A single Vera Rubin NVL72 rack draws up to 227 kW. A 100,000-GPU training cluster draws somewhere between 300 MW and 1 GW continuously. An AI factory the scale Meta is building in Louisiana may eventually draw 5 GW — approximately half the average electricity consumption of New York City. None of that power is free, and none of it arrives at the GPU without passing through a stack of technologies built by a specific set of companies that each own a distinct and non-interchangeable piece of the problem. This essay maps that stack, technically, from the generation source to the rack busbar.
For most of computing history, the constraint on how fast you could build a data center was silicon. You needed chips, memory, and networking gear. The physical building was a secondary concern, and the electrical grid connection was assumed to be available when you needed it.
AI changed that. The transition from conventional server infrastructure to dense GPU clusters — Hopper at 700 W per GPU, Blackwell at 1,000 W, Vera Rubin at densities that put full rack draws above 200 kW — compressed decades of growth in data center power density into three years. The grid was not designed for this rate of change, and it cannot respond to it on the timelines that hyperscalers are building to.
Grid interconnection queues in key US markets now stretch 7 to 10 years. Building a new transmission line from a generation source to a data center campus takes 5 to 10 years. Permitting a new natural gas peaker plant through a traditional regulatory process takes 3 to 5 years. But OpenAI's Stargate project in Texas needed power in 2025. Meta's Prometheus supercluster in Ohio needed power in 2026. These are not timelines that grid interconnection queues can serve.
The bottleneck in AI infrastructure is no longer compute. It is electricity — specifically, how to get large quantities of reliable, continuous power to a location on a timeline measured in months rather than decades.
The consequence is that every AI cluster build at serious scale now requires solving a power procurement problem that is as complex as the compute procurement problem. And solving it means assembling a stack of technologies from multiple vendor categories, each of which serves a distinct function in getting electrons from their origin to a GPU die without interruption.
What follows is a systematic map of that stack — what each layer does technically, which companies own it, and why they matter to the AI scale-out problem specifically.
The power path for a modern AI data center involves at minimum six distinct technology layers, each with different physics, different economics, different deployment timelines, and different vendor ecosystems. No single company owns the full stack. Every layer must be present and operational before the first GPU draws a watt of inference compute.
With the stack mapped, the rest of this essay addresses each major company in this ecosystem — what technology they bring, where in the stack they sit, and what specific role they play in the AI power problem.
GE Vernova is the company that turbine-powers a significant fraction of the AI build-out happening right now. It was spun off from General Electric in April 2024 as a dedicated energy company, inheriting GE's 130-year turbine manufacturing heritage and 75,000 employees across approximately 100 countries. Its core product lines for the data center market are its heavy-duty HA-class gas turbines and its aeroderivative LM2500XPRESS units.
The reason GE Vernova's turbines dominate early data center deployments is deployment speed. When a hyperscaler needs 100 MW of on-site generation capacity and cannot wait seven years for a grid interconnection queue, gas turbines are the only commercially available technology that can close the gap. The LM2500XPRESS is a 35 MW aeroderivative gas turbine — essentially the same engine architecture used in commercial aircraft, adapted for power generation. It has five-minute fast-start capability, 95% factory assembly into simplified modules, and can be installed without requiring access to a transmission line. When Crusoe Energy needed nearly 1 GW of power for its AI data centers, including the Stargate campus in Abilene, Texas, it ordered 29 LM2500XPRESS units.
The HA-class heavy-duty turbine operates at a different scale. A single 7HA unit produces over 400 MW with a combined-cycle efficiency above 64%. Meta's Hyperion data center in Louisiana — targeting 5 GW eventual capacity — uses H-class turbines as part of its co-located generation strategy. The Chevron and Engine No. 1 joint venture with GE Vernova, targeting 4 GW for US data center clusters, uses 7HA turbines as the backbone generation units. The NRG Energy partnership targets 5.4 GW using a similar configuration.
GE Vernova also makes the substations, grid management software (GridOS), and high-voltage equipment at Layer 5 — the transmission and grid interface layer that connects generation to the data center campus. This vertical span from turbine to substation to grid software is relatively unusual in the power industry and gives GE Vernova the ability to offer a more integrated solution to data center developers than a pure turbine manufacturer could.
On the longer horizon, GE Vernova is investing in the BWRX-300 Small Modular Reactor through GE Hitachi — a 300 MW boiling water reactor design targeting commercial operation by the late 2020s. Ontario Power Generation's Darlington site in Canada has received construction approval for a BWRX-300, with a target operation date of 2029. This represents GE Vernova's position in the SMR market, though commercialization timelines remain subject to regulatory and construction risk.
Bloom Energy makes solid oxide fuel cells (SOFCs) — electrochemical devices that convert natural gas or hydrogen directly into electricity through a chemical reaction, without combustion. The fundamental physics of the conversion process gives SOFCs several properties that make them well-suited to data center power applications that gas turbines cannot match.
The most important is load responsiveness. A data center AI workload does not draw constant power. A single server rack may swing from 15 kW to 30 kW within seconds as a prefill burst starts. At the data center level, this creates power swings of 50 to 100 MW on timescales of seconds to minutes — faster than most utility-scale generation can follow and faster than the grid itself can respond without dedicated spinning reserves. Bloom's SOFC systems, combined with supercapacitor buffers, can track these rapid power demands. The fuel cell stack itself does not have rotating machinery that needs to spin up; it adjusts its electrochemical reaction rate in response to electrical load demand with response times that are genuinely competitive for this application.
The second key property is behind-the-meter independence. A Bloom installation generates power on the data center campus, behind the utility meter, without requiring a transmission line connection to the broader grid. For a data center in a location with a seven-year grid interconnection queue, a Bloom fuel cell deployment is not just a faster option — it may be the only technically feasible option for a campus that needs to go live within 18 to 24 months.
Bloom's technical architecture uses stacks of ceramic fuel cells operating at roughly 700–800°C. The high operating temperature enables direct internal reforming of natural gas into hydrogen, which is then oxidized electrochemically to produce electricity. The efficiency of this conversion — approximately 65% electrical efficiency on natural gas — is substantially higher than a simple-cycle gas turbine (approximately 38–42%) and comparable to combined-cycle gas turbine plants. The emissions profile is also substantially lower, with no NOx or particulate emissions from combustion (there is no combustion), though CO2 is still emitted from the natural gas carbon content.
The deployment velocity Bloom has demonstrated for data centers is remarkable. For Oracle's initial AI data center deployments, Bloom delivered and commissioned fuel cell systems within 90 days — a benchmark that no other primary generation technology can match at comparable scale.
By late 2025, Bloom had over 1.5 GW of fuel cells deployed globally, with more than 400 MW serving data centers. Equinix has over 100 MW deployed across 20 sites. The $5 billion Brookfield Asset Management partnership announced in October 2025 represents the largest single capital commitment to SOFC deployment in the technology's history, targeting AI factory deployments globally with GW-scale ambitions. American Electric Power followed with a $2.65 billion unconditional purchase commitment for approximately 900 MW of SOFC capacity for a Wyoming AI campus.
NextEra Energy is the world's largest generator of renewable energy from wind and solar. It operates two primary businesses: Florida Power and Light (FPL), a regulated utility serving 12 million customers in Florida, and NextEra Energy Resources (NEER), the wholesale generation subsidiary that builds, owns, and operates renewable energy projects sold under long-term Power Purchase Agreements. For AI data centers, NEER is the relevant entity.
NEER's technology portfolio covers solar photovoltaic, onshore and offshore wind, battery energy storage systems (BESS), and nuclear power. Its technology advantage is not any single generation source — it is the ability to originate, finance, and deliver multi-gigawatt clean energy packages to hyperscalers who want their AI clusters powered with a low-carbon profile. No other US energy company has the development pipeline, the project origination capability, or the contracted backlog to match this.
The Google Cloud partnership, announced in December 2025, illustrates the model. The centerpiece is a 25-year agreement to restart the Duane Arnold Energy Center in Iowa — a 615 MW nuclear plant decommissioned in 2020 — providing 24/7 carbon-free firm power to Google's AI data center network. Beyond nuclear, NextEra secured a 2.5 GW clean energy package for Meta spanning 11 separate PPAs and two battery storage projects across multiple grid regions. The company's "Bring Your Own Generation" (BYOG) model — developing dedicated power generation sites physically co-located with or directly interconnected to data center campuses — is becoming the architecture for multi-GW AI infrastructure deployments.
NextEra's battery energy storage position is also significant. BESS — large-scale lithium-ion battery installations — serves a crucial function in the AI power stack: it provides the grid services, frequency regulation, and peak shaving that make intermittent renewable generation useful for firm power applications. A 2.5 GW solar portfolio without storage is not a reliable AI data center power source. The same portfolio with 500 MWh of BESS can guarantee power availability across most grid conditions. NextEra originated 13.5 GW of new generation and storage projects in 2025 alone.
The company has committed to a $90–120 billion capital expenditure plan through 2028–2029, explicitly targeting data center power. Its goal of building 15 GW of new power generation for data center hubs by 2035, with an upside case of 30 GW, represents one of the largest capital commitments to AI power infrastructure by any company outside the hyperscalers themselves.
Vistra is one of the largest competitive power companies in the United States and operates the largest merchant nuclear fleet in the country. The relevant word is merchant — these are nuclear plants that sell their power into competitive markets rather than through regulated utility rate recovery. This means Vistra can sign long-term power purchase agreements directly with hyperscalers without requiring regulatory approval, and it can uprate or extend the operating licenses of existing plants with primarily commercial rather than regulatory constraints.
For AI data centers, Vistra's nuclear assets are genuinely special. Nuclear power is the only generation technology that operates 24 hours a day, seven days a week, at high capacity factors (90%+), with zero carbon emissions, with fuel costs that are stable and independent of natural gas price volatility, and that can be contracted at scale on 20-year terms. Those properties are exactly what a hyperscaler building a multi-decade AI infrastructure wants from its power supply.
Vistra's existing nuclear plants in PJM — Perry and Davis-Besse in Ohio, Beaver Valley in Pennsylvania — are among the most valuable baseload assets in the northeastern US grid. Meta's January 2026 nuclear deals included a 20-year PPA with Vistra covering more than 2.1 GW of power from these three plants, plus additional uprates adding 433 MW of incremental capacity. The deal makes Meta one of the most significant corporate purchasers of nuclear power in American history and gives Vistra a 20-year contracted revenue stream that de-risks the capital investment required to extend and uprate the plants.
Vistra also operates a large gas generation and battery storage portfolio — over 40 GW of total capacity — which gives it the dispatch flexibility to provide the ancillary grid services that pure nuclear cannot. A data center with a Vistra nuclear PPA essentially has access to firm, low-carbon baseload power with a flexible generation partner capable of managing grid integration.
Constellation Energy is the largest operator of nuclear power plants in the United States, with 21 reactors across 12 stations producing approximately 10% of the country's zero-carbon electricity. Like Vistra, its assets are merchant nuclear — competitively sold rather than regulated utility recovery. Unlike Vistra, Constellation also pioneered the nuclear restart model with Microsoft's Three Mile Island deal.
The Three Mile Island Unit 1 transaction — a 20-year PPA under which Constellation invested $1.6 billion to restart the 837 MW plant that was decommissioned in 2019 — demonstrated that demand for firm zero-carbon power is strong enough to justify restarting previously shut plants. The economics work because the existing site has approved licenses, built infrastructure, trained workforce, and regulatory familiarity that greenfield nuclear projects cannot replicate. The restart capital cost is a fraction of new construction cost.
Meta's June 2025 deal extended Constellation's Clinton Clean Energy Center in Illinois through a 20-year PPA covering the 1.1 GW plant that was otherwise heading toward retirement in 2027. Google's partnership with NextEra to restart Duane Arnold follows the same model. The pattern is clear: hyperscalers are willing to pay premium prices for nuclear power if it means getting firm, 24/7, zero-carbon baseload electricity on a 20-year contracted basis.
From a technology perspective, the value Constellation brings is not innovation in reactor design — its plants are largely 1970s-era boiling water and pressurized water reactor technology. The value is operational excellence: Constellation's nuclear fleet consistently runs at capacity factors above 90%, with some plants exceeding 95%. For an AI factory that cannot afford grid intermittency, a capacity factor of 90% is meaningfully different from a solar portfolio at 25%.
Oklo represents a fundamentally different approach to nuclear power than Constellation or Vistra. Where those companies operate large existing plants producing hundreds of MW each, Oklo is developing a 75 MW microreactor — the Aurora Powerhouse — that is designed for factory manufacturing, rapid site preparation, and multi-decade operation without refueling.
The technical case for factory-manufactured microreactors is straightforward: every large conventional nuclear plant is built as a one-of-a-kind engineering project, which is why they consistently come in over budget and late. The Vogtle Unit 3 and 4 project in Georgia — the most recent US nuclear construction project — came in $17 billion over budget and seven years late. The hypothesis behind SMRs like Oklo's is that by building the reactor at a factory and shipping it to site, the learning curve that makes aircraft manufacturing cost-competitive can be applied to nuclear.
Oklo's Aurora design is a fast spectrum, metallic-fuel reactor that uses liquid metal (sodium) as coolant. It does not require water for cooling, which simplifies siting significantly. The fast spectrum design can in principle use spent nuclear fuel from conventional reactors as feedstock — a feature that is both technically interesting and politically important in markets sensitive to nuclear waste volumes.
The commercial position in 2026 is pre-revenue but with significant contracted demand. Meta's January 2026 deal includes a 1.2 GW power campus in Pike County, Ohio, with first reactors targeting 2030. Switch has contracted 12 GW of deployment through 2044. Oklo has also begun groundbreaking on a site backed by Department of Energy support. These are real financial commitments, but the timeline to first electricity is several years away.
Caterpillar occupies a position in the power stack that no other type of company can fill: it makes the diesel and natural gas generators that keep a data center running when every other power source fails. The generators inject at Layer 3 — the critical power layer — bypassing the UPS on a 10 to 15 millisecond transfer time when grid power is lost and the UPS battery is depleted.
Data centers are designed to achieve six nines of availability — 99.9999% uptime, which allows approximately 31 seconds of downtime per year. In the context of AI training, downtime is not just an inconvenience. An interrupted training run on a 100,000-GPU cluster can mean losing hours or days of irreplaceable gradient state, depending on checkpoint frequency. A single power interruption without adequate backup can cost millions of dollars in compute time. This is why every serious data center installation has backup generation capacity sized to carry the full critical load indefinitely.
Caterpillar's data center generator product line covers output ranges from a few hundred kW to multiple MW per unit, with modular designs that allow parallel operation of many units to cover full campus loads. The technical requirements for data center generators are distinct from construction site or industrial generators: they must start and reach full load within 10 seconds, synchronize to the bus correctly, and maintain stable frequency and voltage under rapidly varying AI workload demand. Caterpillar's EPG (Electric Power Generation) division has been designing to these requirements for decades.
Cummins is the other primary competitor in this space, with similar scale and technical maturity. Both companies are also investing in hydrogen-capable generator designs that can run on hydrogen fuel when hydrogen supply chains mature — relevant for data centers committed to zero-carbon backup generation.
Vertiv is the company that manages the power between the facility distribution layer and the compute load. Its portfolio spans UPS systems, power distribution units (PDUs), in-rack power delivery, battery energy storage systems, and — as covered in depth in the companion cooling essays — coolant distribution units and thermal management systems.
The UPS function is central. An uninterruptible power supply sits between the utility power input and the IT equipment, providing ride-through power during grid momentary events (microsecond voltage sags, frequency deviations, brief outages) using battery energy stored internally. For an AI cluster, UPS requirements are demanding in ways that conventional enterprise UPS systems are not designed for: the load changes extremely rapidly as GPU workloads ramp up and down, the power factor can be unusual, and the load does not tolerate even a momentary voltage deviation without triggering protective shutdowns in power supplies.
Vertiv's Liebert UPS line — covering single-phase and three-phase systems from a few kVA to multi-MW — is the dominant product in hyperscale data center critical power. The company co-developed the GB200 NVL72 reference architecture with NVIDIA, embedding its power and cooling equipment into the default design that hyperscalers deploying Blackwell use as their infrastructure template.
The 800 VDC transition — moving from the current 48–54 VDC in-rack distribution to 800 VDC for future Kyber-era NVIDIA racks — is a significant technology inflection for Vertiv. Its 800 VDC ecosystem integrating with NVIDIA's Vera Rubin Ultra Kyber platforms is targeted for commercial availability in the second half of 2026. Higher distribution voltage reduces current, reduces copper losses, and enables new power delivery topologies. Vertiv's ability to deliver a complete 800 VDC solution — from the building distribution level down to the rack busbar — is a key competitive differentiator as the industry makes this transition.
Vertiv reported $9.5 billion in backlog as of Q3 2025 with 60% organic order growth — numbers that reflect the fundamental supply constraint in critical power infrastructure at AI scale. Its liquid cooling revenue more than doubled in Q1 2025.
Eaton describes its role as "grid-to-chip" power management. This is an accurate characterization of its breadth: the company makes the electrical equipment that connects the high-voltage grid to the data center campus, the medium-voltage distribution equipment inside the facility, the low-voltage switchgear protecting individual circuits, and UPS and PDU systems at the rack level. It is a vertically integrated electrical infrastructure company with a specifically strong position in switchgear and transformers.
Transformers are a genuine supply bottleneck in the current AI data center buildout. A large data center campus requires multiple transformers — devices that step down high-voltage grid power (138 kV or 345 kV) to medium voltage (13.8 kV) and then to low voltage (480 V) for distribution within the building. The transformer market is substantially supply-constrained: lead times for large power transformers in the US are 2 to 3 years, driven by demand from both data centers and grid modernization. Eaton is investing $340 million in a new transformer manufacturing facility in Jonesville, South Carolina, targeting production beginning in 2027.
Medium-voltage switchgear — the equipment that protects, controls, and isolates electrical systems at the medium voltage level — is equally critical and equally constrained. Eaton is opening a new 370,000 square foot switchgear manufacturing facility in Bellevue, Nebraska, targeting production beginning in H1 2027. The switchgear market is projected to grow from $17.8 billion in 2024 to $31.8 billion by 2034, driven primarily by data center and grid modernization demand.
For 800 VDC, Eaton is developing a medium-voltage solid-state transformer (SST) that will sit at the heart of a DC power distribution system — converting medium-voltage AC directly to 800 VDC for distribution through the rack busbars. This is a technologically significant product because it eliminates one conversion stage from the conventional AC-to-DC architecture, improving efficiency and simplifying the distribution chain.
A reasonable question after mapping this ecosystem is whether any single company could vertically integrate enough of it to reduce the number of partners a hyperscaler needs to work with. The answer is no — not because of commercial or regulatory reasons, but because of physics and technology maturity.
Gas turbines and nuclear reactors address completely different deployment timescales. Gas turbines can be installed in 90 to 120 days. Nuclear plants take decades. A hyperscaler planning AI infrastructure for the next 20 years needs both: gas turbines or fuel cells to generate power while transmission and nuclear infrastructure is being built, and nuclear or large-scale renewables to replace gas over the medium and long term as carbon commitments tighten.
Fuel cells and gas turbines address completely different operating profiles. Fuel cells provide behind-the-meter independence and rapid load following. Gas turbines provide massive capacity at lower cost per MW. A data center that needs 500 MW of primary generation with grid independence uses gas turbines for base capacity and fuel cells for load-following and rapid deployment.
UPS systems and backup generators address completely different failure scenarios. The UPS handles microsecond to minute-scale grid events. Backup generators handle extended outages lasting hours or days. Both are required because the probability distribution of grid failure includes both short events (very frequent) and long events (rare but catastrophic without coverage).
Switchgear and transformers are not interchangeable with anything else. They are physical infrastructure that must be present for any of the other technologies to function. Their lead times are the binding constraint on how quickly new AI capacity can be brought online, which is why Eaton is building new manufacturing capacity as fast as it can.
The ecosystem is not redundant. It is layered. Each company owns a piece of the physics that cannot be replicated by any other technology.
The shift from the current 48–54 VDC in-rack power distribution to 800 VDC — driven by NVIDIA's Kyber rack platform — is not just an electrical engineering detail. It restructures the entire power chain between the utility grid and the GPU die, with material consequences for every company in the stack.
The fundamental physics reason to raise distribution voltage is to reduce current. Power equals voltage times current. If you want to deliver 200 kW to a rack, you can do it at 48V with 4,167 amperes, or at 800V with 250 amperes. Lower current means smaller conductors, lower resistive losses, less copper in the busbar, and less heat generated in the distribution path. NVIDIA has stated that 800 VDC allows over 150% more power to be transmitted through the same copper cross-section versus current architecture — which means a Kyber rack can potentially use a conventional air-cooled busbar rather than the liquid-cooled busbar that Vera Rubin requires at 5,000 amperes.
For the power stack companies, the implications are: Eaton's solid-state transformer becomes a critical enabling component in the new architecture. Vertiv's 800 VDC UPS and power distribution systems become the standard for hyperscale deployments from 2027 onward. The conventional AC UPS systems that currently dominate will be supplemented by or replaced with DC architectures that are inherently more efficient for data center applications. Delta Electronics and SolarEdge are also developing competing 800 VDC solutions, creating a new competitive tier in what was a relatively stable product category.
Power has always been an infrastructure concern for data centers. What changed with AI is the order of magnitude. A conventional 2022-era data center drew 20 kW per rack and could be powered from municipal utility service with standard electrical engineering. A Vera Rubin NVL72 rack draws 187–227 kW, needs liquid-cooled busbars, requires a dedicated CDU loop, and sits within a rack cluster that may draw 100 MW or more. That rack exists within a campus that may draw 500 MW, within an AI factory that may eventually draw multiple gigawatts.
At that scale, power is not a utility commodity you order from the local electric company. It is a multi-year infrastructure program requiring gas turbines, nuclear PPAs, fuel cells, UPS systems, switchgear, transformers, and backup generation — each from a different specialist company with different technology, different lead times, and different deployment constraints. The companies in this essay are not supporting cast in the AI story. They are the prerequisite infrastructure without which the GPU racks cannot run.
The grid interconnection timeline is the most important constraint almost no one in the compute industry talks about. A hyperscaler can order and receive Vera Rubin NVL72 racks within 6 to 12 months. It can hire the engineers to run them. It can sign the land leases. The thing it cannot do quickly is connect that campus to power. Grid interconnection queues in PJM and MISO stretch 7 to 10 years. The AI infrastructure build-out of the 2020s is, at its core, a race against that queue — using gas turbines, fuel cells, nuclear PPAs, and behind-the-meter generation to create AI capacity faster than the grid can deliver it.
Every company in this essay owns a piece of the solution to that race. None of them can solve it alone.
Manish KL writes about AI infrastructure, memory systems, accelerator architecture, cooling systems, and power. Related essays: AI Cluster Reliability Beyond Fault-Tolerant Parallelism · The Real AI Bottleneck Is Moving From Compute to Interconnect Power Density · The Cooling Stack Is the New Critical Path: How Blackwell GB300 NVL72 Racks Manage 142 kW
© 2026 Manish KL