AI Data Center Power: Firma Bundles Renewables Into One Contract

A startup argues that clean power has a paperwork problem

As AI data centers draw more electricity, the usual explanation for why so many of them run on natural gas is that wind and solar are too slow, too unreliable or too costly. Firma Power, a young company, says that explanation is mostly wrong. In its view, the clean electrons already exist. What's missing is an easy way to put them together. The company built software that sizes a planned data center's power needs, searches proposed clean-energy projects and assembles a portfolio of wind, solar and storage that can serve the site. The data center developer then signs one contract instead of negotiating with each project separately.2

The two outlets that covered the launch tell the same story. Both describe the barrier as fragmentation: many small projects have to be stitched together into a single, reliable supply, and that is hard work.12 Jesse Jenkins, a Princeton energy-systems engineer and Firma's CTO, said the company started from a simple observation. About 98% of what sits in interconnection queues is wind, solar and storage, yet that capacity is not the default way large new loads get power. That gap, he said, points to market friction, not a shortage of supply.2 Jenkins co-founded the company with president Scott Burger, a Form Energy veteran, and CEO Brent Alderfer, a wind and solar developer. Firma raised about $1.4 million in seed funding in February 2026.2

The money is small for a sector where projects cost billions. Firma, though, is not trying to build power plants. It wants to be the layer that matches developers who need buyers with data centers that need power. It says it is working on roughly a dozen projects with hyperscalers, AI cloud providers and data center developers, though it has not disclosed details of any specific deal.2

The scale problem behind the pitch

Firma's argument rests on how much larger data centers have become. A data center once counted as "large" at around 50 megawatts. A single AI facility can now need 500 megawatts or even a gigawatt, which is roughly what San Francisco uses.2 Technical coverage traces this to the chips. Conventional server racks draw about 5 to 10 kilowatts, while AI racks need 40 to 100 kW.5 Nvidia's GB200 NVL72 racks draw about 120 to 140 kW, which means older facilities can't host them without a full redesign.22

Forecasts of total demand vary widely. Gartner expects global data center electricity use to reach 565 terawatt-hours in 2026, up 26% from 447 TWh in 2025. It also expects AI-optimized servers to use more power than conventional servers in 2027.27 Other analyses claim global consumption will pass 1,000 TWh by the end of 2026 and attribute that figure to the IEA.30 A peer-reviewed review, however, cites the IEA projecting about 945 TWh by 2030, not 2026.8 The more conservative Gartner trajectory is the better guide here. The 1,000-TWh-in-2026 claim seems to mix up a 2030 projection with a current-year figure. Even the lower numbers describe a power system under strain.

In the United States, Goldman Sachs projects data center demand rising from 31 gigawatts in 2025 to 41 GW in 2026, which would take data centers from 4.1% to 5.3% of peak summer demand.30 Lawrence Berkeley National Laboratory expects U.S. data center consumption of 325 to 580 TWh by 2028, up from about 176 TWh in 2023.28

The grid is feeling it

The stress is already measurable. The North American Electric Reliability Corporation issued its highest-urgency Level 3 alert on May 4, 2026, after gigawatt-scale data center loads disconnected within seconds. That kind of event can cause frequency swings similar to a large generator tripping offline.28 In PJM, the largest U.S. grid operator, the independent market monitor attributed roughly 40% of a record capacity-auction cost to data center demand. PJM has said typical customer bills could rise about 1.5% to 5% starting in June 2026.28 Deloitte reports that some fast-growing data center regions have already seen harmonic distortions, load-relief warnings and near-miss incidents.29 Gartner has predicted that power shortages will limit 40% of AI data centers by 2027.21

Under this pressure, developers have mostly turned to gas. One energy-markets analyst said large gas generators are what most data centers use today. Builders often look for sites near pipelines, and turbine makers such as GE Vernova and Siemens have seen orders rise.5 Some projects skip the grid entirely. Chevron and Microsoft agreed to a 20-year deal for off-grid natural gas supplying a 2.67-gigawatt power plant and data center near Pecos, Texas.7 One analysis described on-site gas generation as likely to remain a permanent part of digital infrastructure through at least the end of the decade.7 Nuclear has had a lot of attention, including Constellation's $1.6 billion restart at Three Mile Island for Microsoft. One analyst, though, does not expect new nuclear options to be ready for AI data centers until 2030 to 2035.57

Firma's pitch targets exactly this gap. The U.S. is on pace to add a record 86 GW of wind, solar and battery capacity this year, roughly the size of the U.K.'s entire grid.2 Burger said many queued projects fail for an ordinary reason: they have no buyer, and without a long-term contract a developer can't get financing.2 Firma calculates that around 200 GW of clean capacity could be added across just three grids (PJM, SPP and MISO) by 2030 while still meeting reliability requirements.2 Burger also said customers report being told by utilities to wait until 2032 for a new gas plant, while Firma says it can produce workable power strategies in days.2

Where the reporting diverges

The coverage agrees about the problem but differs on the solutions. Firma's approach keeps customers on the shared grid, which it presents as an advantage over behind-the-meter gas plants.2 Other startups are going the opposite way. Rune, which raised a $40 million Series A, places modular data centers directly at solar farms to use power that would otherwise be curtailed. It says a typical solar plant wastes up to 20% of what it generates, and it focuses on inference workloads.9 In China, Envision says more than 80% of the power for its Galaxy Campus in Ulanqab will come straight from its own wind farms, bypassing the grid.10

A Harvard Business School working paper adds a market view. From 2023 to 2025, data center returns explained about 34% of daily movements in clean-energy stocks, compared with under 9% for traditional energy.6 Investors, in other words, seem to treat renewables as a real beneficiary of AI load growth, even though gas dominates actual new hookups.

Cooling is part of the power bill

Energy supply is only half of the picture. Gartner's breakdown shows cooling and other infrastructure using about 195 TWh in 2026, up 22.6%, and that line is growing faster than conventional servers.27 As rack densities rise, air cooling is running out of headroom. JLL's data center lead said densities beyond 70 kW have outrun what air can physically handle, leaving direct-to-chip or immersion liquid cooling as the realistic options.20

Adoption estimates vary here too. TrendForce expects liquid cooling to reach 53% of AI chips in 2026, up from about 33% in 2025.18 Goldman Sachs forecasts that 76% of AI servers will be liquid-cooled in 2026.12 Cold-plate, or direct-to-chip, systems make up roughly 70% of the liquid-cooling market. Immersion is gaining ground in the densest facilities.15 One study suggests advanced cooling could cut cooling energy by up to 50%.11 Retrofits are difficult, though, so purpose-built AI facilities are better placed to adopt it.20

The reading

Firma's main claim is a reasonable one. The constraint on clean power for AI may now be deal-making and contracting rather than physics or cost. A record build-out of renewables alongside utilities quoting 2032 for gas suggests the bottleneck is in how deals are structured. Firma's case is still unproven: the round is small, the company hasn't named any deals, and its 200 GW figure is a theoretical ceiling, not a forecast.

The more useful takeaway is that power for AI is splitting into several approaches: gas plants, grid-tied clean portfolios, co-location at solar farms and direct wind supply. Other levers could make each of those go further. Deloitte cites a study finding that if data centers cut just 1% of their annual load, grids could connect about 126 GW of new demand with little expansion.29 Liquid cooling also reduces how much power each unit of computing needs. If Firma can turn a scattered set of queued projects into supply that data centers can actually contract for, it won't replace those levers, but it could make clean energy the quicker option before 2030 rather than the more principled one.