Anthropic is spending at an extraordinary scale. The AI company has warned investors that its models could resist shutdowns and cause catastrophic harm, while its financial figures show a business pouring billions into computing resources. That combination makes the company’s trajectory difficult to miss—and harder to describe as ordinary software growth.
Anthropic recorded an operating loss of more than $8 billion last year as it increased spending on computing power. Its revenue jumped 12-fold to almost $4.6 billion, but mounting costs for resources used to train and run its models pushed operating expenses to almost $13 billion.
Revenue Is Rising Faster Than the Bill
The gap between revenue and spending defines Anthropic’s current position. The company generated almost $4.6 billion in revenue last year, yet its operating expenses reached almost $13 billion, leaving an operating loss of more than $8 billion.
That arithmetic is the less glamorous side of frontier AI. Every improvement in model capability demands more computing resources, and Anthropic’s figures show that the cost of building and running those systems remains enormous—even after revenue multiplied 12-fold.
The latest figures show a business moving toward profitability, at least by one accounting measure. Revenue in the second quarter of this year was $11.5 billion, and Anthropic is on course for its second consecutive quarter of operating profit, albeit on an adjusted basis.
An estimated annual revenue of $60 billion for 2026 points to a much larger business than the one reflected in last year’s results. But revenue forecasts do not erase the cost structure; they underline how much scale Anthropic needs to support its models.
Compute Commitments Meet Model Risk
Anthropic has signed $517 billion in compute agreements over the past 11 months. That commitment dwarfs the company’s almost $4.6 billion revenue last year and sits beside operating expenses that already approached $13 billion.
The agreements show that Anthropic is preparing for demand that requires vast computing capacity. They also make the company’s safety warnings harder to separate from its financial strategy: the systems it says could resist shutdowns and cause catastrophic harm are the same class of systems driving its need for more compute.
That does not make the warnings less credible. It makes them more consequential. Anthropic is not describing a distant research concern detached from its business; it is discussing risks inside a company expanding model development, revenue, and infrastructure commitments at the same time.
The result is a stark bargain. Anthropic is chasing revenue that reached $11.5 billion in the second quarter of this year and may reach an estimated $60 billion annually in 2026, while committing $517 billion to compute agreements and acknowledging the possibility of catastrophic model behavior.
Adjusted operating profit may offer a cleaner near-term headline, but the larger picture remains expensive and unsettled. Anthropic’s numbers describe a company racing toward commercial scale while its own safety disclosures warn that control cannot be treated as a footnote.
Based on
