From gigawatt superclusters and orbital data centres to 2MW edge deployments and geothermal campuses, the AI infrastructure buildout is reorganising around a single constraint: clean, firm power.
Every conversation about artificial intelligence eventually ends with recognition of one hard physical fact: the machines behind it are demand high levels of electricity.
Training and running frontier models has turned the data centre into one of the fastest-growing sources of power demand on the planet, and the hyperscalers, cloud providers, developers and grid operators racing to feed it are all bumping into the same ceiling.
Clean power that runs around the clock, not just when the wind blows or the sun shines, has become the scarcest resource in the AI economy.
Goldman Sachs Research estimates US data centre power demand will more than double from roughly 31GW in 2025 to about 66GW by 2027, by which point data centres would consume 8.5% of the country’s peak summer electricity, up from 4.1% in 2025.
Globally, it has forecast that data centre power demand could climb 165% by the end of the decade against 2023 levels. The catch is delivery: Goldman’s analysts reckon only around half to 60% of the capacity scheduled over the next year or two will actually come online on time, hobbled by permitting delays, equipment shortages and a grid never built for load growth on this scale.
Nothing illustrates how central that constraint has become quite like the roadshow that preceded SpaceX’s June listing on Nasdaq. The company, best known for rockets, devoted an entire slide of its investor presentation to the proposition that “terrestrial capacity faces significant limitations”, telling prospective shareholders that US data centre demand hit 62GW in 2025 against just 49GW of supply, a shortfall it expects to widen further.
US electricity generation, SpaceX noted, has grown at less than 3% annually since 2023 after fifteen years of near-total stagnation, while China’s output has compounded at roughly 6%.
The economics on display in the same deck explain why so much capital is chasing the power problem.
SpaceX disclosed that Anthropic is paying $1.25 billion per month for access to compute capacity across its Colossus clusters, which the company describes as the world’s largest coherent supercomputer, with roughly a gigawatt of nameplate compute draw through May 2029, while Google has agreed to pay $920 million per month from October 2026 for access to 110,000 Nvidia GPUs.
SpaceX’s capital expenditure nearly doubled to $20.7 billion in 2025, with AI infrastructure the largest and fastest-growing component. When compute can be pre-sold at those rates, years ahead of delivery, the bottleneck is no longer demand or chips.
Meta joining the party by saying it would sell excess compute shows there is ample opportunity in compute supply. The question is, how will it be powered?
An industry reorganising around the constraint
The rest of the industry is contorting itself around the same shortage. And the activity now spans from gigawatt superclusters to modular edge sites.
Nebius, the Nasdaq-listed AI cloud company, this month launched an asset-light partnership model that effectively outsources the hardest part of its growth: partners finance, own and operate the data centres, while Nebius supplies the architecture, software stack and, crucially, the customers.
The pitch to data centre developers, infrastructure investors and national AI projects is that Nebius brings the demand, so partner capacity starts generating a return the moment it goes live.
It is a telling inversion: the scarce input is no longer software or sales, but powered floor space, and Nebius is willing to share economics with anyone who can deliver it.
Demand-side behaviour tells the same story. Reflection AI, the US startup building open frontier models, signed a $1 billion compute agreement with Nebius just weeks after inking a separate deal for access to SpaceX’s computing resources, a double-sourcing strategy that has become standard practice as AI labs scramble to lock in capacity wherever it exists.
Compute is being bought like a commodity in shortage: from multiple suppliers, on long contracts, well in advance.
The same dynamics are playing out at the small end of the market. Nasdaq-listed Duos Technologies this month signed a Master Services Agreement with Nistar, a Chicago-based development platform for AI and hyperscale data centre infrastructure, to deploy up to 2MW of critical IT-load capacity at Duos’ Columbus, Georgia campus, enough, Nistar says, to support an institutionally backed AI-compute client’s planned deployment of 1,024 Nvidia B200 GPUs. The agreement adds to an existing 10MW contract at the site, expected to be billed by the end of August, with a further 10MW planned for installation between October and November.
Nistar, which straddles grid-connected and behind-the-meter energy strategies, powered land development and real-asset finance. described the deal as “the first of what we expect will be multiple deployments” and said it plans to replicate the model at additional locations.
The very existence of firms like Nistar, structuring power, real estate, GPUs and capital into a single financeable package, shows how thoroughly the industry has reorganised around energy availability: even a 2MW deployment now comes power-anchored by design.
The space layer
Connectivity is the other piece of infrastructure being rebuilt for the AI era, and here, too, the startups are scaling fast. SWISSto12, the Swiss manufacturer using 3D printing to produce satellite payloads and its compact HummingSat geostationary platform, this month closed a $70 million Series C on the back of $140 million of 2025 revenue, more than $500 million in contracted orders and a 110% compound annual growth rate since 2022, alongside an $84.8 million award from European Space Agency member states for HummingSat’s development and in-orbit validation.
The company has secured seven HummingSat contracts with operators, including SES and Viasat, and has more than 2,000 of its HummingLink payload and antenna solutions deployed in orbit, supporting missions from direct-to-device connectivity to intersatellite data relays. “Space is increasingly recognised as essential infrastructure for the global economy,” said CEO and founder Emile de Rijk.
SpaceX is making the same argument at a different order of magnitude.
Its answer to the terrestrial power shortfall is characteristically extreme: put the data centres in orbit, where solar energy is unlimited and radiative cooling is free.
The company told IPO investors it intends to begin deploying AI compute satellites on Starship as early as 2028, delivering 100KW of compute power per metric ton of payload, with the existing Starlink network handling data routing between compute clusters and end users on Earth. Between SWISSto12’s intersatellite relays, Starlink’s laser mesh and the prospect of orbital compute itself, the satellite layer is evolving from a communications afterthought into a load-bearing part of AI infrastructure.
Why geothermal, why now
For the terrestrial buildout, though, the mismatch remains: abundant demand chasing scarce, firm, around-the-clock supply. That is precisely the gap a new generation of geothermal companies is wagering it can fill, and on 17 July, one of them recruited someone well positioned to know whether the bet can work.
Vesari Inc., a Miami-based startup majority-owned by London-listed intellectual property group Tekcapital plc (AIM: TEK), announced that Dr Joseph N. Moore, among the world’s foremost authorities on the geology and geochemistry of geothermal systems, has joined its newly formed Science Advisory Board. For more than a decade, Moore ran Utah FORGE, the roughly $300 million Department of Energy-funded field laboratory that has become the flagship of America’s push to make enhanced geothermal systems commercially viable.
“Dr Moore’s work has helped define the modern scientific understanding of geothermal systems,” said Clifford M. Gross, Vesari’s executive chairman. “As Vesari seeks to build a new generation of geothermal-powered, hyperscale AI infrastructure with its proprietary technology, having a scientist of Joe’s standing to help guide our subsurface and reservoir strategy is invaluable.”
Geothermal has long been the quiet member of the renewable family. Enhanced geothermal systems, the approach Moore has spent his career advancing, aim to change that by drilling into hot rock and engineering the permeability needed to circulate water and harvest heat almost anywhere.
The appeal for AI is obvious: geothermal produces firm, baseload power that doesn’t stop at sundown.
“Geothermal energy is uniquely suited to deliver the clean, firm, around-the-clock power that AI computing now demands,” Moore said, calling Vesari’s plan to wire baseload geothermal generation directly into high-density computing “a compelling application of the science.”
Rather than draw from a grid already straining under AI’s appetite, the company wants to build “behind-the-meter” campuses in which the power plant and the compute live on the same site, sidestepping the interconnection queue that Goldman identifies as the industry’s chief delivery risk, and taking the power-anchored logic that developers like Nistar are applying at the edge to its natural conclusion.
The Fervo proof point
The public markets have already shown they will pay for exactly this thesis. In May, Houston-based Fervo Energy, the best-known name in enhanced geothermal, pulled off the largest clean-energy IPO in Wall Street history, raising around $2.2 billion, including the over-allotment and closing its first day with a market value just above $10 billion.
That valuation was handed to a company with roughly $138,000 of 2025 revenue, a $57.8 million net loss, and a flagship plant not expected to deliver first power until late 2026.
An act of faith underwritten by backers including Bill Gates’s Breakthrough Energy Ventures and offtake commitments spanning a framework agreement with Google for up to three gigawatts, plus deals with Shell and Southern California Edison. The shares have since settled around $35, valuing the company at nearly $10.3 billion.
Fervo proved a geothermal company at the intersection of clean firm power, and AI compute can command a multibillion-dollar valuation years before profitability. And as SpaceX’s pre-sold compute agreements and Reflection’s forward purchases demonstrate, the industry has normalised paying for capacity long before it exists.
Eleven patents, one integrated bet
Moore’s appointment lands two days after Tekcapital, which holds 51% of Vesari, told the London market that the company had acquired the rights to 11 non-provisional patent applications filed with the U.S. Patent and Trademark Office (USPTO).
The portfolio doesn’t protect a single gadget: the applications span integrating geothermal generation with the data centre; using the geothermal loop to cool dense racks of AI chips; recovering waste heat to augment power output; running the site as an “islanded” electrical system independent of the grid; scheduling AI workloads to match available energy; even routing traffic over low-Earth-orbit satellite links in place of terrestrial fibre, an architecture that slots directly into the satellite connectivity layer companies from SpaceX to SWISSto12 are now building out, plus a commercial layer for selling verified carbon-free compute at a premium.
Vesari describes a closed-loop system in which power, cooling, compute, pricing and utilisation form “a reinforcing feedback loop”, with the patents designed to fence off the interaction between the parts, not just the parts themselves.
“AI is the fastest growing technology sector in the world. To meet its promise and continue its exponential growth requires abundant clean energy that neither increases the cost of electricity nor damages the environment,” said Louis Castro, a Tekcapital director, adding that the filings put “the intellectual property foundation for the business in place” in the part of the market “where significant value in this emerging sector will be created.”
Tekcapital also disclosed that Vesari has opened preliminary talks with US investment bankers about reaching the public markets, potentially through a reverse merger into a listed shell or a de-SPAC transaction paired with a concurrent financing — conversations the company stressed are early and exploratory, with no certainty any deal will proceed.
As with all fast-moving companies and technologies employed to meet AI compute demand, there are risks in Vesari and geothermal. The patents are filed, not granted. Vesari has yet to build a campus, and enhanced geothermal, despite genuine momentum at sites like FORGE, is still moving from demonstration toward scale.
But the timing is right for the company, with geothermal looking set to slide nicely into the suite of options detailed here that will power the AI revolution.