Research
中 RSS

Data Center Power: Supply Side and Demand Side

AI has overnight turned electricity into the most expensive bottleneck in data centers. For the past decade this industry obsessed over bandwidth, latency, and PUE; now the top question is blunter — where does the power come from, is there enough of it, and when will it actually be available.

Following that question, I worked through a macro survey from both the supply side (where power comes from) and the demand side (where power goes), and I’m writing it up as a stage-gated research note.

Data center construction site in Utah, USA
Aerial photograph of the NSA Utah Data Center during construction, showing several large facility buildings — complexes of this scale, measured in hundreds of thousands of square feet, typically carry steady-state power demands in the hundreds of megawatts. Source: Wikimedia Commons / Swilsonmc (CC BY-SA 3.0).

Supply Side: Where the Power Comes From

Data center power supply can be divided into two modes, delineated by the utility meter.

Front-of-the-Meter: Direct Connection to the Utility Grid

Power comes from the local grid and is paid for by usage — this is still the dominant model. As recently as a year ago, facilities that actually relied on on-site generation as their primary power source numbered only about 13%, meaning roughly 87% still depended primarily on the grid; a Bloom Energy industry survey of around 100 decision-makers projects that the on-site share will rise to about 38% by 2030, with roughly 60% of data centers still grid-dependent at that point[1].

Worth calling out separately is the “nuclear PPA” trend that has been generating significant buzz over the past two years:

Three Mile Island nuclear power plant with cooling towers in the distance
Three Mile Island nuclear power plant — Unit 2 (retired) is on the left; Unit 1 on the right is the Crane Clean Energy Center that Microsoft contracted to restart. Source: Wikimedia Commons (CC BY 2.0).

One thing worth clarifying: these “nuclear + AI” headline deals are structurally still front-of-the-meter — the data center does not own the reactor, the arrangement is a PPA (Power Purchase Agreement), and power still travels over the grid. Truly behind-the-meter nuclear means small modular reactors (SMRs) located within the data center campus itself, and that remains essentially in the planning and first-unit construction phase.

Behind-the-Meter: Data Centers Building Their Own Generation

Generation equipment is installed on the data center’s own side of the meter, and power is consumed directly without passing through a utility meter. This is the fastest-growing segment right now, and the core driver is simple — grid interconnection queues often mean waiting several years, making on-site build-out the fastest path to meeting project timelines[1].

Behind-the-meter on-site solutions, arranged roughly by deployment speed and scale, span a spectrum like this:

SolutionDeployment TimelineCharacteristicsRepresentative Players
Fuel cellsFastest (~90 days)Most agile, smallest footprint, near-zero noise/NOx, suited for rapid emergency deploymentBloom Energy (leader); FuelCell Energy, Plug Power, Ballard, Doosan
Small gas turbines / reciprocating engines~12–24 monthsSolid and reliable, suited for mid-scaleGE Vernova (LM2500/LM6000), Caterpillar, Cummins; Solaris (leasing)
Large gas turbinesNew orders queued to ~2028–2030Highest per-unit capacity, most economical at scaleGE Vernova, Mitsubishi Power, Siemens Energy
Nuclear (incl. SMR)Restart ~3 years / SMR post-2030Largest scale, most stable, cleanest, long-term solutionOperators: Constellation, Vistra; SMR developers: Kairos, NuScale, Oklo
Energy storageSupplementalPeak shaving, load smoothingTesla (Megapack), Fluence

Fuel Cells: Fastest and Most Flexible

Fuel cells play the role of “speed-first option.” In Bloom Energy’s partnership with Oracle, the target was on-site data-center-scale power delivery in 90 days — in practice, the first batch of modules was delivered in just 55 days[4]. The physical form factor is a set of shipping-container-sized modules that can be combined into installations ranging from 20 MW to 500 MW+, running on natural gas, biogas, or hydrogen, with availability of 99.9%–99.999%[5].

The trade-off is a per-watt capital cost roughly 10%–15% higher than small gas turbines, but in islanded microgrid deployments requiring 99.9% availability, that gap is offset by the over-build cost gas turbines incur to maintain standby redundancy[6].

Four Bloom Energy Server fuel cell modules
Bloom Energy Server (colloquially the “Bloom Box”) — a solid oxide fuel cell (SOFC) unit roughly the size of a large refrigerator. Source: Wikimedia Commons (CC BY 3.0).

Gas Turbines: The Backbone of Large-Scale Behind-the-Meter Power

What truly underpins large-scale behind-the-meter deployments is gas turbines. GE Vernova’s LM2500 has now surpassed 2,100 cumulative deliveries and 75 million operating hours; the LM6000 has more than 1,200 deliveries and 60 million operating hours, and can synchronize to grid within 5 minutes of startup[7]. Both are aeroderivative models — derived directly from commercial aviation jet engines — which makes them rugged, reliable, and backed by a mature maintenance ecosystem.

GE LM2500 gas turbine undergoing maintenance
The LM2500 gas turbine — derived from the civil aviation CF6 engine, widely used in naval vessels, offshore platforms, and land-based power generation. Source: U.S. Navy (public domain).

That said, order books at large gas turbine manufacturers are already filled out to 2028–2030, so new data centers hoping to use this path simply cannot afford the wait.

Nuclear and SMR: The Long-Term Optimal Solution, but Still a Distant Horizon

SMRs are the central narrative for “behind-the-meter nuclear”: shrink a reactor into factory-manufactured modular units, assemble them on-site, and locate them directly within the data center campus. The most aggressive plan — Oklo’s Aurora — aims for its first unit to go online in 2027–2028, but the mainstream estimate is that the first commercial SMRs will enter data center service power around 2030; NuScale has signed a 6 GW deal with TVA with a target of power generation from the first plant by 2030; Kairos’s Hermes-2 molten-salt reactor is also targeting 50 MW for TVA in 2030[8].

NuScale Power Module small modular reactor cutaway diagram
NuScale Power Module cutaway — a single integrated unit containing the reactor, containment vessel, and steam generator, designed to output 50–77 MWe. Source: NuScale Power, LLC.

Business Model: Short Hardware Tail, Long Service Tail

Behind-the-meter equipment vendors typically operate on a model of “one-time hardware sale + 15–20 year long-term service contract.” Taking Bloom Energy’s backlog as of end-2025 as an example: product backlog roughly $6 billion, service backlog roughly $14 billion, totaling roughly $20 billion — a 3:7 hardware-to-service split, with services forming the much longer tail[9].

This model is structurally identical to the playbook of aviation engine manufacturers (GE, Rolls-Royce) and industrial gas companies (Air Liquide): deploy the expensive hardware first, then collect maintenance, spare parts, and fuel contract revenue over the following two to three decades — long-term contracts lock in customers, and service cash flows are smooth and counter-cyclical.

Demand Side: Where the Power Goes

Having covered the supply side, the next question is: once power enters the data center, exactly where does it go?

Both Epoch AI and American Compute have broken this down using NVIDIA’s official GB200 reference architecture. Taking a reference AI data center of approximately 171 MW (NVIDIA GB200 NVL72 architecture), the approximate power share by module is as follows[10][11]:

ModuleShare
GPUs~40%
Other server components (CPU, memory, NVLink, PSU losses, etc.)~22%
Cooling + power distribution~29%
Networking equipment~9%

The 29% for cooling is based on air cooling; switching to direct liquid cooling can bring this below 15%[10].

GPUs are the dominant power consumer by far: H100 700W → B200 1,000W → B300 1,400W — each generation more demanding than the last. Eight GPUs in a single chassis, without accounting for CPU or networking, already draws 11.2 kW; starting with the B300, liquid cooling is not optional[12]. For comparison, a single Intel/AMD server CPU draws roughly 300W — not remotely in the same league.

NVIDIA DGX GB200 NVL72 rack
NVIDIA DGX GB200 NVL72 — 72 Blackwell GPUs and 36 Grace CPUs in a single rack, with a total rack power draw of approximately 120 kW, liquid-cooled as standard. Source: Wikimedia Commons (CC BY-SA 4.0).

Metrics: From PUE Toward Tokens-per-Watt

PUE (Power Usage Effectiveness) = total facility power ÷ IT equipment power, closer to 1.0 is better. This metric has been in use for nearly two decades; the global industry weighted average has hovered around 1.54 for six years with essentially no movement. Top operators, however, can push it to extremes — Google’s trailing twelve-month PUE across its global data center fleet in 2024 was 1.09[13][14].

Water and carbon have their own parallel metrics: WUE (water) and CUE (carbon), with the same logic as PUE, just different denominators.

Genuinely new in the AI era is a shift in metrics from “energy consumption efficiency” toward “compute output efficiency”:

The rack density gap is fundamentally the question of whether legacy general-purpose cloud infrastructure can physically accommodate the next generation of GPUs — many older data centers are not short on power per se, but their cooling is inadequate, their power distribution was never designed for it, and their per-rack limits simply cannot reach those levels.

Summary

In the next phase I want to dig separately into several threads: where the actual bottlenecks are in gas turbine production capacity, how large a real market share Bloom Energy can capture in this cycle, and whether the SMR cost curve can ultimately compete with gas — all left for future notes.

References — sources

  • data center
  • power
  • fuel cell
  • gas turbine
  • nuclear
  • GPU
  • cooling