What Really Limits How Fast AI Can Grow?

AI companies can now buy chips faster than they can plug them in. The limit on how fast AI grows has moved from the chip to the power line, the transformer and the cooling system.

Last updated September 2026
Figure 1 · The 2030 scenario

A new industry the size of a country's grid

For most of computing history, the hard part was the chip. Today a large AI campus can have its servers delivered in months and then wait years for the electricity to run them. The scale explains why: data centers are on track to become one of the largest single sources of new electricity demand in the United States.

649 TWhBerkeley Lab's 2030 reference case for U.S. data-center electricity, about as much as all of Canada uses in a year.
11.8%That scenario's share of all U.S. electricity, up from about 4% in 2023.
521–843 TWhThe range once chip shipments, utilization and efficiency are allowed to vary.

These are modeled scenarios, not commitments or measured 2030 demand. Outcomes depend heavily on shipments, utilization, chip life, and efficiency.

The story in one paragraph

Think of an AI data center as a power plant running in reverse: it takes in a steady river of electricity and turns it into computation and heat. The chip gets the headlines, but a chip sitting in an unpowered building, or in a hall that cannot shed its heat, produces exactly nothing. So the real unit of AI capacity is not chips installed but chip-hours that are powered, cooled and connected. Over the past three years that unit has become scarce for reasons that have little to do with semiconductors: utility queues, transformer backlogs, transmission lines and local permission. The good news is that each of these is an engineering and coordination problem, not a law of physics, and the fixes are already visible.

  • The demand is real: U.S. data centers went from 58 TWh in 2014 to 176 TWh in 2023, and could reach 649 TWh by 2030.
  • The bottleneck moved: a transformer that took three to six months to order in 2019 took one to two years by 2024.
  • Heat is the hidden cost: nearly every watt that enters a rack leaves it as heat, and denser racks concentrate that heat into less space.
  • The way out is flexibility: campuses that can shift or pause work connect sooner, while onsite power and storage buy time for the grid to catch up.

Measured results, derived quantities, projections, targets, and editorial inference are identified by context. Announced capacity is never treated as operating performance.

Part I: What a data center actually is

A chip is useful only when every layer beneath it works

Useful capacity is where five things overlap: electricity that is actually energized, enough memory and network to feed the chips, room to remove heat, software that keeps the machines busy, and jobs that finish. Explore the layers below; the weakest one sets the output of the whole building.

Figure 2 · The stack map

Accelerators are the visible layer

An AI data center is a power plant in reverse. Each layer can strand the ones above it, and the scarce unit is reliably powered, networked, cooled accelerator-hours, not installed chips.

Usually the longest lead time

Grid and onsite power

What it does
Generation, transmission, interconnection, substations, storage, and backup establish electrical capacity.
Measure
MW · time to energize

Binding constraint: Interconnection queues and firm supply, not equipment delivery.

Berkeley Lab's 2026 scenarios describe electrical demand at the campus level. Facility, compute fabric, and workload figures are architectural descriptions, not measurements of any single deployed site.
Part II: The physics nobody can negotiate with

Every watt that goes in comes out as heat

Better power electronics can trim losses, but the first law of thermodynamics does not bend. Almost all the electricity a campus uses ends up as heat that must be carried out of the chip, the rack and the building, every second it runs.

Grid electricity→computation + losses→heat removal

Two ways the same watt gets wasted

TrainingThousands of chips work in lockstep, so one slow or failed node can stall the whole job. This favors huge, tightly networked campuses.
InferenceAnswering users can be spread across many smaller sites, trading tight synchronization for being close to people.
Either wayA building with no power and a building full of idle chips produce the same thing: zero useful work.

That is why the slowest step has moved. Servers can be assembled in weeks. Grid studies, transformers, transmission lines, substations and cooling plants take years.

Part III: How builders are responding

AI companies are turning into infrastructure companies

The labs that grow fastest are no longer just buying chips. They are reserving transformers, signing power deals, building their own generation and redesigning workloads so the grid can say yes sooner. Four levers matter most.

Flexible workloads

Training runs and batch jobs that can pause or move let a campus connect before the grid is fully upgraded, because it can promise to back off at peak hours.

Balanced systems

Buying enough memory and network alongside accelerators keeps expensive chips busy instead of waiting on data.

Thermal co-design

Liquid cooling, rack density, climate and water supply are chosen together, because every watt in comes out as heat.

Firm interconnection

Grid upgrades, onsite power and realistic ramp schedules are negotiated years before the servers arrive.

The evidence that power is now the constraint

The demand curve is steep and documented. Lawrence Berkeley National Laboratory’s U.S. series rises from 58 TWh in 2014 to 176 TWh in 2023; its earlier 2028 range was 325–580 TWh, and a later DOE update gives a 649 TWh 2030 reference case. These are successive model vintages, not one measured trajectory, but they all point the same way. Local grids feel it first: Ireland’s energy agency counted data centres at 21.2% of national electricity in 2024, and new connections around Dublin have been sharply restricted.

The supply side is where the delay lives. DOE’s March 2026 transformer review says distribution-transformer orders moved from roughly three to six months in 2019 to one to two years or longer in 2024. A transformer is only one part of energizing a 100 MW campus, which also needs land, a substation, generation and interconnection studies, but it shows how equipment built for a slow-growing grid is struggling with a sudden new customer.

Builders are responding by bringing their own power. The IEA tracks a 2026 U.S. pipeline of onsite gas projects for data centers. That is project intent rather than delivered power, and much of it is fossil-fueled, which is the tension at the heart of this story: the fastest electrons to deliver are often not the cleanest ones.

The global picture is large but the pinch is local

Worldwide, the IEA's 2026 update estimates 485 TWh of data-centre electricity in 2025 and about 950 TWh in its 2030 case, roughly Japan's entire consumption today. That is still only a few percent of global electricity. The problem is not that the world runs out of power; it is that demand arrives in huge, concentrated blocks at specific substations. A single 100 kW rack with a PUE of 1.2 draws about 120 kW and rejects about 120 kW of heat, and a campus holds thousands of them.

Water follows the same local logic. Evaporative cooling towers use water on site, dry coolers save water but burn more power, and the power plants feeding the campus use water elsewhere. There is no honest universal “litres per chatbot answer”; the right question is always which cooling design, in which climate, on which grid.

Who is building what

Liquid cooling manifolds, immersion tanks, medium-voltage switchgear and dedicated nuclear or geothermal plants are all aimed at the same wall. Search the record, or filter by infrastructure domain.

8 programmes
VertivLiebert XDU & Liquid SolutionsCoolant Distribution Units (CDUs) and modular manifold systems delivering treated liquid coolant directly to multi-kW compute racks and cold plates
Reported evidence
Core cooling partner for NVIDIA Blackwell NVL72 architectures; high-volume manufacturing of multi-megawatt CDU systems.
Announced next step
Standardized liquid infrastructure supporting 100+ kW per rack compute densities with dry-cooler heat rejection.
Unresolved risk
Fluid leakage risks inside high-voltage compute racks, quick-disconnect coupling reliability, and pump cavitation under variable pump speeds.
Schneider ElectricEcoStruxure Data CenterIntegrated medium-voltage switchgear, modular prefabricated power skids, and liquid-ready busway architectures designed for rapid multi-megawatt deployment
Reported evidence
Deploying turnkey electrical packages across global hyperscale data centers; contracted billions in power infrastructure equipment.
Announced next step
Reducing data center electrical installation and commissioning timelines from years down to months.
Unresolved risk
Supply chain shortages in heavy medium-voltage transformers and circuit breakers with delivery lead times extending over 3 years.
CoolIT SystemsDirect Contact Liquid CoolingCustom micro-channel cold plates mounted directly on GPUs, CPUs, and memory modules, coupled to stainless steel rack manifolds
Reported evidence
Cooled many of the world's fastest supercomputers (Frontier, Aurora) and deployed broadly across hyperscale AI clusters.
Announced next step
Liquid cold plates supporting 1,500W+ thermal design power (TDP) per individual silicon package.
Unresolved risk
Corrosion in mixed-metal fluid loops, micro-channel clogs from particulates, and tight mechanical tolerances on wafer packaging.
SubmerSmartPod ImmersionSingle-phase immersion cooling submerging entire server chassis in biodegradable synthetic dielectric fluid, eliminating heatsinks and fans
Reported evidence
Pilots and commercial deployments with Intel, hyper-scalers, and European research computing centers achieving PUE below 1.05.
Announced next step
Ultra-dense AI superclusters running >100 kW per tank with zero water consumption.
Unresolved risk
Technician maintenance handling dripping fluid-soaked boards, component cable degradation in hydrocarbon oils, and vendor warranty voiding.
EatonEnergyAware UPSHigh-efficiency double-conversion uninterruptible power supply (UPS) systems capable of fast bidirectional grid frequency regulation
Reported evidence
Operating in large hyperscale facilities; allows data centers to monetize multi-megawatt battery reserves to support local grid stability.
Announced next step
Grid-interactive data center electrical architectures acting as dispatchable virtual power plants.
Unresolved risk
Battery cell degradation from frequent micro-cycling and coordinating fast inverter response with utility protection relays.
Crusoe EnergyDigital Flare Mitigation & Off-Grid ComputeModular data center containers placed directly at oilfield stranded flare gas sites, wind farms, and geothermal plants, bypassing the electrical grid
Reported evidence
Operates hundreds of modular compute units across North America; expanded into gigawatt-scale AI data center parks backed by long-term clean power.
Announced next step
Multi-gigawatt purpose-built AI data center campuses paired with dedicated behind-the-meter generation.
Unresolved risk
High-speed optical fiber connectivity to remote off-grid locations and fuel gas quality variability causing generator maintenance issues.
Constellation Energy & MicrosoftCrane Clean Energy Center20-year power purchase agreement to restart the 835 MW Unit 1 nuclear reactor at Three Mile Island, dedicating 100% of power to Microsoft data centers
Reported evidence
Landmark corporate nuclear PPA executed in 2024, setting precedent for hyperscalers directly funding baseload nuclear capacity.
Announced next step
Commercial restart by 2028, providing 24/7 carbon-free power matching round-the-clock AI training workloads.
Unresolved risk
Nuclear Regulatory Commission (NRC) licensing timelines, turbine and steam generator refurbishment costs, and public opposition.
Open Compute Project (OCP)Open Rack v3 (ORV3)Open industry specification standardizing 48V DC busbar power delivery, blind-mate liquid cooling connectors, and universal chassis form factors
Reported evidence
Adopted by Meta, Google, Microsoft, and server OEMs globally; eliminated redundant individual server power supplies.
Announced next step
Establishing universal standards for high-pressure blind-mate liquid quick disconnects and 100 kW+ rack cooling loops.
Unresolved risk
Proprietary server architecture divergence by hyperscalers undermining open hardware interoperability.

Power Usage Effectiveness (PUE) ratings (e.g. 1.1) describe overhead outside the IT equipment; they do not reduce the heat each accelerator kilowatt must shed.

The optimistic view, with conditions

The data center becomes a good grid citizen

The most hopeful outcome is not endless new power plants. It is data centers that behave like flexible customers: ramping up in stages, pausing work at peak hours and paying for the grid upgrades they need. Done well, that lets AI connect sooner and helps pay for a stronger grid everyone uses.

Now

Grow in stages

Campuses that ask for power in realistic steps, matched to construction, connect far sooner than those demanding a gigawatt on day one.

Near term

Move what can move

Training and batch work shifts across hours and regions, so the campus uses spare grid capacity instead of adding to the peak.

Structural

Build with the grid, not around it

Storage and onsite generation act as a bridge, while new transmission and clean firm power catch up with demand.

What usable AI capacity actually needs

  1. Power that is actually connectedGrid investment, onsite assets and ramp schedules agreed years before the servers arrive.
  2. A balanced machineMemory and networking bought with the chips, not as an afterthought.
  3. Cooling designed for the loadRack density, liquid loops, climate and water chosen together.
  4. Workloads that can flexScheduling that shifts delay-tolerant jobs when power is scarce.
  5. Honest accountingReporting delivered chip-hours, not nameplate megawatts or chips purchased.

Sources, method, and boundaries

Energy quantities are Berkeley Lab and IEA scenarios, not commitments or measured 2030 demand. Flexibility mechanisms come from Berkeley Lab's 2026 grid-integration review; no scenario is presented as a forecast certainty. The stack decomposition describes how AI campuses are built, not a benchmark of any single site. Country comparisons are approximate and for scale only.

PUE
Power usage effectiveness: total facility energy divided by IT equipment energy; lower is more efficient.
Interconnection queue
The utility process for studying and approving a new large electrical load before it can be energized.
Stranded accelerator
Installed compute that cannot be productively used because power, cooling, memory, or network capacity is missing.

Read More

Essays, books, talks, and research that shaped this field’s arguments. Influence is not endorsement; company communications and advocacy are labeled. Some publisher links require a subscription.

  1. Advocacy essay

    Leopold Aschenbrenner — Racing to the Trillion-Dollar Cluster

    The influential infrastructure-growth scenario; its power and capital assumptions remain assumptions.

  2. Industry analysis

    Dylan Patel and colleagues — AI Datacenter Energy Dilemma

    The supply-chain and power-delivery perspective on scaling complete accelerator facilities.

  3. Industry analysis

    SemiAnalysis — What is So Hard About Behind-The-Meter Power For Datacenters? Part 1

    Examines the practical barriers to supplying data centers outside ordinary grid connections.

  4. Institutional report

    IEA, “Energy and AI” (2025)

    The most comprehensive recent system-level treatment of data-center electricity demand, grid connections, cooling, generation, and AI-driven infrastructure growth.

  5. Paper / report

    Patterson et al., “Carbon Emissions and Large Neural Network Training” (2021)

    Influential in shifting discussion from model FLOPs alone to hardware, datacenter efficiency, energy source, and system choices.

Across the fields: learning curves, deployment, and rebound