AI isn’t just eating software; it’s eating electricity. The IEA now projects global data centers will consume about 945 TWh by 2030, roughly double today—nearly Japan’s total power use. The surge is overwhelmingly driven by AI inference and training. That changes the grid’s optimization target from “cheapest kWh” to “reliable 24/7 clean MWh, at scale.” IEA+1

And that’s exactly why three technologies are suddenly in the pole position for the AI era: fusion (yes, with a real PPA), enhanced geothermal, and small modular reactors (SMRs). Here’s how each stacks up—and who’s most likely to keep GPUs lit without torching climate targets.

Bottom line: AI has turned firm, carbon-free electricity from “nice to have” into the currency of digital growth.

Fusion grows up: Helion’s Orion + Microsoft

Last week, Helion began construction on Orion, a fusion power plant in Malaga, Washington, with plans to deliver electricity to Microsoft under a previously announced PPA target date of 2028. The project sits near existing grid infrastructure (Rock Island Dam) and will migrate learnings from Helion’s Polaris prototype to the commercial machine. Even if early output is modest, the signal is seismic: fusion is moving from R&D into customer-driven delivery, pulled by AI demand. Reuters+1World Nuclear NewsS&P Global

Strengths: Potentially compact, zero-combustion, and—if it works—dispatchable with high power density close to loads. Risks: Physics and engineering timelines; first-of-a-kind cost curves; regulatory and insurance unknowns. In other words: promising, but not bankable at hyperscale until Orion proves net-positive output, capacity factor, and LCOE at commercial scale.

Geothermal gets specific: Fervo’s 24/7 data-center corridor

Geothermal’s renaissance is less flashy but far more grid-real. Fervo Energy closed $206M in new financing to accelerate Cape Station (Utah) and laid out a “Data Center Corridor” concept: co-locating enhanced geothermal systems (EGS) with fiber and transmission to feed 24/7 clean power straight into AI campuses. Enhanced geothermal has always promised firm, weather-proof output; what’s new is execution aligned to AI siting and interconnection realities. fervoenergy.com+1

Authoritative analyses back the thesis. The IEA’s 2025 special report underscores geothermal’s vast untapped potential as firm clean power; NREL’s ATB charts cost trajectories; MIT’s canonical study laid the groundwork for EGS’ baseload promise. Together, they sketch a path where drilling tech and oil-and-gas capabilities translate directly into scalable, 24/7 carbon-free electricity (CFE) near load. IEA Blob Storageatb.nrel.govMain

Strengths: Firm CFE, modular well-by-well scaling, compatible with behind-the-meter siting that dodges interconnection queues—catnip for data centers. Risks: Subsurface uncertainty, induced seismicity management, and supply-chain depth for high-temperature tools. Still, among near-term options, EGS looks most “productizable” for AI campuses.

SMRs leave the powerpoint: Ontario’s BWRX-300

Canada just crossed a concrete threshold. OPG received a License to Construct the first BWRX-300 SMR at Darlington; provincial approval for construction followed, with the first unit targeted around 2030. It’s the clearest Western deployment track for modern nuclear, with Poland actively progressing a parallel fleet strategy via OSGE. For cloud operators, SMRs offer industrial-grade reliability with known regulatory pathways and long asset lives. ans.orgWorld Nuclear News

Strengths: Proven nuclear fundamentals with simplified, modularized design; grid-scale capacity that scales to multi-GW campuses. Risks: First-of-kind delivery risk on schedule/cost; public acceptance; long development cycles vs. AI’s breakneck buildout.

Callout: If you need hundreds of MW of firm CFE in one place before 2030, SMRs are the only mature nuclear path with a real construction license in the West today. ans.org

The supporting cast: LDES becomes default, not novelty

Even with firm power, long-duration energy storage (LDES) is the buffer that makes portfolios work—bridging weather events and price spikes, firming hybrid supply, and squeezing more out of congested interconnects.

  • Form Energy has started initial deliveries/build-out for its 100-hour iron-air system serving Minnesota, with first commercial operation targeted by end-2025. If costs land as projected, this is the missing piece for wind-still winter weeks in the Midwest and Northeast. Great River EnergyWest Virginia Public Broadcasting
  • Energy Vault brought a 57 MW / 114 MWh BESS online in ERCOT (Cross Trails) in June 2025—further proof that storage is standard grid infrastructure now, not just pilots. Solar Power World OnlineBusiness Wire
  • On chemistry diversity, CATL’s sodium-ion “Naxtra” heads to mass production in December 2025, which could lower costs and de-risk supply chains for grid storage. Reuters

A quick reality check: Carbon removal is not baseload

Policy headwinds and technology realities remind us: DAC is not a power plant. Occidental’s 1PointFive won Class VI permits for its STRATOS DAC site in Texas with operations eyed for 2025—a regulatory milestone, yes—but Climeworks’ layoffs and delays show how fragile the business model remains without strong policy and bankable offtakes. Important for net-zero portfolios, but irrelevant to keeping AI racks online tomorrow. ReutersUS EPAThe Guardian

So…who wins the AI baseload race?

Near term (now–2028):

  • EGS geothermal is the most deployable firm CFE that matches the AI siting problem (corridors, behind-the-meter, build-as-you-drill).
  • SMRs offer scale and durability for mega-campuses, with Ontario’s BWRX-300 as the pathfinder—valuable signal to regulators and financiers.
  • LDES becomes mandatory glue across portfolios.

Medium term (2028–2032):

  • If Helion’s Orion demonstrates meaningful, price-competitive output into the Microsoft PPA, fusion jumps from science project to category disruptor. Until then, expect EGS + SMR + LDES hybrids to dominate AI power procurement. Reuters
Hot take: In the 2020s, AI won’t pick a single winner—it will pay for redundancy. Expect layered PPAs: EGS baseload + SMR tranche + LDES hedges, with peakers only as last-resort insurance.

Why This Matters

AI isn’t “just another load spike.” It’s persistent, location-constrained, and brand-sensitive—and it’s arriving faster than grid upgrades. That forces hard choices: permit firm clean power near people, or watch AI migrate to wherever it’s easiest to plug in—even if that means dirtier grids. The winners will be regions that streamline siting for firm CFE (EGS/SMR), fast-track interconnects, and treat LDES as critical infrastructure, not a pilot line. IEA

What leaders should do next

1) Procure for 24/7 CFE, not annual averages

Shift from “100% renewable energy by certificates” to hourly-matched power contracts. It aligns incentives with real grid decarbonization and reliability for AI workloads.

2) Build dual-track firm power portfolios

Don’t make a religion out of a single tech. Pair EGS for near-site baseload with SMR capacity for scale, and LDES to handle weather and price volatility. Fusion becomes a high-alpha option when proven.

3) Co-locate data + energy

Follow Fervo’s corridor playbook: co-site fiber, substations, and firm CFE. Behind-the-meter projects can dodge interconnection queues and politics. fervoenergy.com

4) Get serious about permitting

Copy Ontario’s predictable licensing cadence and absolute clarity on step-gates. Don’t strand developers in multi-year limbo. ans.org

Sources for deeper grounding (external, non-news)

  • Read the IEA’s 2025 “Energy & AI” analysis for the 945 TWh data center demand pathway. IEA
  • Explore NREL’s 2024–2025 geothermal ATB and technical baselines for cost curves. atb.nrel.gov
  • Revisit MIT’s foundational “Future of Geothermal Energy” report—still the EGS north star. Main

Conclusion

AI is forcing a return to baseload—but cleaner and smarter. In the scramble, EGS geothermal feels like the pragmatic starting gun, SMRs deliver industrial reliability, and fusion is the wildcard with a customer breathing down its neck. The real winners won’t pick one—they’ll orchestrate portfolios that match hourly demand, site intelligently, and move at the speed of AI.

“The grid doesn’t care how cool your model is. It cares whether you can keep the lights on—cleanly, every hour.”