Welcome to The Calibre Brief, a Calibre Code USA production.

Electricity capacity costs nine times more than it did two years ago. One AI data center claims it can fix this—by behaving like a power plant.

In Manassas, Virginia, construction is underway on a 96-megawatt facility inside PJM's transmission grid, where interconnection queues stretch years long and capacity prices have climbed ninefold in two years. Aurora is not just another data center build. It is designed to answer whether a computer facility can earn the same grid identity as a gas peaker plant.

The engineering proof of concept came from a May 2025 demonstration in Phoenix. A 256-GPU cluster running Emerald AI software profiled each running job, sorted workloads by interruptibility, and applied power caps to shed roughly 25 percent of load during simulated peak demand while completing its work. The underlying data and code were released publicly. The coalition's claimed capabilities—multi-hour curtailment, carbon-aware operation, geographic load shifting—trace to this demonstration and to a paper under IEEE review.

Aurora is 700 times larger than the largest reported real-world deployment, and the evidence comes from teams whose commercial interests align with a positive result. The coalition's estimate of 100 gigawatts of unlocked grid headroom describes peak-hour potential, not round-the-clock energy. More critically, a gas peaker earns its capacity payment because it accepts penalties for failure to perform on every emergency hour PJM might declare—including multi-day winter storms that exceeded the longest curtailment any demonstration has yet reported. No independent certification of Aurora's capabilities has been completed. Whether PJM will accredit software-flexible load as firm capacity, whether baselines can be measured without gaming, and whether flexibility revenue plus interconnection priority double-recovers the same attribute remains open, and will be decided in proceedings still in progress.

Aurora's DCFlex testing determines whether power-flexible AI factories become a standardized grid asset or remain a vendor pilot. The engineering has shown it can bend the load. The certification answers a different question: whether anyone will write that behavior into a tariff, assign it a price, and collect when it does not show up. The answer shapes how fast the AI buildout proceeds within existing grid constraints, and who pays for the transmission upgrades new data centers require.

The Aurora project has already established that software-orchestrated load reduction is technically real at small scale. What remains genuinely undetermined is whether the measurement, accreditation, and penalty frameworks that make a gas peaker a trustworthy grid asset can be rebuilt for a different kind of machine—one whose firmness is a software policy rather than a physical turbine. That institutional translation, not the engineering, is what Aurora is actually testing.

The Calibre Brief, a Calibre Code USA production, asks: should software-flexible data centers earn partial capacity credit in PJM's market, or do they require a separate, narrower flexibility product with different pricing and different penalties? Share your analysis in the YouTube comments. Please like this episode and subscribe to The Calibre Brief wherever you listen.