A percentage is not a grid service
AI data centers are often discussed as if a fixed slice of their electricity demand could be switched off whenever a power grid needs help. A new preprint argues that this shortcut hides the variables a grid operator actually needs: how many megawatts can be released, for how long, with what probability, and from which collection of computing clusters. The study does not demonstrate a data center responding to a grid instruction. It converts a large production workload trace into a testable estimate of eligible load relief, then shows why eligibility shrinks when a promise must last longer or be dependable more often.
That distinction is constructive for both electricity systems and AI infrastructure. A vague flexibility claim is difficult to plan around. A duration-specific, reliability-specific quantity can be measured, contracted and later compared with performance. If data-center operators can validate such quantities with meters and controlled curtailment trials, flexible computing could help absorb grid stress while preserving high-priority work. The preprint supplies a framework for stating that promise, not evidence that the analyzed operator already offers it.
Reconstructing electricity from a production trace
Louisiana State University researcher Meiyi Li analyzed a public Alibaba production trace covering 155,410 GPUs on 37,707 servers across 17 clusters for 185 days. The underlying records describe hourly GPU allocation and utilization, workload type, priority and other scheduling information. They do not include an electricity meter. Li therefore reconstructed power with workload-specific GPU curves, estimates for host computers and idle accelerators, and a central power usage effectiveness factor of 1.2 to represent facility overhead such as cooling.
The resulting dataset contained 4,439 valid hourly observations. Across 300 draws of uncertain power-model parameters, the paper reports a time-averaged median facility demand of 55.83 megawatts. Its average pointwise range from the fifth to the 95th percentile was 45.74 to 69.10 megawatts, while the central-parameter reconstruction averaged 52.32 megawatts. Those differences matter. They show that 55.8 megawatts is a modeled estimate anchored to workload records and published measurements, not a reading from the site's utility connection.
The analysis then sorted workload power by operational meaning. Its central rule treated low-priority jobs as eligible for preemption, but it did not assume that stopping computation would turn an allocated GPU completely off. Keeping the device's idle draw in the calculation reduced average immediate eligible relief from 5.99 megawatts of attributed workload power to 3.55 megawatts. That smaller number equals about 12.1 percent of workload power and 6.35 percent of the modeled median facility demand. The denominator and the electrical boundary therefore change the apparent flexibility substantially.
Eligible load is not dependable relief
Eligibility describes work that might be interrupted under the study's classification. Dependable relief asks a harder historical question: how much eligible power remained available throughout an event window in at least a chosen share of possible starting hours? For the 13 clusters with enough eligible load to enter this calculation, mean eligible power was 3.53 megawatts. Under the paper's full-realization assumption, the amount available in 95 percent of sampled windows fell from 2.51 megawatts for one hour to 2.32 megawatts for four hours and 1.95 megawatts for 24 hours.
Full realization is a ceiling, not a measured control result. The paper represents actual delivery with a factor called q, ranging from zero to one. If an operator can realize only half of the eligible response, every dependable-megawatt estimate is halved. The trace cannot determine q because it does not record grid dispatches, successful checkpoints, service-level consequences, control latency, cooling response or the extra energy used when paused jobs resume. Those quantities require field trials and facility metering.
A single average percentage also misrepresented the shape of the result. When calibrated to mean eligibility, it overstated the paper's 95-percent-available quantities by 17.4 percent at one hour, 25.5 percent at four hours and 46.6 percent at 24 hours. Calibrating a percentage to match the four-hour tail simply moved the error: it understated one-hour relief and overstated 24-hour relief. The lesson is not that percentages are always useless. It is that one percentage cannot encode duration and reliability at once.
Combining clusters helped, but less than an independence assumption suggested. Li reports that aggregation raised four-hour firmness, meaning dependable relief divided by mean eligible power, from 0.38 for the median cluster to 0.66 for the 13-cluster portfolio. Residual changes across clusters were positively correlated, so weak availability could occur together. The public records could not identify why. Grid planners should therefore credit only diversification demonstrated by a particular portfolio rather than presuming geographically or administratively separate clusters fluctuate independently.
Scheduling does not create a free battery
The paper separately examined whether jobs that had waited in the production scheduler formed a ready reserve of new work that could be delayed. In the better-covered portion of the trace, newly arriving high-priority training and offline-inference jobs with at least the required observed delay averaged only 0.00794 megawatts. Their one-hour capacity at 95 percent availability was zero. This finding describes the scheduler and matching rules in this dataset. It does not prove that a different scheduler, workload mix or customer agreement could never create a useful delay product.
That restraint is important because shifting computation is not the same as discharging a battery. A delayed training run eventually needs electricity, and interruption can require checkpointing, migration or restart. A rebound may arrive at an inconvenient hour. Online services may have latency obligations that make them poor candidates for interruption even when utilization looks low. A useful contract would identify eligible workload classes, the duration and reliability target, the measured realization factor, recovery limits and a rule for recalibration as the fleet changes.
A useful screen that still needs field validation
The preprint's practical contribution is a falsifiable screening tool. Instead of assuming that 10 or 20 percent of a facility is flexible, an operator can publish a surface of dependable megawatts for defined durations and reliability levels. A utility can then ask whether meter data and dispatch tests reproduce it. This moves the conversation toward accountable grid participation and could help useful AI capacity connect without asking power systems to trust an unqualified headline percentage.
Yet the evidence remains narrow. The manuscript is a first-version preprint and has not been peer reviewed. It covers one operator, one 185-day period and hourly rather than sub-hourly records. Hardware, workload labels, priority policy and cluster coordination may differ elsewhere or change within the same fleet. The host-power model accounts for a large share of reconstructed facility demand but was not independently validated against the traced site. Treating every low-priority pod as preemptible also does not confirm that each job can checkpoint successfully.
The next step is an instrumented demonstration, not a larger claim. Facility meters should record GPU, host and cooling response during requested curtailments of several lengths. Operators should track failed checkpoints, service effects, rebound energy and repeat-event performance, then estimate q by workload class and horizon. Replication across operators would reveal whether the duration curve is common or highly local. Until that evidence exists, the reported megawatts are estimates of historical eligibility under explicit assumptions, not dependable power already deployed to a grid.