Anthropic Gets $5B From Amazon, Promises to Spend $100B Back

Look at the numbers cold and something feels off. Amazon just put another $5 billion into Anthropic. Anthropic, in turn, committed to spending $100 billion on AWS compute. Twenty dollars out for every dollar in. And somehow, this has become the most common deal structure in AI in 2026.

The $5B-in, $100B-out Math

The mechanics are simple. Amazon tops up its Anthropic stake with an additional $5 billion. Anthropic signs a long-term agreement to consume $100 billion worth of AWS compute — and commits to running its frontier models primarily on Amazon’s in-house Trainium chips rather than Nvidia silicon.

On paper, it’s mutual. Anthropic locks in the GPU and custom-silicon capacity it needs to keep training and serving models. Amazon locks in a multi-year revenue stream for AWS and gets a marquee customer validating its own AI accelerators.

Dig a layer deeper and the symmetry starts looking strange.

This Is Circular Financing

The phrase making the rounds on Hacker News and in analyst notes is circular financing. The pattern:

Company A invests in Company B. Company B spends that investment buying Company A’s product. Company A’s revenue jumps. Company A’s stock goes up. Company A raises more capital. Repeat.

Nvidia committing $100 billion to OpenAI, which turns around and buys Nvidia GPUs. Microsoft pouring tens of billions into OpenAI, paid largely in Azure credits. Now Amazon and Anthropic. Same loop, different logos.

The uncomfortable part: the amount of actual cash moving through these deals is far smaller than the headline numbers suggest. The money essentially circulates between two sides of the same ledger.

What Amazon Actually Wants

So why structure it this way? Three motives line up neatly.

Defending AWS’s narrative. While Azure has ridden the OpenAI workload to blockbuster growth, AWS has been painted as the laggard of the AI boom. A $100 billion Anthropic commitment rewrites that story in one press release.

A Trainium reference customer. Having a frontier lab train production models on Amazon’s own chips is the single most valuable marketing asset Amazon could buy in its campaign to chip away at Nvidia dependence.

Option value on AGI. If Anthropic actually builds something transformative, Amazon — as a major equity holder — gets front-row seats and a claim on the upside.

Real Demand or Dot-Com Redux

This is where the room divides.

Optimists point to real pull-through. Anthropic’s API revenue is growing fast. Enterprises are paying actual money for Claude. Circular or not, if there’s genuine end-user demand at the edge of the loop, the loop holds.

Skeptics reach for a less flattering analogy: vendor financing circa 1999. Cisco and Lucent lent money to telecom customers who promptly bought their equipment. Revenues ballooned on paper. Then the customers went under, and the dominoes followed.

The arithmetic invites the comparison. Anthropic’s annualized revenue is estimated somewhere around $10 billion. A $100 billion consumption commitment on top of that is not a rounding error — it’s an order of magnitude.

The Numbers That Actually Matter

Whether this structure is substance or bubble will reveal itself through a few signals. Anthropic’s paid-user revenue growth. Stability of API pricing. Enterprise lock-in and net retention. What matters is not the announced contract value but what end customers are actually paying at the end of the chain.

One thing is clear: AI infrastructure in 2026 has drifted past ordinary capitalism. A handful of hyperscalers and a handful of labs are inflating each other’s ledgers in a multi-hundred-billion-dollar game of pass-the-parcel. Telling the companies creating real value from the ones just moving numbers around has never mattered more.

If you saw a $100 billion compute contract hit the wire tomorrow — would you applaud, or would you ask where the cash is?