From Socket to Token: Where AI's Electricity Actually Goes

01 The scale of the draw

A single query to a chatbot feels weightless, but the machine answering it may sit inside a campus that draws as much power as a small city. The newest AI training sites are specified in the hundreds of megawatts, a scale that puts them alongside steel mills and mid-sized towns on a utility planning sheet. That is a genuinely new class of electrical load for the computing industry.

Scale changes siting math. For most of the internet's history, a data center could follow tax incentives and fiber routes almost anywhere, because even a large facility fit comfortably inside the grid's slack. A 300-megawatt AI campus cannot. It needs a dedicated substation, transmission-grade feeders, and often years of queuing before the utility can guarantee delivery.

The result is that electricity, not real estate or networking, has become the first question operators ask about a candidate site, and the hardest one to answer quickly.

02 From socket to server rack

Power moves through a chain before it ever reaches a GPU. The grid interconnect and an on-site substation step transmission voltage down to distribution levels; switchgear routes it; uninterruptible power supplies condition it against dips and outages; power distribution units take it the final hop to the rack, where server power supplies convert it once more to the voltages silicon actually wants.

Every stage takes a cut. AC-to-DC conversion, double-conversion UPS topologies, and transformer losses each burn a few percent, and the losses compound: a facility can surrender five to ten percent of its incoming electricity before a single computation runs. Multiplied together, the stage efficiencies define the ceiling of what the building can ever deliver to a chip.

The design goal is boring but unforgiving: keep each stage near 97 to 98 percent efficiency, because fractions of a percent at campus scale are whole megawatts.

03 Inside the silicon

Accelerator-dense racks moved the power equation. A single modern GPU can pull on the order of 700 watts at full tilt, and a rack packs eight or more of them alongside CPUs, memory, and networking. Where a traditional server shelf might draw a few kilowatts, an AI rack can draw ten times that, which is why operators now talk about busways and direct-to-chip delivery that would have looked exotic five years ago.

Nameplate thermal design power is only the anchor. Real workloads swing below TDP, but the important fact is directional: nearly every watt a GPU draws becomes heat inside the room. There is no exhaust stack. Electricity in, computation and heat out, and the heat must be carried away before silicon throttles.

That tight coupling between compute and thermal load is what makes cooling the co-equal of the power system rather than an afterthought.

04 The cooling tax

Power usage effectiveness, or PUE, is the industry's plainest measure of overhead: total facility power divided by the power the computers actually use. A PUE of 2.0 means half the electricity never reaches a computation. For decades cooling was the largest piece of that overhead, and older facilities still run ratios far above modern norms.

Hyperscale operators compressed the tax. Careful airflow management, hot-aisle containment, higher supply temperatures, and economizers that draw in outside air pushed modern facilities toward PUE figures of roughly 1.1 to 1.3, meaning ten to thirty percent overhead rather than a matching share. Liquid cooling extends the trajectory for AI racks, because liquid carries heat away far more effectively than air at the densities accelerated hardware now produces.

The trend explains a quiet accounting shift: a decade ago a facility's biggest engineering fight was against its chillers; today the fight is against the physics of ever-denser silicon.

05 Training versus inference

Training and inference draw power in different shapes. A frontier training run is a dense, bursty block load: thousands of accelerators pinned near their thermal limits for weeks, with checkpointing and synchronization creating a recognizable rhythm on the grid. A utility can read a training cluster the way it reads a factory shift.

Inference is distributed and steady. Millions of users generate continuous traffic that spreads across facilities and time zones, with utilization that hums rather than spikes. Individually small, inference is becoming the dominant long-term draw because it scales with usage rather than with research campaigns.

The planning consequence is significant: training builds peak-demand stories, but inference builds the base-load curve a grid must serve every hour, and the industry is only beginning to size interconnects around it.

06 The grid frontier

Interconnection queues have become the industry's gatekeeper. In many markets a new large load waits years for transmission studies, upgrade assignments, and permission to connect, and those timelines now dictate AI capacity schedules more than chip supply does. Operators negotiate directly with utilities over dedicated substations, tariff structures, and curtailment commitments.

The workarounds are telling. Some campuses pair with on-site generation, gas turbines today with fuel cells and nuclear arrangements under study, to bypass parts of the queue. Others site next to retired plants specifically to inherit their grid connections, because an existing interconnect is the scarcest asset in the buildout.

Power availability, not land or fiber, now gates where AI capacity lands, and that inverts decades of siting logic.

07 Efficiency outlook

Efficiency per token keeps improving. Each hardware generation delivers more computation per watt, better cooling multiplies the gains, and scheduling software squeezes utilization out of fleets. On a per-unit-of-work basis, AI keeps getting cheaper in energy terms.

The rebound question shadows that curve: cheaper compute invites more compute. If falling cost per token simply expands demand, total electricity use can climb even as efficiency improves, the same dynamic that played out in lighting and transport. The data-center consumption curve bending upward through 2026 is consistent with exactly that pattern.

Through 2027 the signals worth watching are concrete: interconnection queue positions for AI campuses, PUE disclosures from hyperscalers, the ratio of announced capacity to energized capacity, and whether inference demand settles into the steady base load grids can plan around, or keeps outrunning it.