The AI Moats Are Now Measured in Megawatts

One morning, Anthropic launched a model that matches its flagship's performance at 40% lower operating cost. Ninety minutes later, OpenAI answered with new models at half the API price. ^[1] The whole thing fit inside a lunch break. Frontier AI repriced itself in ninety minutes.

Improvements in caching and inference let us serve these models at lower cost, and we're passing those savings directly on to users and customers. — OpenAI

The pressure had a name before the discount did: Chinese developers have been shipping open-weight models that close the gap with the American frontier, undercutting the API and subscription revenue the labs were built on. ^[2] Morgan Stanley's read on what follows is supermarket logic — the cheap model is the milk at the front of the store, priced to pull customers in so they buy everything around it: the storage, the security, the managed services, where the margin actually lives. The model becomes the loss leader and the profit migrates to the infrastructure underneath it, which is why Amazon and Alphabet are already defending the store with chips of their own. ^[3] Cisco shows what the customer is doing with the discount. It carries $17.17 billion in inventory purchase commitments for AI infrastructure, and in the same breath it ships a tool called Tokenomics that compares models by cost against quality and lets customers switch to the cheaper one. ^[4] A company betting billions on physical capacity is also selling the means to treat the model layer as interchangeable. Cheaper tokens, meanwhile, get used more, so volume keeps climbing even as the margin on each one dies. ^[5] None of this is joined to the next part of the story by anything a company has said out loud. The price cut and the groundbreakings are two ends of one observable shift, not a stated cause and effect. But read the units the big deals are now written in, and the shift is unmistakable. In Ohio, SoftBank is building what it calls the world's largest AI campus — $1.5 trillion on a Cold War uranium-enrichment site the U.S. Energy Department is spending billions to scrub clean, powered by a $33 billion gas plant and 18 planned small nuclear reactors. ^[6] Uranium scrubbed for chips. In France, SoftBank pledged up to €75 billion for 5 gigawatts, a choice Son tied to the country's nuclear fleet. ^[7] His logic ran wider.

We are clearly in the process of closing the gap we had in terms of computing capacity in Europe. — Emmanuel Macron

India is buying the layer with the tax code — holidays that run to 2047 and concessional financing, aimed at 6.5 gigawatts by 2030 — and Adani has committed $100 billion to a 5-gigawatt platform. ^[8] A European sovereign-AI market is now estimated at 40 gigawatts ^[9], and the vendors are already selling into it: NVIDIA offers "sovereign environments" as a product line — sovereignty itself, offered as a SKU. ^[10] The state, meanwhile, has stopped merely permitting this and started angling to own a piece of it. Washington is exploring taking direct equity in OpenAI and Anthropic — funding for shares, even taxes payable in stock — though the exploration has not moved past the exploration stage. ^[11] And the people who already own compute are hoarding it: SpaceX keeps roughly half its capacity for its own projects and refuses long-term rental contracts, pricing the rest at spot. ^[12] Now read the same catalog against the grain. Exelon, a utility that has to actually connect these projects, went through its own pipeline and struck 18 gigawatts of "high-probability" data-center demand down to 11, canceling a $20 billion agreement near Joliet. ^[13] Its CFO explained what changed.

Even at the highest allowed price, the market is not attracting the level of new supply the system needs. — Calvin Butler

Goldman Sachs reckons only about half of planned capacity may actually get built, and Oracle declared force majeure on a New Mexico project after regulators blocked its gas pipeline. ^[14] Kenya shows the gap from the buyer's side. Microsoft and G42 announced a $1 billion campus — the largest private digital investment in the country's history — and never broke ground, because the guaranteed demand the hyperscalers needed was more than the grid could carry. ^[15] President Ruto put the arithmetic plainly.

We would need to switch off half the country for the data center to be powered. — William Ruto

The money underneath the buildout is starting to price the doubt. Private credit defaults in the sector are creeping toward 5%, and a co-CEO at one of the lenders says the terms never compensated for the risk. ^[16]

the equity captures the upside while the credit has the downside. — Mysterio

In the physical economy, the boom is already measurable — data-center construction is up more than 60% year over year while the rest of construction shrinks. ^[17] Here is the asymmetry at the bottom of all of it. The uranium site, the French gigawatts, the Indian tax holidays all exist on paper — in contracts, pledges and public commitments. They are real and transacted. What nobody has produced is the one ledger underneath all of them: verified demand. The only people who have actually counted it are the ones who must build to it, and their first count struck 40% of it as speculative. The demand curve is being written somewhere it cannot be spun — in the utility connection queues and the default rates — and the direction it has moved so far is down.