Megawatts Over Models: AI's Next Constraint Is Power, Not Algorithms

For most of the last decade, the story of artificial intelligence was a story about software: better architectures, larger datasets, cleverer training tricks. The constraint that mattered was ideas. In mid-2026, the deals crossing our AI Deal Tracker tell a different story. The binding constraint is increasingly physical — power, cooling, and the interconnects that move data between chips. The frontier is being poured in concrete and wired to the grid.

Consider the week's most telling agreement. Nvidia and Valar Atomics have agreed to explore a 30 MW nuclear-powered AI data center in Utah, built around the Ward250 reactor and designed for minimal water use. A chipmaker partnering with a nuclear startup would have read as a novelty two years ago. Today it reads as a roadmap. When a single training cluster can draw more power than a small town, "where does the electricity come from" stops being an operational footnote and becomes a strategic decision made years in advance.

The same logic runs through the compute market. Reflection AI signed a multi-year lease worth up to $6.3 billion for priority access to Nvidia GB300 "Blackwell Ultra" GPUs inside SpaceX's Colossus 2 data center. Read that structure carefully: a model lab is not buying chips, it is renting guaranteed capacity, for years, from an owner of physical infrastructure. Compute is being financialized the way real estate or power generation is — long-dated commitments, priced in billions, that lock buyers and suppliers together well past the horizon of any single model release.

Even the quieter deals point down the stack toward physics. Angel funding into SmartCore's silicon-photonics optical I/O chiplets targets one of the least glamorous but most punishing bottlenecks in modern AI: moving data between processors fast enough to keep them busy. As models scale across tens of thousands of GPUs, the interconnect — not the individual chip — is often what caps effective throughput. Photonics is a bet that the next performance gains come from the wiring, not the transistor.

Zoom out and a pattern forms. The public sector has noticed too: the European Commission's backing of a Tarragona AI gigafactory frames sovereign compute as industrial policy, not merely procurement. Across our compute and hardware deal categories, the throughline is the same: capital is flowing to whoever can secure megawatts, floor space, and supply.

Why does this matter for what comes next? Because a constraint that shifts from ideas to infrastructure changes who can compete, and how fast. Algorithms diffuse quickly — a clever technique is a paper away from being everywhere. Power plants and gigawatt data centers do not diffuse; they take years to permit, finance, and build. If the frontier is gated by physical buildout, the advantage accrues to actors who can marshal energy and capital on multi-year timelines: hyperscalers, sovereign wealth, utilities, and the handful of labs able to sign nine-figure capacity commitments. The moat is no longer only the model. It is the substation next to it.

This is not a wholesale rejection of the software story — algorithmic efficiency still compounds, and a breakthrough that halves training cost would reshuffle every projection here. Hedge accordingly. But the direction of the 2026 deal flow is hard to argue with. When a chip company is scouting reactors, a lab is leasing capacity through 2029, and a government is subsidizing a gigafactory, the market is voting that the scarce resource is physical.

The near-future implication is uncomfortable for anyone who assumed AI would stay a light, abstract industry. The next phase looks heavy: land, power-purchase agreements, grid interconnection queues, cooling water, and the slow physics of construction. The winners of the next 18 months may be decided less by who has the best research team and more by who secured the megawatts in 2026. You can watch that race unfold, deal by deal, in the AI Deal Tracker — and increasingly, the most important numbers on the page are measured not in parameters, but in watts.