Small reactors meet big AI ambitions

A surge of corporate interest is steering billions toward small nuclear power—even as timelines and costs lag the AI boom.

Tech giants are signing power from reactors that don’t exist yet. Google has reserved up to 500 megawatts from Kairos Power by 2035. Oklo outlined a 12-gigawatt supply deal with data-center developer Switch. Amazon invested more than $1 billion in nuclear projects last year. Since 2023, companies have announced commitments totaling over 32 gigawatts of proposed small modular reactor capacity—roughly the output of 30 large plants. The promise is firm, carbon-free baseload for the cloud. The reality is trickier.

None of these advanced reactors will deliver power before 2027 at the earliest; most point to the early 2030s. That timing collides with AI’s immediate load growth, forcing near-term reliance on efficiency gains, renewables, and the grid as it is. Long term, though, the math is hard to ignore: data-center electricity demand is expected to more than double by 2035, approaching ~945 terawatt-hours a year—more than France or Canada use in total. For executives staring at uptime SLAs and climate pledges, nuclear looks like insurance. Reliability matters.

What’s actually new

Small modular reactors aren’t a new idea; the shift is who’s showing up to buy. In the past four years, U.S. agencies have committed more than $6 billion to SMR developers through cost-shares and loans, while private investors added over $3 billion. Corporate buyers are now attaching themselves to first units, not just studies.

A flagship example: Google, the Tennessee Valley Authority, and Kairos Power structured a plan to route a first 50-megawatt Kairos unit into Google’s Southeast data centers around 2030, backed by hundreds of millions in federal support. That would be the first U.S. utility purchase from a Generation IV reactor for commercial use. It’s a beachhead, not a flood. Still, it’s new.

New designs, old goal: firm power

The reactor toolbox is diverse. Kairos uses molten fluoride salt at atmospheric pressure, removing high-pressure steam risks and aiming for “passive” safety. Its fuel is golf-ball-sized pebbles that encase uranium in ceramic and carbon layers designed to tolerate high heat without failing. Oklo goes smaller: 15-megawatt microreactors using sodium coolant and heat pipes, sized to fit a shipping-container footprint with multi-year refueling intervals. TerraPower, backed by Bill Gates, is building a 345-megawatt sodium-cooled fast reactor in Wyoming. X-energy’s 80-megawatt high-temperature gas design targets industrial hosts like Dow.

The shared pitch: factory-built modules, simpler systems, faster installs, and round-the-clock output that matches data-center duty cycles. In theory, that’s perfect for AI inference farms that hate downtime. In practice, first-of-a-kind projects rarely run like slide decks. Caveat noted.

The cost problem surfaces

Independent projections peg early SMR power around $182 per megawatt-hour by 2030—above large nuclear (~$133/MWh) and combined-cycle gas (~$126/MWh), and roughly 30% higher than onshore wind or solar paired with batteries on average. Scale still matters.

History helps explain the caution. Russia’s and China’s operating SMRs ran 300–400% over original estimates. In the U.S., NuScale’s first project unraveled in 2023 when costs jumped from roughly $58 to $89/MWh and the total bill reached $9.3 billion for 462 megawatts. Municipal partners walked; the plan died. That wasn’t about physics. It was about finance, supply chains, and risk.

Developers counter that first units are tuition. The curve improves with replication—dozens of identical reactors per year, not one bespoke colossus per decade. If a real manufacturing cadence emerges, 60-year life and near-perfect capacity factors could narrow the premium. That’s the thesis. It’s unproven.

The fuel supply bind

Many advanced designs need HALEU fuel—uranium enriched to 15–20% U-235 versus the 3–5% used today. HALEU shrinks cores and stretches refueling intervals, which is great for modular plants and isolated campuses. It’s also scarce. Russia was the key supplier before 2022; sanctions slammed that door.

The U.S. is funding a domestic HALEU restart and offering limited federal stockpiles to seed first builds. Even so, logistics bite. No high-capacity transport cask for HALEU is yet certified in the U.S.; moving enough fuel with small, approved containers could take thousands of shipments for a single core. Some developers have floated temporary use of lower-enriched fuel to stay on schedule. Fuel is the choke point.

Reliability anxiety meets grid reality

Data-center operators live in fear of blips. A 2022 substation failure near Heathrow grounded more than a thousand flights; airports and hospitals are now studying on-site firm power, including SMRs. The logic for hyperscalers is similar: co-locate a dedicated unit, island the campus during grid stress, and turn an energy liability into a controllable asset. That’s the dream. The gap is real.

The calendar, however, won’t bend. Oklo’s earliest target is 2027 for a 15-megawatt unit, assuming smooth licensing and siting. Kairos is aiming at 2030 for its first commercial module. TerraPower and X-energy point to the turn of the decade as well. Between now and then, hyperscalers will squeeze efficiency, lock in more wind and solar, extend existing nuclear, and, in some regions, install new gas turbines as a reliability backstop. It’s a portfolio, not a silver bullet.

The credibility test

This sector’s fate hinges on a handful of first builds. Hit cost and schedule, and SMRs can graduate from MOUs to purchase orders—and from purchase orders to campus staples by the mid-2030s. Miss, and the dot-com analogy that skeptics invoke will harden into memory. The dollars are real. So are the hurdles.

Why this matters

• Tech is pre-buying firm, carbon-free power to secure AI-era capacity, effectively underwriting first-of-a-kind nuclear builds years before electrons flow.
• The first SMRs are a referendum on nuclear’s new playbook—factory modules, novel fuels, faster schedules. Their cost and timing will decide whether small nuclear becomes standard data-center gear or a cautionary tale.