Anthropic's "historic profitability" and the $1.25 billion monthly bill that changes everything
Anthropic's first operating profit made headlines — but it was built on a two-month discounted compute deal with xAI worth $15B a year. Here's the structural reality behind the narrative.
Last quarter, Anthropic reported its first operating profit. The AI world applauded. The narrative machine ran at full speed.
Here's what the narrative left out.
That profitability was built on a two-month discounted rate. In May, Anthropic signed a deal with xAI — yes, Elon Musk's company — to access the entire Colossus 1 data center in Memphis: 220,000 Nvidia GPUs, 300 megawatts of raw compute power. The price tag, disclosed in SpaceX's SEC filing: $1.25 billion per month. $15 billion per year.
The first two months came at a reduced rate while xAI completed its infrastructure ramp-up.
Those two months are now over.
Why does compute cost this much?
Because frontier AI is, at its core, an energy and silicon business. 300 megawatts is the power draw of a mid-sized city. 220,000 of the world's most expensive chips, running continuously, cooled at scale, depreciated over years. There is no magic here — only physics and balance sheets.
Anthropic needed that capacity immediately. It was compute-constrained, users were hitting rate limits, Claude Code was throttled. xAI had Colossus 1 sitting underused while building Colossus 2. A deal was struck. Both sides got what they needed — but Anthropic got a clock, too.
The structural reality
Anthropic's annualized revenue has grown from $9 billion at end-2025 to over $30 billion by early April 2026. An extraordinary trajectory.
But $15 billion per year in compute from a single vendor — before AWS, before Google TPUs, before salaries and research — means the margin story is not what it appears.
The "historic profitability" of Q1 was real. It was also a two-month accounting window purchased at a discounted rate.
Q3 will be the honest number.
Our read at Seerius
This isn't a criticism of Anthropic's ambition. Building at this scale requires exactly this kind of infrastructure aggression. And the revenue growth is real — $9 billion to $30 billion annualized in under six months is not noise.
But step back and look at the architecture of the business.
Anthropic depends on xAI's data centers to train its models. It depends on AWS and Google to serve them. It depends on Nvidia's supply chain to exist at all. Every layer of the stack — compute, inference, silicon — is owned by someone else, priced by someone else, and can be repriced at renewal.
This is not a software company. It is a tenant in someone else's building, paying rent at a scale that makes real estate look quaint.
The narrative has been carefully constructed: safety-first, values-driven, the responsible path through the AI transition. We find that positioning genuinely more credible than OpenAI's. Dario Amodei is a serious operator with a serious research agenda. And yet — OpenAI, Anthropic, Google DeepMind, xAI — they are all running the same race, burning the same capital, dependent on the same Nvidia supply chain.
The difference isn't the business model. It's the communications department.
But credibility doesn't pay a $1.25 billion monthly invoice.
And here is the deeper issue: the entire frontier AI layer is engaged in a race where the finish line keeps moving. Each model generation requires more compute. More compute requires more capital. More capital requires more revenue. More revenue requires more users. More users require better models. The loop is closed, accelerating, and structurally unprofitable at the margin until someone blinks or consolidates.
What we build at Seerius sits one layer below this war — and that is intentional.
We invest in Swiss and European SMEs where AI is not a survival condition but an operational multiplier. Companies with real cash flows, tangible assets, and defensible market positions — where deploying the right AI tools compounds value without requiring a data center in Memphis.
Our Augmented Operator Program doesn't ask portfolio companies to win the compute race. It asks them to use its output intelligently: automating workflows, sharpening commercial execution, accelerating decision-making. The infrastructure bill belongs to Anthropic. The productivity gain belongs to our companies.
This is the arbitrage we see clearly.
Frontier AI labs are building the roads. We are helping businesses drive on them — faster, leaner, and without a $15 billion annual toll.
The compute wars are a VC sport. Operational transformation is an operator's game.
And in Europe, that game is just beginning.
— The Seerius Team