draft IPO filing reviewed by Reuters. Compute and infrastructure spending, up about 291%, is the separate question.
Anthropic, the developer of Claude, recorded a net loss of $80.7 billion in the first half of 2026. That is larger than the annual revenue of many big global companies.
Why did the loss grow as the company's value rose?
According to a draft IPO registration statement reviewed by Reuters, Anthropic's first-half 2026 net loss was $80.7 billion. A year earlier, the loss for the same period was about $13 billion. It has grown more than sixfold.
About $72 billion of the total came from revaluing convertible bonds and related financial instruments.
A convertible bond can be exchanged for stock under set conditions. When the issuer's value rises, the value of the conversion right embedded in the bond can rise too.
The issue is accounting. When some convertible bonds or conversion-related liabilities are measured at fair value, a rise in that value lets the company book an expense for the larger liability on its income statement. The loss appears on paper even though no cash leaves the company. The treatment is not the same for every convertible. It depends on contract terms and accounting classification.
Anthropic's first-half net loss, by component
- Convertible bonds and related revaluation expense: $72 billion, or 89.2% of the total
- Remainder of the net loss: $8.7 billion, or 10.8% of the total
$80.7 billion minus $72 billion leaves $8.7 billion. That remainder is not the same as operating loss or cash burn.
One more distinction matters. A non-cash charge is not economically meaningless. If the instruments convert to shares, existing shareholders can be diluted. Depending on the contracts, a cash repayment obligation could also arise.
That no cash has gone out today does not mean there is no cost to future shareholders.
The real concern for AI companies is compute
Anthropic's first-half spending on compute and infrastructure was reportedly about $9 billion. That is up roughly 291% from about $2.3 billion in the same period a year earlier.
AI companies use enormous computing resources to train models. But training is not the only cost. Each time a customer asks Claude a question, tells it to write code or runs a complex task, inference compute costs are incurred.
If the cost of serving customers rises faster than revenue, profitability may not improve.
The key test for investors is not the simple growth rate of spending. It is this question: is the compute cost needed to earn $1 of revenue falling?
Is the $518 billion in future commitments all debt?
Reuters reported that Anthropic has planned or committed to spend at least $518 billion on cloud and compute infrastructure and similar items in the years ahead.
That figure should not be read as financial debt due immediately, or as investment already made. The financial burden depends on when long-term contracts are performed, whether they can be cancelled, minimum purchase obligations, prepayments and how the contracts link to revenue.
On the other hand, if customer demand is weaker than expected or price competition in AI services intensifies, contracts that locked in capacity early could become a burden.
For cloud providers such as Amazon, Google and Microsoft, large compute demand can be a revenue opportunity. For an AI model company, it is a cost that must be paid first in order to generate revenue.
Even within the same AI spending boom, the profit structures of infrastructure suppliers and model developers differ.
Insight Times Editorial Desk
