On July 4, a widely-circulated post by @AYi_AInotes highlighted a line buried in SpaceX’s revised IPO filings: SpaceX is supplying Anthropic with $1.25 billion of compute every month, under a contract running through May 2029, with either side able to terminate on 90 days notice. The same window saw reports that Anthropic has locked in 1.4 GW of data-center capacity in Australia, with $15 billion in build-out cost.

Read these three together and the story is not “Anthropic bought more compute.” It is that AI compute has quietly moved from being a cloud resource you rent by the token, to a piece of industrial infrastructure you buy on multi-year fixed contracts. In industrial-age language this is called a power purchase agreement (PPA). In the AI era it is still the same shape: a customer locks in capacity at a fixed price, the supplier locks in cash flow to build. I am calling this “compute as power plant,” and that is the thesis of this post.

1. $1.25 billion a month: why this is not SpaceX doing cloud

SpaceX is not a cloud vendor. AWS, Azure, and GCP sell you compute, storage, networking, SDKs, ops, compliance, billed by the token or by the vCPU-hour. That is cloud.

What SpaceX is selling is merchant compute. They run AI compute for their own internal needs (X, Starlink ops, internal R&D), and when they have surplus capacity, they sell it. The contract format is a fixed-monthly-fee, fixed-capacity, terminable-on-90-days. That is not a SaaS contract. That is the shape of a power purchase agreement — the same one used for electricity in regulated markets.

When industrial users sign PPAs, they are doing two things at once: locking in price and locking in capacity. The downstream buyer gets predictability; the upstream supplier gets cash flow they can take to a bank to finance a power plant. AI companies just started signing PPAs for GPU clusters the same way, and that means GPUs are now being financed, built, and locked like power plants, not sold like cloud.

The risk SpaceX is taking on is not “tenant churn on a managed service.” It is “what if our GPU fleet underperforms and we still owe the capacity.” That is a generation-asset risk, not a cloud-business risk.

2. Anthropic’s 1.4 GW: the buyer becoming its own power company

Anthropic’s Australian move looks like a purchase. It is actually a self-build. 1.4 GW maps, on GPU-cluster math, to roughly 700k–1.4 million H100-equivalents — a fixed-asset commitment in the trillion-yuan range at today’s spot prices.

The $15 billion is build cost, which means Anthropic is buying land, transformers, cooling, long-dated power, construction crews. Anthropic just walked across the line from customer to utility — from “I rent power” to “I own a plant.”

Putting the two together: SpaceX’s $1.25B/month contract is “power plant selling power to industrial user.” Anthropic in Australia is “industrial user starting to build its own plants.” The AI industry’s supply topology is flipping from centralized cloud + many small customers, to a few large actors who lock in capacity on multi-year terms — and everyone else still buying the spot market.

This is not the first time this inversion happened. US wholesale electricity markets did the same thing in the early 2000s: a wave of large industrial users (aluminum smelters, then data centers, then crypto mines) signed PPAs, then started building their own generation. AI is at the same fork today.

3. OpenAI’s “Stargate UK” was a $30B fiction

As a contrast: on July 4, The Guardian revealed that OpenAI never set foot on the Crown Estate Cobalt Business Park, never met the local council, and that of the £30B the UK government had been advertising as “potential investment,” £20B was just construction cost estimate, not committed money. The whole thing was rolled out in time for Trump’s UK visit.

Compared to SpaceX and Anthropic’s paper — IPO filings, sealed bid documents, electricity long-term agreements — OpenAI’s UK project is press conference + policy theater without a real PPA behind it.

The simple filter for builders: who is putting real cash on the table for a real PPA, and who is announcing $30B at a press conference? If the latter, treat it like vaporware.

4. What this means for you

If you are at a frontier-model company’s finance or operations team, the next 12 months change three things:

  1. Re-read your GPU procurement contracts. Check whether your supplier has quietly moved from token billing to monthly fixed contracts. If not, push for a PPA proposal — if you don’t sign, your competitor will, and they’ll have locked out the next 18 months of capacity.
  2. Move “power long-dated agreements” onto your balance sheet, not under “cloud services.” Most finance teams today have no clean line item for this; expect a Q3 switch in how this shows up.
  3. Watch non-traditional compute suppliers. SpaceX is not the only one. Power utilities with surplus hydro, steel mills with captive electricity, data-center REITs — all of them will try to bypass hyperscalers and sell compute directly in the next 12 months.
  4. Plan compute, not just tokens. The question you used to ask (“can AWS give me a better rate?”) is being replaced by (“can NVIDIA / TSMC / a local utility give me a PPA?”). That’s a different relationship graph.

If you are an AI founder outside the frontier-model tier, the answer is simpler and unchanged: rent by the token. PPAs are rational only when energy is more than ~25% of your cost base. Below that, spot is fine. Don’t sign a 10-year power contract because it’s fashionable — that’s how you end up the next WeWork.

5. The thing I almost left out

Compute as a PPA is not a small bump. It is a load-bearing piece of infrastructure, on the same maturity curve as wholesale electricity in the 1990s and fiber backhaul in the 2010s. Once two flagship customers have signed, the rest of the industry will follow within 18 months — not because it is fashionable, but because the alternative is being shut out of capacity at peak.

The thesis to watch over the next year is not “who announced another data center.” It is “who has signed a real PPA, with paper behind it.”

References