Watts Are the New Spec Sheet for AI Data Centers

The most telling infrastructure news this week came with no ribbon-cutting. NVIDIA told operators that AI factories are power-limited, Reuters tallied the governors now rationing data-center electricity, and US reporting raised fresh doubts about who is actually cleared to buy H200 chips. Three different constraints — watts, grid access and export vetting — are converging on the same bottleneck: not the chips themselves, but everything around them.

Power Is Now the Design Constraint

NVIDIA's November 5 post on its Vera CPU states the problem plainly: "AI factories are power limited," so operators need infrastructure that maximizes "the work done for every watt" S2. The company claims that in a 100-megawatt data-center comparison, a single Vera CPU beats both the fastest x86 CPU and the most energy-efficient one — 1.6x more agents per watt against the former, 2x faster agent speed against the latter. These are vendor-reported comparisons, not independent benchmarks, and the named partner results come without full methodologies. Still, the framing is the signal: when a CPU vendor's headline metric is agents per watt rather than peak performance, the power budget has become the spec everything else serves.

More Tokens, Barely More HBM

The same logic runs through NVIDIA's companion post on memory. From Blackwell Ultra to the Vera Rubin platform, HBM capacity per GPU rose roughly 3%, yet NVIDIA claims up to 30x higher throughput per dollar on a SemiAnalysis agentic-inference workload S3. The stated trick is hierarchy: NVFP4 checkpoints shrink model weights, inactive KV cache is offloaded to cheaper storage tiers, and retrieval indexes stay out of HBM entirely. The post's own summary — that what matters is how many useful tokens a factory can deliver "within its power budget" — reads as an admission that raw capacity stopped being the selling point. Teams evaluating platforms should ask for the workload definitions behind such multipliers; "up to" figures are sensitive to token-rate and service-level settings.

Grids and Governors Set the Schedule

On the demand side, the meter is being capped publicly. A Reuters factbox published November 5 lists Texas pausing new data-center approvals through the grid-interconnection process over reliability concerns, and New York imposing a one-year moratorium on facilities drawing 50 megawatts or more — the first full state-level moratorium in the US S4. Pennsylvania has pulled data centers from its fast-track permitting program and now requires local community approval, while Portugal is preparing rules that would demand renewable power and cooling systems that do not compete with household water supplies. Denmark's new law goes further, giving data centers the lowest priority in the grid queue. The practical consequence: interconnection timelines and local approvals, not hardware lead times, increasingly determine when a cluster goes live.

Export Vetting Is the Quiet Risk

Finally, who gets the chips is under scrutiny. The Verge, citing New York Times reporting, describes approved Chinese buyers of Nvidia's H200 chips including firms with ties to blacklisted entities — one, New Matrix, appears to be little more than an empty office with wires dangling from the ceiling S1. The Trump administration relaxed sales restrictions after Nvidia agreed to hand the government a 15-percent cut of China revenue. For infrastructure operators, this is not just foreign-policy news: allocation flows are becoming a compliance surface, and counterparties two hops away can carry sanctions exposure.

The through-line is uncomfortable but clarifying. Compute is still scaling; the hard limits have moved to watts, grid queues and vetting paperwork. Teams that plan in megawatt-years rather than GPU counts will see fewer surprises.