AI's New Power Brokers Don't Make AI

The price of AI intelligence is collapsing while the price of housing it is exploding, and the gap between those two lines is where the race has actually moved. On the intelligence side, the story is almost boring now. OpenAI cut fees on its Luna model by 80%, and Anthropic halved its own rates to counter Chinese rivals ^[1]. Hugging Face's download data shows 83% of all models pulled are under a billion parameters — the workhorse models, not the frontier ones ^[2]. Microsoft, which should know, now concedes that large language models are becoming commoditized ^[3]. The thing everyone was racing to build is turning into a cheap input. The thing that is not cheap is the building itself. PJM, the grid operator for the mid-Atlantic, saw capacity prices jump more than 1,000% in a year ^[4]. The average wait to connect a new project to the grid has stretched from 36 months to 61 ^[5]. Microsoft — the same company conceding its models are commodities — is carrying $175 billion in capital spending and $329 billion in lease commitments, more than half a trillion dollars pointed at physical assets ^[3]. Intelligence is getting cheap; matter is getting expensive. And the companies winning the matter race are not the ones you would have filed under "AI" six months ago. Bitcoin miners, who already hold power contracts and substations, are converting them. CleanSpark signed a $6.6 billion, 20-year lease for 175 megawatts, and its CEO, Matt Schultz, described the pivot in his own terms.

This lease is a transformational moment for CleanSpark as we complete our evolution into a diversified digital infrastructure platform and begin monetizing our power portfolio at institutional scale. — Matt Schultz

Ionic Digital, a miner that crawled out of the Celsius bankruptcy, is pivoting to a 234-megawatt AI lease with an 89-megawatt expansion ^[6]. Oil companies are building private gas plants for data centers. Chevron, ExxonMobil, Diamondback and Pacifico are developing "power islands" in Texas, where the grid operator projects 24 gigawatts of new data center demand by 2031 — roughly the peak load of the Houston metro area ^[7]. Chevron has already signed a 20-year deal with Microsoft for a 2.67-gigawatt project ^[8]. A contract manufacturer, Flex, is spinning off its cloud and power infrastructure unit as a standalone company next year, with more than 90% of its projected revenue already booked ^[9]. And in Europe, the beneficiaries are industrial firms most people have never heard of — Atlas Copco's vacuum pumps, Alfa Laval's heat exchangers, Air Liquide's specialty gases — whose cooling and gas businesses are suddenly being revalued as AI infrastructure ^[10]. None of these companies makes a model. That is the point. The reason they matter is that the grid itself is being routed around rather than fixed. Google's head of energy calls transmission the number one challenge, citing one utility that quoted 12 years just to complete an interconnection study ^[11]. So Google is colocating data centers next to power plants, skipping the transmission line entirely. Meta deployed more than 800 mobile mini-turbines in El Paso to keep things running, and OpenAI's data center developer is buying 29 jet-engine gas turbines from Boom Supersonic in a $1.25 billion deal ^[12]. There are 58 gigawatts of gas power in development in Texas alone ^[12]. The grid operator itself says data centers can be built faster than the generation needed to serve them ^[4]. Nations have noticed, and they are treating electricity as a strategic input alongside chips. China is offering up to 50% electricity subsidies to push domestic chips that are 30-50% less efficient than the Nvidia parts they cannot buy — the state paying the power bill to close the compute gap ^[13]. Japan is buying 27,500 GPUs as national infrastructure, and Jensen Huang made the stakes explicit.

The next frontier of AI is in the physical world, and this is a once-in-a-generation opportunity for Japan. — Jensen Huang

India took the other fork: it deliberately chose small, open-weight models that avoid the energy costs of massive models, betting on cheap deployment over frontier scale ^[14]. Even the remaining bottleneck is physical. Elon Musk says the limiting factor right now is memory ^[15]. Of 3,969 planned data centers, only 802 are under construction ^[8]. The question is no longer whose model is smartest. It is who can power and house it — and the companies that already held the power, the land, the cooling and the gas are the ones quietly winning a race that was supposed to be about software.