The AI Labs Are Buying Atoms

The AI industry's number of the quarter is a markdown: OpenAI cut fees on its Luna model by 80 percent, Anthropic brought out a high-performance model at half the cost of its top system, and Meta released a cheap one of its own ^[1]. The billing is being rebuilt to match. The industry is moving off software licenses and onto metered, utility-style rates, charging for each unit of work and shunting the routine jobs to whichever model is cheapest ^[2]. What does a company buy when the thing it sells gets cheap? The big banks have an answer, and it is: nothing in the software. Morgan Stanley expects open-weight models — AI systems published free for anyone to run — to keep grinding prices down; it names Meta's Muse its bellwether, and it reads cheap model access as a loss leader: bait for the profitable side of the cloud, where storage, security and managed services earn the money ^[3]. Bank of America, watching Chinese labs move the fight from the price of tokens to the price of a finished task, puts the long-run profit in chips, memory and large cloud platforms, and pointedly not in standalone model developers ^[4]. What makes this week unusual is what one lab went out and bought. In the five days from Sunday, September 13, to Friday, Anthropic signed the Danish drugmaker Novo Nordisk as a drug-discovery partner, bought a biotech called Coefficient Bio, opened a working biology laboratory in the San Francisco Bay Area, and filed for a $2 trillion Nasdaq listing that discloses $65 billion in annualized revenue and more than $500 billion already committed to computing capacity ^[5][6][7]. A company that sells intelligence by the token spent the week buying atoms. Start with the bench, the strangest line item. The new facility is a wet lab — real laboratory space, where experiments happen in liquids and living cells rather than on a screen — and the plan is for Claude, Anthropic's model, to run biology experiments there through robotic arms, microscopes and liquid handlers ^[6]. It did not appear from nowhere: three weeks earlier, Anthropic published a research preview of what it calls a Model Hardware Standard, a common language for its agents to drive exactly that equipment around the clock, with AWS, Danaher, Doosan Robotics, Hugging Face and Raspberry Pi already testing it ^[8]. The company also seated Vasant Narasimhan, chief executive of the Swiss drugmaker Novartis, on its board ^[6]. Anthropic's life-sciences lead was blunt about where the software's job ends.

We believe that to do biology, the final test is still and will be for a while in real lab work. — Eric Kauderer-abrams

That is an odd concession from a company whose product is software, and it happens to be the field's honest boundary. Stanford's AI virtual lab, which found a lung-cancer treatment approach in a single day for $46 of computing, states outright that its agents cannot perform experiments on cells ^[9]. Amazon hit the same wall this spring and rented past it: its Bio Discovery service routes the bench work through partners like Ginkgo Bioworks, while AWS sells the intelligence on top ^[10]. Anthropic bought the bench. The pitch has moved with the purchases. By Wednesday, Dario Amodei was selling Novo Nordisk on compressing a century's worth of medical breakthroughs into a decade, a promise he had called a deceptive cliché in August ^[5]. OpenAI's half of the inventory is electrons, and it has been accumulating for months rather than days. It is negotiating a claim on 12.5 percent of Helion Energy's fusion output, five gigawatts by 2030 scaling to fifty by 2035 ^[11]. Meanwhile the data centers themselves are being wired around the grid: behind-the-meter generation, meaning turbines built on-site and piped straight into the building, skipping the utility and its connection queues. The buildout includes 29 jet-engine turbines for OpenAI's Robinson Crusoe data centers, 800 mobile mini-turbines staged for Meta's El Paso site, and nearly 58 gigawatts of gas in development in Texas alone ^[12]. Altman himself treats the two — artificial intelligence and cheap energy — as twin cost-down projects.

My vision of the future ... is that if we can drive the cost intelligence and the cost of energy way, way down, the quality of life for all of us will increase incredibly. — Sam Altman

The supplier list runs the same direction. SpaceX has turned its Colossus supercomputers into a commercial leasing business, and Anthropic pays it $1.25 billion a month for capacity: a frontier lab renting the machines its intelligence runs on, from a rocket company pushing the business toward a $100 billion annual pace ^[13]. Wall Street has noticed, and it is selling the migration as a product. Brookfield is raising $10 billion of equity, with Nvidia as seed investor, to buy as much as $100 billion of AI infrastructure: land, energy systems, data centers ^[14]. Nvidia's chief executive, Jensen Huang, frames AI in exactly those terms.

AI infrastructure demands land, power, and purpose-built supercomputers--and our partnership with Brookfield brings all of these elements together in a ready-to-deploy AI cloud. — Jensen Huang

For smaller investors there is a ticker: the Defiance AI & Power Infrastructure ETF holds GE Vernova, Eaton, Vertiv and Quanta Services, the equipment makers whose order books run years deep on the cloud giants' construction budgets. It holds precisely what the labs are racing to secure ^[15]. Across twenty-nine institutional mid-year outlooks, the consensus favored AI infrastructure and electricity grids over software ^[16]. One caution flag from the same street: Goldman Sachs' chief equity strategist warns of an AI earnings bubble, with data-center spending and government borrowing competing for the same scarce capital. AI companies already account for 44 percent of the $135 billion in US convertible bonds sold this year, debt that can turn into stock ^[17]. The street that packages AI as electricity is also counting the debt underneath it. Then the week closed on a split screen. On Friday, the same day the wet lab opened, software stocks rallied on their earnings, and the fear that AI would eat them was written off as overdone ^[18]. The same morning, Meta launched Muse, a personal assistant with a free tier and subscriptions at $20 and $100 a month, priced, in the company's own framing, to recover its AI capital spending ^[19]. The labs are spending as though intelligence were becoming a utility; the public market just paid up for the software. One side of that split is early, and nothing in this week's numbers says which. The next round of filings is where the two readings meet: the labs' capital budgets on one side, the software companies' margins on the other.