The rise of flexible data centers

Start with the intuition the episode exists to overturn. Data centers have been treated as the most inflexible load on the planet, historically declining demand response and demanding extreme redundancy. The economic version is sharper: an AI accelerator turns power into extremely valuable tokens, so curtailing one looks obviously irrational, and Sivaram says that intuitive blocker is part of why he has so little competition. Kann’s framing is that most people in the market would locate the constraint there, on the compute side. Both then argue the compute side is the part that is moving. Google has announced flex and priority inference tiers, where a developer chooses between immediate and delayed delivery at different prices and, more importantly, at genuinely different service levels. Anthropic has peak and off-peak periods where rate limits bite earlier. And a March 2026 paper from Emerald’s chief scientist and co-authors finds 18 to 55 percent power flexibility across representative AI workloads. On the utility side, by contrast, you get one service level, which is that you get power.

Why that gap matters is the stranded capacity argument. The grid runs at roughly 50 percent utilization or less for most of the year, which Sivaram translates into more than 100 gigawatts of capacity that could serve AI factories today if they would ramp down partially during rare peaks, perhaps 50 to 200 hours a year. Set that against the demand numbers he cites, a NERC forecast in January of a 224 gigawatt summer peak increase almost all from data centers and an EPRI projection that data centers could use up to 17 percent of American power by the end of the decade. If the only response is building, he argues, you get higher rates and slower AI growth at the same time.

The economics are where the argument turns, and where the two men partly disagree. Both accept that traditional demand response does not work here: if the reward for curtailing is a smaller electricity bill, the value of the tokens you did not generate dwarfs the saving, which is the watt-bit spread Kann attributes to Brian Janous. What flips the calculation is not money but access. Sivaram’s example is a 200 megawatt data center able to reach 230 megawatts a year ahead of schedule while swapping in liquid cooling and next-generation accelerators, which he puts at billions of dollars of value, easily covering the cost of curtailing during rare peak hours. He half-disagrees with Kann on the operating-cost half of it, arguing some workloads already pencil today because a data center can sit at a point on the power-performance curve where it sheds much more power than performance, and that as the cost of generating tokens falls toward the cost of power, the only real operational input into intelligence, responding to prices will increasingly pay for itself.

The obstacle he does not minimize is that a data center is not one entity. There may be a developer and operator, a tenant cloud provider, an inference-serving company above that and an end customer above that, and all of them have to cooperate for net withdrawal at the grid connection to move. His answer is that the incentive has to come from the top: if utilities, system operators, regulators and governors say a flexible data center connects faster and bigger, everyone below aligns. Behind-the-meter equipment then becomes an amplifier rather than an escape. Sivaram argues AI factories belong on the grid and calls the opposite view shortsighted, because a load heading toward a quarter or more of national electricity is the system’s anchor tenant and losing it would be a catastrophe for everyone’s rates. Bridge power is rational in the near term and should be folded back into grid service once the connection arrives. Kann proposes a mini dispatch curve as the organizing idea: inside a single site, workload flex is the cheapest increment and should be used first up to the point where customer performance suffers, then generators, batteries or fuel cells. Sivaram likes the analogy but insists the curve is dynamic rather than static.