Which startups control the power behind AI today?
In our AI infrastructure market deck, you will find everything you need to understand the market
SUMMARY
Crusoe controls the broadest slice of the power behind AI today, with Lancium and VoltaGrid holding the next most important choke points.
Electricity became the main physical bottleneck because AI campuses can be financed and built faster than grids can connect them. The national power supply matters, but the immediate constraint is usually a particular substation, transmission corridor or equipment delivery date.
No startup controls the whole electricity stack. The strongest companies own something that can decide whether a campus switches on: live capacity, power-ready land, reserved generation equipment, an approved grid position or the systems that keep a huge electronic load stable.
Crusoe leads because it has already converted energy access into more than 200 megawatts of live AI infrastructure, then used that operating proof to support several gigawatts of contracted development. Its edge is execution across the campus, not ownership of every power asset underneath it.
Lancium’s influence is less visible but unusually hard to replace. Grid-connected land with substations and a credible energization path can be more valuable than a technically impressive generator that still lacks a site, permit or buyer.
VoltaGrid shows why natural gas is ahead in the near term. Modular engines can arrive in blocks, sit behind the meter and energize one phase while the utility is still working on the rest of the campus.
The electrical-control layer is becoming a bottleneck of its own. ON.energy can determine whether dense AI loads ride through faults safely, while GridCARE can help utilities release capacity that already exists but is trapped inside planning and interconnection processes.
Clean firm power is gaining credibility, but mostly on a later timetable. Fervo has the strongest geothermal build, Kairos has the clearest nuclear execution, and Form Energy offers long-duration resilience, yet none currently matches gas or grid-connected sites for speed at multi-gigawatt scale.
Hyperscalers create winners without surrendering control. They fund several technologies and regions at once, which gives startups enormous contracts but leaves most of them dependent on a small number of buyers with plenty of alternatives.
The headline gigawatts need heavy filtering. Operating capacity, financed construction, binding contracts, equipment reservations and long-range aspirations are not interchangeable, and the ranking changes quickly once those categories are separated.
The practical hierarchy today is therefore Crusoe first, Lancium second and VoltaGrid third, followed by narrower control points such as ON.energy, GridCARE and Fervo. Nuclear developers still control valuable options; they do not yet control delivered electricity.
This market map, featured in our AI infrastructure market deck, highlights top companies and startups in the AI infrastructure market
Why did electricity become AI’s biggest physical bottleneck?
AI electricity demand is now growing faster than the grid can connect the campuses being proposed.
The latest International Energy Agency outlook estimates that global data-center electricity use rose to about 485 terawatt-hours in 2025 and could reach roughly 950 terawatt-hours by 2030. In the United States, EPRI currently projects data centers could consume 9% to 17% of national electricity by 2030. The range is wide, but even its lower end would more than double today’s share.
The immediate headache is local rather than national. The US Department of Energy has seen connection requests for individual hyperscale sites ranging from 300 megawatts to more than one gigawatt, often with requested delivery in one to three years. Utilities can build that much generation over time; they rarely have the substation, transmission and equipment ready where a customer wants it now.
Money is arriving much faster than electricity. According to the IEA’s latest update, capital spending by five large technology companies exceeded $400 billion in 2025 and was expected to rise another 75% in 2026, while utilities still need years to add transmission corridors, turbines or nuclear plants.
What does it really mean to control AI power, and does any startup control the whole stack?
No startup currently controls the full chain. Real control comes from owning a bottleneck that can decide whether a data center switches on, where it gets built or how reliably it runs.
For us, an announcement starts to count when the site, permit, customer and delivery path begin to line up. Operating generation carries the most weight, followed by financed construction, an approved grid position, scarce equipment already reserved and contracts that commit real money.
Unglamorous assets move near the top of the ranking. Lancium’s land and interconnection rights can carry more practical weight than a reactor design. ON.energy’s medium-voltage systems can determine whether a volatile AI load meets grid requirements. VoltaGrid’s gas generators may look old-fashioned beside advanced nuclear, but they can arrive years earlier.
Crusoe develops and operates AI campuses, but its largest sites still need utility connections, power plants and equipment from established manufacturers. VoltaGrid can place generation behind the meter, while Energy Transfer supplies the gas and INNIO supplies many of the engines. Kairos Power can build a reactor only after regulatory approvals and with a utility willing to buy the output.
The buyers also keep their options open. Amazon is supporting X-energy nuclear projects, Zanskar geothermal power, solar farms and batteries at the same time. Google has agreements spanning geothermal and advanced nuclear. A hyperscaler can move capital between technologies and regions more easily than most startups can replace a hyperscaler customer.
The real power sits at individual choke points. A startup becomes powerful when its absence can delay a campus by years, force it to another location or leave billions of dollars of computing equipment unable to run. By that test, only a small group has meaningful influence today.
| Type of control | What we look for | Leading startup examples | Main limitation |
|---|---|---|---|
| Usable power today | Energized capacity and operating equipment | Crusoe, Redwood Materials | Still relies on outside generation or grids |
| Fast new generation | Financed equipment, fuel access and near-term delivery | VoltaGrid | Fuel and engine supply remain external |
| Power-ready locations | Land, substations and approved interconnections | Lancium, GridCARE | Utilities retain final authority |
| Future firm power | Permits, construction and binding buyers | Fervo, Kairos Power | Most capacity is still being built |
| Electrical reliability | Proven systems that manage faults and load swings | ON.energy, Form Energy | Depends on outside generation |
If you want more recent data on this point, please see our latest AI infrastructure market report.
As this chart shows, and as featured in our AI infrastructure market deck, search interest in AI infrastructure has risen sharply
Is Crusoe the clear AI power leader right now?
Crusoe is currently the strongest overall startup in AI power because it has already turned energy access into live computing infrastructure at unusually large scale.
Its Abilene campus is the clearest proof. Crusoe broke ground in June 2024 and energized the first two buildings, totaling more than 200 megawatts, within roughly a year. The site is live for Oracle Cloud Infrastructure and built around dense NVIDIA systems rather than ordinary colocation demand.
Since then, the site has become much bigger. The original Lancium-owned campus is planned to reach 1.2 gigawatts, and a neighboring 900-megawatt Microsoft development would take the wider Abilene footprint to about 2.1 gigawatts. Crusoe and Lancium have also announced a separate one-gigawatt campus in Childress.
Crusoe now reports 4.9 gigawatts of contracted AI infrastructure capacity. That figure is contracted rather than live, yet it carries more weight than a loose development pipeline because named campuses, customers and construction already sit underneath it. A recent 750-megawatt agreement with Bergen Engines and the new five-gigawatt ON.energy partnership show Crusoe reserving the machinery needed for the buildout.
No other startup currently combines that range of work with more than 200 megawatts already live.
Why is Lancium so hard for AI developers to bypass?
Lancium controls the asset AI developers struggle most to find: a large site with a believable path to power.
At Abilene, Lancium owns the campus while Crusoe develops and operates the data centers. The first phase is already live, and the campus is expanding toward 1.2 gigawatts. Lancium later secured a $600 million debt package tied to the site, giving the project more substance than an unfinanced interconnection reservation.
Its model is now repeating across Texas. Lancium owns the land for Crusoe’s newly announced one-gigawatt Childress campus and is involved in a separate QTS project in Hall County. Across Abilene and Childress alone, it sits underneath 2.2 gigawatts of named development before counting the newer Microsoft site beside Abilene or the Hall County project.
Lancium can shape who gets a realistic path to energization because the transmission, substations and flexible-load systems have already been worked through at its sites.
This chart, included in our AI infrastructure market deck, shows annual VC investment in AI infrastructure startups
Is VoltaGrid winning the race for fast data-center power?
VoltaGrid is currently the leading startup for large, fast, behind-the-meter generation.
Its disclosed data-center commitments include 2.3 gigawatts for Oracle Cloud Infrastructure and more than one gigawatt with Vantage Data Centers. Those two relationships alone exceed 3.3 gigawatts, far above the near-term commercial pipelines of advanced nuclear developers.
VoltaGrid has backed those sales claims with hardware and money. It ordered 1.5 gigawatts of INNIO gas-engine capacity, expanded its work with ABB and secured a $1 billion strategic equity investment led by Blackstone and Halliburton. A previous financing package included $2 billion of secured notes and a $3 billion asset-based facility.
Its QPAC platform can combine nodes of up to 20 megawatts into larger systems. That lets a campus energize one phase while later buildings remain under construction. The approach suits AI developers that cannot wait for a utility to finish a one-gigawatt upgrade all at once.
VoltaGrid still depends on pipeline capacity, permits and outside engine manufacturers. Even so, none of the startups we reviewed has assembled a larger credible package of near-term on-site generation for AI customers.
If you want more recent data on this point, please see our latest AI infrastructure market report.
Why is natural gas ahead of advanced clean power for now?
Natural gas is winning the current speed-to-power race because AI customers need firm megawatts before most new clean technologies can deliver them.
VoltaGrid’s multi-gigawatt customer book and the rush to reserve engines show where near-term demand is going. Gas systems can be installed in blocks, run continuously and sit behind the meter where grid capacity is missing. A customer can start with tens of megawatts and add more as the campus grows.
The clean alternatives are moving, but their timelines remain smaller or later. Fervo’s first 100 megawatts at Cape Station are due to reach operating scale around early 2027, with 500 megawatts planned by 2028. Kairos Power’s first electricity for Google is scheduled around 2030. Form Energy can store power for days, though it still needs another source to charge the batteries.
Renewables will still provide a large share of data-center growth. The IEA expects wind and solar to meet nearly half of the additional electricity demand through 2030. Those projects usually need firm generation, storage or grid support beside them, especially for campuses expected to run expensive accelerators around the clock.
Gas still faces emissions, permitting, fuel-price and equipment constraints. For campuses racing to open within the next few years, speed wins.
This chart, included in our AI infrastructure market deck, shows why CoreWeave is winning in AI infrastructure
Which geothermal companies are closest to real AI scale?
Fervo is the clear geothermal leader today, while Zanskar has emerged as the most credible private challenger.
Fervo’s Cape Station is fully contracted and backed by $421 million of non-recourse project financing. Its first 100 megawatts are moving toward operation, with the wider development planned to reach 500 megawatts. Fervo has also become public, so it is now better described as a startup-born scale-up than a private startup.
The freshest operational result is more important than the IPO. Fervo reported that its third-generation well design drilled 143% faster than its first Cape Station well. Geothermal economics are heavily influenced by drilling time, so repeated improvement across wells gives us more confidence than one successful demonstration ever could.
Zanskar is much smaller, but it now has both operating experience and a serious AI-linked contract. The company operates the Lightning Dock plant and recently signed a 100-megawatt power agreement with NV Energy. Amazon says that geothermal output will support its future Nevada data centers alongside 600 megawatts of solar and 600 megawatts of battery storage. Zanskar also raised a $115 million Series C and added a separate development-finance facility that can scale to $100 million.
That puts Zanskar well behind Fervo on scale but ahead of most private geothermal peers on operating proof.
Which nuclear startup is actually furthest ahead?
Kairos Power is furthest ahead on execution today, X-energy has the deepest industrial coalition, and Oklo still leads mainly in headline demand.
Kairos has broken ground on Hermes 2 after receiving Nuclear Regulatory Commission construction permits. The plant is designed to supply up to 50 megawatts through TVA around 2030, becoming the first project under Google’s broader plan for as much as 500 megawatts by 2035. A permitted site under construction deserves more weight than a much larger agreement without a permitted commercial plant.
X-energy’s advantage is its supply chain. Amazon, Energy Northwest, Korea Hydro & Nuclear Power and Doosan are supporting a program aimed at more than five gigawatts by 2039, starting with a Washington project that could grow from 320 to 960 megawatts. X-energy has also reserved graphite capacity and is building its own TRISO fuel infrastructure.
Oklo has the largest customer figures: a 1.2-gigawatt Meta-backed project in Ohio and a non-binding agreement covering up to 12 gigawatts with Switch through 2044. Those numbers show extraordinary demand for nuclear power, but they describe a long development runway rather than electricity available now.
| Company | Strongest evidence today | AI-linked scale | What keeps it from current control |
|---|---|---|---|
| Kairos Power | NRC permits and construction underway at Hermes 2 | 50 MW first project; up to 500 MW with Google | First delivery is still around 2030 |
| X-energy | Amazon backing, utility partner and expanding fuel supply chain | More than 5 GW targeted by 2039 | First commercial fleet remains under development |
| Oklo | Meta funding mechanism and huge customer interest | 1.2 GW with Meta; up to 12 GW non-binding with Switch | Commercial licensing and construction remain ahead |
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows annual funding in AI infrastructure startups
Can batteries and power-control systems really keep an AI campus running?
Batteries and power-control systems can already keep AI campuses stable through faults and outages, but they cannot replace continuous generation.
Form Energy’s agreement with Crusoe covers 12 gigawatt-hours of iron-air storage, with deployments expected to begin in 2027. At a steady one-gigawatt load, that amount stores about 12 hours of electricity. Spread across Form’s advertised 100-hour duration, it would support roughly 120 megawatts. Either use is valuable during grid stress, yet the batteries must eventually recharge.
Redwood Materials provides the stronger operating example today. Its 12-megawatt, 63-megawatt-hour solar-and-battery microgrid for Crusoe was built in under four months and has reported 99.2% availability. The partners are now expanding the site to support 20 additional modular data centers, nearly seven times the original computing capacity.
Redwood is proving that a smaller off-grid AI deployment can run now, while Form Energy is targeting multi-day gaps at much larger campuses.
ON.energy is becoming one of the most important electrical-control startups because its systems sit between unstable AI loads and the grid.
Crusoe has selected the company for five gigawatts of medium-voltage AI uninterruptible-power systems across several campuses, with commissioning beginning in 2026 and continuing into 2027. ON.energy also secured a separate transformer supply agreement covering five gigawatts. Large transformers can hold up an otherwise finished site for months or years.
ON.energy now has a real operating reference. It deployed an AI UPS at the US Department of Energy’s National Laboratory of the Rockies and later published independent testing showing zero-voltage ride-through performance against proposed ERCOT requirements for large electronic loads. The Crusoe rollout will test whether that performance holds at a vastly larger scale.
AI clusters can change power demand abruptly as thousands of GPUs start, stop or synchronize work. ON.energy’s equipment smooths those swings, carries loads through faults and helps a campus connect without destabilizing the surrounding network. Outside suppliers provide the energy; ON.energy controls the last electrical step into the GPU racks.
| Battery project | Disclosed scale | Practical meaning today |
|---|---|---|
| Form Energy and Crusoe | 12 GWh | Multi-day resilience for part of a large campus, with delivery still ahead |
| Redwood Materials and Crusoe | 12 MW / 63 MWh | Operating microgrid with about five hours of storage at full output |
Can grid startups unlock more AI power than a new plant?
Grid startups can sometimes release hundreds of megawatts faster than a new generator can be built, although utilities still decide whether those gains become available.
GridCARE has the most directly AI-focused example. Its work with Portland General Electric is expected to make more than 80 megawatts of incremental Hillsboro capacity available in 2026 and more than 400 megawatts by 2029. The company has since added a collaboration with National Grid to speed large-load connections in other territories.
TS Conductor tackles the physical wire. Salt River Project replaced conductors along an 8.5-mile transmission route in Phoenix and lifted the continuous current rating from 2,210 to at least 3,800 amperes. The upgrade raised continuous capacity by roughly 72% without rebuilding the existing towers, in a region where data-center growth is a major source of new load.
LineVision uses sensors and software to measure how much electricity a line can safely carry under current weather conditions. A recent Duquesne Light pilot found about 25% more available capacity, while an earlier National Grid deployment helped manage a corridor connected to a 600-megawatt offshore wind farm.
Their role is to help utilities release megawatts already trapped inside the network, often years before a new transmission line would arrive.
This chart, included in our AI infrastructure market deck, compares the main business model options for AI cloud infrastructure providers
Are hyperscalers creating energy winners or keeping them dependent?
Hyperscalers are creating powerful suppliers while making sure none becomes impossible to replace.
A large technology customer can transform a startup’s finances. Google’s commitment gave Kairos a path from demonstration reactors to a commercial fleet. Amazon gave X-energy capital and a first deployment program. Meta’s agreement lets Oklo receive development funding before the Ohio campus produces power. Oracle and Microsoft have helped Crusoe turn power-ready land into multibillion-dollar infrastructure.
The same buyers deliberately spread their bets. Amazon’s Nevada package combines Zanskar geothermal, Primergy solar and batteries, and utility delivery. Google has backed Fervo geothermal as well as Kairos nuclear. Crusoe is combining grid power, gas engines, long-duration storage and second-life batteries across different projects.
Buying several technologies at once leaves the hyperscaler with the better hand. A delayed reactor can be replaced by more gas or grid capacity at another site. An energy startup usually cannot replace a multi-gigawatt hyperscaler order so easily.
A supplier gains real bargaining power only by owning a local bottleneck that cannot be recreated quickly, such as Lancium’s grid position or VoltaGrid’s reserved engines.
Which announced AI power gigawatts should we distrust?
We should distrust every gigawatt figure until we know whether it describes operating power, contracted capacity, a development target or a non-binding option.
Crusoe’s figures show the difference. More than 200 megawatts at Abilene is live. Its 4.9 gigawatts is contracted infrastructure capacity. The wider pipeline exceeds 40 gigawatts, but that last number includes sites under negotiation and advanced development. Treating all three as equivalent would turn a strong operating story into an exaggerated one. The live figure remains the best evidence.
Nuclear announcements require even more discipline. Oklo’s 12-gigawatt Switch agreement is non-binding and stretches through 2044. X-energy’s five-gigawatt target reaches to 2039. Kairos has the smallest headline number, yet its first plant has permits and construction activity behind it.
Equipment orders sit somewhere in the middle. VoltaGrid’s 1.5-gigawatt engine order does not prove those units are installed, though it reserves scarce manufacturing capacity and supports named customer commitments. Form Energy’s 12-gigawatt-hour Crusoe agreement is also meaningful, while delivery still begins later.
We use five labels: operating, under construction, financed and contracted, contracted only, and aspirational. Startups often publish the last four together, which makes their pipelines look much more solid than they are.
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, featured in our AI infrastructure market deck, shows the share of revenue generated by each customer segment in the AI infrastructure market
What could knock today’s AI power leaders off course?
Execution, equipment and regulation pose a bigger threat to the current leaders than a shortage of customer demand.
Crusoe must repeat its Abilene build across several campuses without losing speed or reliability. Lancium must turn grid positions into energized projects while communities and utilities scrutinize cost, water and local infrastructure. VoltaGrid needs enough engines, transformers, gas connections and air permits to support a customer book that already runs into several gigawatts.
Fervo and Zanskar have to show that drilling gains continue across different geology. Form Energy must manufacture unprecedented volumes of a new battery chemistry. ON.energy must deliver a five-gigawatt rollout after proving the system at much smaller scale. Nuclear developers face the longest chain: fuel, licensing, first-of-a-kind construction and customer patience.
A hyperscaler can also postpone a campus or shift spending to another region, leaving a startup with equipment and land designed around one enormous buyer. Modular systems and diversified customers offer the best protection.
Which startups control the power behind AI today?
Crusoe leads today, followed by Lancium and VoltaGrid, while ON.energy, GridCARE, Fervo and a smaller group control narrower but increasingly valuable parts of the system.
Crusoe takes first place because it combines live AI capacity, large contracted projects and the ability to coordinate energy with data-center construction. Lancium follows because its grid-connected land enables several of those projects. VoltaGrid earns third place with the largest credible near-term package of behind-the-meter generation.
ON.energy is the strongest emerging control-layer company. Its equipment can determine whether gigawatt-scale AI loads survive faults and connect safely. GridCARE can bring forward usable capacity without waiting for a new plant. Fervo is the clean firm-power leader, though its first hundreds of megawatts remain far below the scale of the largest AI campus plans.
The nuclear startups still control options rather than electricity. Kairos has the best execution record, X-energy the strongest industrial coalition and Oklo the biggest customer headlines. Their influence becomes real when commercial reactors begin delivering power.
Utilities, pipeline owners, regulators and hyperscalers still set the outer limits of the market. The companies at the top have found the places where those incumbents move too slowly: power-ready sites, modular generation, integrated campuses and grid-safe electrical systems. They now have real control over which AI projects get built, even though none controls the whole machine.
| Rank today | Startup | What it controls now | Why it ranks here |
|---|---|---|---|
| 1 | Crusoe | Integration of power, campuses and live AI infrastructure | Live hyperscale campus plus multi-gigawatt contracts |
| 2 | Lancium | Gigawatt-scale sites and grid access | Owns power-ready land beneath several major Texas campuses |
| 3 | VoltaGrid | Fast behind-the-meter generation | Largest named behind-the-meter order book among startups |
| 4 | ON.energy | Power quality and grid-safe delivery | Large Crusoe rollout plus independent testing |
| 5 | GridCARE | Faster access to existing grid capacity | PGE project plus a newer National Grid collaboration |
| 6 | Fervo | New clean, firm generation | Financed and fully contracted Cape Station build |
If you want more recent data on this point, please see our latest AI infrastructure market report.
This chart, included in our AI infrastructure market deck, shows how GPU cloud infrastructure technology has evolved over time
OUR METHODOLOGY
This analysis asks which startups can actually influence whether an AI campus gets powered, connected and kept online. We separated the market into the physical bottlenecks that decide delivery: power-ready land, grid access, on-site generation, electrical reliability, storage, clean firm power, transmission upgrades and construction speed.
We prioritized recent operational evidence over reputation. Energized capacity carried the most weight, followed by construction, project financing, permits, approved grid positions, reserved equipment, utility partnerships, binding customer commitments and independent technical validation.
We ranked companies by control rather than by valuation or technological novelty. The central test was whether a startup’s absence could delay a campus, force it to another location or leave installed computing equipment unable to operate reliably.
We also separated announced capacity into five categories: operating, under construction, financed and contracted, contracted only, and aspirational. This prevents a live megawatt, an equipment order and a non-binding 2040s development target from being treated as the same thing.
Key sources used for the market backdrop and company evidence include the International Energy Agency’s Energy and AI work, EPRI’s Powering Intelligence outlook, the US Department of Energy’s work on data-center electricity demand, company disclosures from Crusoe, Lancium, VoltaGrid, ON.energy, GridCARE, Fervo Energy, Zanskar Geothermal, Kairos Power, X-energy, Oklo, Form Energy, Redwood Materials, TS Conductor and LineVision, alongside regulatory and grid materials from the Nuclear Regulatory Commission and ERCOT.
In our AI infrastructure market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- Which startup is redefining how data centers are powered?
- Is power replacing chips as AI’s bottleneck?
- What is the biggest AI bottleneck today?
