AI Data Centers Waste Significant Power Smoothing GPU Load Swings, Industry Warned

The public debate over AI energy consumption has largely centered on how to build enough generation capacity to keep pace with demand. A commentary published by Data Center Knowledge argues that framing skips a more fundamental question: why is consumption so high to begin with? According to the piece, authored by Taavi Madiberk, CEO of energy storage firm Skeleton Technologies, a substantial share of the answer lies in how operators manage the volatile power draw produced by large-scale AI training — not in the compute workloads themselves.

What's driving the volatility

Modern AI training clusters typically use a bulk-synchronous execution pattern in which thousands of GPUs compute in parallel, then pause simultaneously to exchange and reconcile data before the next computation cycle begins. At hyperscale, those coordinated pauses translate into abrupt, facility-wide drops in power demand. The swings are frequent and steep enough to stress transformers, power distribution hardware, and grid-connected equipment upstream of the facility — raising the risk of costly instability or outages.

To prevent demand from collapsing during idle intervals, operators commonly inject secondary workloads timed to fill the gaps. Oracle, cited in the piece, uses a millisecond-resolution monitoring system described as a "GPU heartbeat" to detect idle periods and trigger fill workloads almost instantaneously. The technique keeps the facility's aggregate power draw artificially flat — but at the cost of running computation that would not otherwise be needed.

Why the fix creates new problems

Secondary workloads split into two varieties, each with distinct downsides. Operators sometimes slot in genuinely useful deferred tasks, but these compete with the primary training job for memory bandwidth and thermal capacity, stretching training timelines and reducing throughput. When that tradeoff is unacceptable, the alternative is dummy computation — calculations that produce nothing, performed purely to hold power draw at a stable level. Across a facility with tens of thousands of GPUs, the cumulative energy spent on meaningless arithmetic is, the article argues, a largely invisible but material source of waste.

The consequences extend beyond the electricity bill. Facilities that declare higher peak power requirements face longer grid interconnection reviews, because utilities must verify that sufficient generation and transmission capacity exists before approving the connection. Sustained operation at peak load also accelerates wear on GPUs, cooling systems, and electrical infrastructure, compressing equipment lifespans and raising maintenance costs. Each factor compounds at the scale typical of hyperscale AI buildouts.

What to watch

Madiberk's argument is that the industry needs purpose-built demand-smoothing technology — his firm sells energy storage systems — rather than workarounds that inflate the consumption figure regulators and communities are already scrutinizing. Whether that means ultracapacitor-based buffering, software-level scheduling changes, or revised training parallelism strategies, the article does not prescribe a single answer, but it frames the volatility problem as urgent and addressable independent of new generation capacity.

Grid interconnection delays are already holding up data center projects in multiple markets, and political pressure over electricity costs is rising. If secondary workloads are as prevalent as the article suggests, addressing them would reduce both peak capacity requirements and total consumption without waiting for new power plants — a potentially faster path to relief than the supply-side investments currently dominating policy conversations.

Note that the piece is authored by the CEO of a company with a commercial interest in the problem space. The operational mechanics described — bulk-synchronous training pauses, Oracle's heartbeat system — are consistent with publicly known industry practices, but independent validation of the energy-waste scale cited is not available from the single source.