AI labs can turn a profit, but Anthropic shows the gap to close
Anthropic’s reported 2025 results show rapid revenue growth alongside larger operating losses, leaving AI-lab profitability unproven.
By Sofia Marchetti · Columnist
· 3 min read
Can AI labs turn a profit? The available evidence says they can in principle, but a durable route has not yet been demonstrated. For investors, Anthropic’s reported 2025 figures show why: revenue grew quickly, while the cost of computing and infrastructure remained higher still.
Reuters, citing an Anthropic IPO prospectus it reviewed, reported that the company generated nearly $4.6 billion in revenue in 2025. It also spent $7.33 billion on compute and infrastructure, part of $12.65 billion in total operating expenses, and recorded an $8.06 billion operating loss. Anthropic declined to comment to Reuters.
A limited comparison puts the current challenge in plain terms. Dividing reported revenue by compute and infrastructure spending produces a ratio of about 0.63. In other words, that spending category alone was roughly 1.6 times revenue. This is not a gross-margin or cash-flow calculation, and it cannot determine long-term economics because it excludes other details about the business. It does show the scale of the gap in the reported year.
What would let AI labs turn a profit?
Profitability would require revenue and usage to rise faster than the cost of serving customers, costs to fall sufficiently, or some combination of both. Reuters reported that AI companies expect broader adoption of large language models, or LLMs, to increase spending on tokens.
Price trends complicate that case. Reuters reported that Silicon Data’s LLM Token Expenditure Index, which estimates average token prices, had fallen more than 40% since June 30. Lower prices could encourage enough added usage to lift overall spending, a possibility often described as Jevons’ paradox. Reuters said there was little evidence at the time that this was happening. It also cautioned that the index is new, may miss parts of the market and does not separate individual models.
Customer concentration adds another variable. Reuters reported that almost one-quarter of Anthropic’s 2025 revenue came from two customers. Its prospectus also warned that many of its biggest clients did not have long-term contracts and could reduce or end spending.
Why the $42 billion net loss needs context
Reuters reported a nearly $42 billion net loss for Anthropic in 2025, but said roughly $34 billion of that figure was an accounting charge tied to an increase in the estimated value of financing that could convert into company shares. Reuters said that charge was not money spent operating the business. The $8.06 billion operating loss is therefore the more useful of the two reported figures for examining the company’s annual operating performance, though neither figure establishes a future profit path.
One proposed alternative is to use advanced models inside businesses owned by, partnered with or affiliated with an AI lab, rather than selling access broadly. In an essay, Marcus Abramovitch suggested areas such as quantitative trading and pharmaceuticals as possible applications. That remains a hypothesis, not evidence that AI labs have achieved profitability through internal deployment.
For now, the relevant test is straightforward: whether AI labs can convert adoption into enough revenue to cover the large and continuing costs of compute, infrastructure and other operations. Anthropic’s reported figures show strong growth, but not that the test has been passed.
This story draws on original reporting from Klement on Investing.
