AI data centres need 300 MW to 1 GW+ of continuous power per campus — running around the clock, zero tolerance for intermittency. Every infrastructure planner eventually hits the same question: if nuclear is the ideal firm power source, why not just deploy solar and wind with batteries instead?
This article answers that question with a structured comparison of every realistic option across five dimensions that actually matter: deployment timeline, cost per MWh, 24/7 reliability, carbon profile, and scalability. It’s part of our comprehensive series on the nuclear-AI power story driving the buildout, where we examine every dimension of the nuclear renaissance reshaping AI infrastructure.
One thing you probably haven’t heard enough about: enhanced geothermal systems are deploying faster than any other firm-power alternative. And yes, we address the AI bubble risk directly — not buried at the bottom where you might miss it.
What does “baseload power” actually mean and why does it matter for AI data centres?
Baseload power — also called firm or dispatchable power — is electricity generated continuously at stable output regardless of weather or time of day. AI training workloads behave like a flat 24/7 industrial load. GPU clusters run continuously. You cannot pause them for clouds or a windless night without wasting compute and breaking latency SLAs.
The key metric here is capacity factor: the percentage of time a source generates at rated output. Nuclear runs at approximately 90%. Geothermal hits approximately 85%. Onshore wind reaches 35–40%. Solar manages only 20–25% — which means a 500 MW solar farm actually delivers roughly 100–125 MW of firm power, and nothing after dark.
LCOE (levelised cost of energy, in $/MWh) is the standard lifetime cost metric used throughout this comparison. Its main limitation: it does not capture the cost of intermittency. A lower-LCOE intermittent source can easily cost more all-in once you factor in storage or backup generation.
Why can’t solar and wind solve the AI power problem?
Solar and wind are the cheapest electricity sources ever built — wind at $30–60/MWh LCOE, solar at $26–50/MWh. But cheapness only matters if the power arrives when the workload needs it. And AI training runs 24/7.
Battery Energy Storage Systems (BESS) are the obvious pairing. The problem is the maths. At current prices, BESS for a 5 GW facility would run over $5 billion — before degradation. And that still doesn’t cover multi-day low-generation events: cloudy weeks, calm wind patterns. Tim Fist of the Institute for Progress puts it plainly: solar and wind with BESS for full data centre reliability is not feasible at current technology and cost.
Wind has additional headwinds in the US. The Trump administration has actively challenged offshore wind permits, removing it from near-term planning for most US operators.
None of this means solar and wind have no role. They’re appropriate for flexible-workload facilities and supplemental capacity. The argument here is specifically about 24/7 high-uptime AI training campuses.
How does nuclear compare on timeline, cost, and reliability?
Nuclear’s case rests on one defining characteristic: approximately 90% capacity factor, carbon-free, available at scale. Everything else in this comparison is measured against it. The question is which version, on what timeline, at what cost.
Existing plant restarts are the fastest nuclear path. Microsoft’s 20-year PPA with Constellation Energy to restart Three Mile Island Unit 1 secured 835 MW at roughly $1.6 billion — far cheaper than a new build. Meta signed a comparable deal with Constellation for the Clinton Clean Energy Center; Google is purchasing 615 MW from Iowa’s Duane Arnold plant via a 25-year NextEra PPA, back online by 2029. The catch: this supply is finite. The number of sites amenable to reopening is perhaps a handful, and the best opportunities are already taken.
New nuclear and SMRs are a different timeline entirely. Google signed the world’s first corporate SMR purchase agreement with Kairos Power — 500 MW across six or seven molten salt reactors, first unit by 2030, full deployment by 2035. Amazon is backing X-energy‘s gas-cooled SMR programme, targeting 5 GW by 2039. TerraPower broke ground on its 345 MW Natrium reactor in Wyoming in June 2024, targeting 2030.
Current SMR LCOE sits at $89–102/MWh — competitive when you factor in capacity factors exceeding 95%, but NuScale‘s cancellation is the cautionary reference case. Projected build costs rose from $3.6 billion for 720 MWe in 2020 to $9.3 billion for 462 MWe in 2023 before the project was cancelled. Treat SMR cost targets as aspirational until a commercial plant actually demonstrates them.
For a deeper dive on reactor designs and the hyperscaler deals driving them, see SMR technology in detail.
Is geothermal a genuine alternative to nuclear for AI infrastructure?
Most nuclear-AI coverage treats geothermal as a footnote. It shouldn’t. Enhanced Geothermal Systems (EGS) operate at approximately 85% capacity factor, are carbon-free, and are dispatchable — they behave like nuclear, not like solar. And they can be deployed in 2–4 years, not 2030–2035+.
EGS uses hydraulic stimulation to create permeability in hot dry rock, extracting heat without natural hydrothermal resources. It’s the same directional-drilling playbook that unlocked unconventional oil and gas. The DOE is direct about this: EGS offers the greatest potential for power generation per drilled foot of next-generation geothermal technologies, opening approximately 25% of global land area to viable development.
Fervo Energy is the commercial story right now. Its December 2025 $462 million Series E — oversubscribed, with Google among new investors — is funding Cape Station in Utah, a 500 MW project investors are treating as the sector template. Fervo drilled a well at Project Blanford in under 11 days and filed for an IPO in January 2026. Institutional capital is paying attention: TechCrunch’s 2026 climate tech analysis names geothermal the top capital deployment candidate for 2026–2027, ahead of nuclear and offshore wind.
The geography constraint is real and shouldn’t be glossed over. Viable EGS resources are concentrated in the western United States; northern Virginia’s data centre cluster faces a mismatch with current EGS geology. But that coverage area is larger than most analyses acknowledge — and the technology keeps improving.
What role do gas turbines play in the near term?
Gas turbines are what is actually being deployed at AI data centres today. Not what anyone prefers — what works given the timelines. Meta has an agreement with Entergy to power a $10 billion data centre using gas with potential CCUS integration; Google is partnering on gas-plus-CCS projects at Broadwing Energy in Decatur, Illinois.
The key mechanism is behind-the-meter (BTM) deployment — turbines sited on or adjacent to the campus, bypassing grid interconnection queues that run 5–10 years. BTM gas is the only option for operators who need power in 1–3 years.
The carbon cost is real. Gas-fired generation produces approximately 400–500 kg CO2 per MWh, in direct conflict with net-zero commitments. Gas is a bridge, not a destination.
Carbon Capture and Storage (CCS) changes the carbon profile but adds cost and timeline. Only amine-based post-combustion capture is commercially available today, with an efficiency penalty of 15–25% of net output and an LCOE addition of $30–60/MWh — pushing gas+CCS to $80–120+/MWh. BCG positions gas with CCS as “the leading scalable option that can deliver” firm, low-carbon power in the near term.
One more thing worth knowing: equipment lead times have extended to 3–4 years with prices doubling in some markets. The window for fast gas deployment is narrowing.
What is the AI bubble risk and how does it affect the nuclear demand thesis?
The nuclear-AI power thesis — the same thesis explored across nuclear power’s AI renaissance — rests on a demand assumption: AI infrastructure investment will sustain current growth rates for the 10+ years it takes to bring new nuclear online. If that investment stalls or corrects, the demand thesis weakens. This is not a fringe concern.
The same CTOs being asked to commit to 20-year nuclear PPAs are managing capex budgets under board scrutiny of AI ROI. Amazon, Alphabet, Meta, and Microsoft spent nearly $300 billion on capex in 2025 — roughly 1.3% of US GDP, substantially debt-funded. GMO‘s analysis puts it starkly: AI investment already surpasses TMT spending in 1999–2000, but the cash flows to justify it haven’t appeared yet.
A 20-year nuclear PPA is a firm obligation regardless of demand. Over-commitment creates stranded assets; under-commitment means competing for constrained capacity. Geothermal and gas partially mitigate this — shorter deployment timelines mean committing closer to demand you can actually see rather than forecasting out to 2035.
Hyperscaler PPA activity is the best public signal of large-buyer demand confidence. Not certainty — but it’s the signal you’ve got.
How do the options compare — and what should infrastructure planners actually do?
Here’s the structured comparison across all five dimensions.
Existing nuclear — available now — ~$30–50/MWh — 24/7 yes — carbon-free — limited scalability (fixed plants)
New nuclear / SMR — 2030–2035+ — $50–130/MWh (target) — 24/7 yes — carbon-free — modular scalability
Solar + BESS — 3–5 years — $40–80+/MWh (with storage) — 24/7 no (with limits) — low carbon — high scalability
Wind — 4–6 years — $30–60/MWh — 24/7 no — low carbon — high scalability (onshore)
Geothermal (EGS) — 2–4 years — $50–80/MWh (est.) — 24/7 yes (firm) — carbon-free — growing
Gas turbines — 1–3 years (on-site) — $60–100+/MWh — 24/7 yes — high carbon — high scalability
Gas + CCS — 5–8 years — $80–120+/MWh — 24/7 yes — lower carbon — limited scalability
Figures are indicative ranges; actual costs vary significantly by project, region, and market conditions.
Geothermal sits in an unusual position — firm power, carbon-free, deployable in 2–4 years — that makes it the most attractive option for planners who cannot wait for new nuclear but cannot accept gas carbon exposure.
Here’s the practical framework:
- 1–3 years: On-site gas is the only realistic option. Plan for carbon offset commitments.
- 2–4 years, western US site: Geothermal via EGS (Fervo and peers) is worth evaluating first.
- Existing nuclear PPA available: Lowest-risk, lowest-cost firm power option — supply is finite, move quickly.
- 2030+ horizon: Engage the SMR pipeline now; queue positions at Kairos, X-energy, and TerraPower matter.
- Solar + BESS: Right for flexible-workload facilities or supplemental capacity, not a replacement for firm power at AI training scale.
For applying this comparison to cloud vendor evaluation, see applying the power source comparison to cloud vendor evaluation.
FAQ
What is the difference between baseload power and peak power for AI data centres?
Baseload is always-on generation at consistent output — the minimum a data centre requires continuously. AI training collapses the baseload/peak distinction: it runs at high intensity 24/7. Inference is more variable.
Can a data centre run entirely on solar power?
Yes, with real trade-offs. For a 5 GW facility, BESS at ten times operating power backup would cost over $5 billion before degradation. Solar-only is feasible for lower-uptime facilities; it’s not economical for 24/7 AI training campuses at 100 MW+ scale.
What is LCOE and how should I use it to compare power sources?
LCOE (levelised cost of energy) equals total lifetime plant cost divided by total lifetime energy output, in $/MWh. Its main limitation: it doesn’t capture the cost of intermittency. A lower-LCOE intermittent source may cost more all-in once backup generation or storage is added. For firm-power comparisons, look for all-in system LCOE.
Why are hyperscalers signing 20-year nuclear PPAs when plants take 10+ years to build?
Two situations are being conflated here. For existing plant restarts — the Three Mile Island model — the plant is already built and power can flow within 1–2 years of the PPA. For new SMR builds, the 20-year PPA provides revenue certainty that makes project financing viable. PPAs are enabling instruments, not just purchasing contracts.
What happened with NuScale and what does it mean for SMR timelines?
NuScale’s Carbon Free Power Project was cancelled in 2023 after projected build costs roughly doubled — from under $3.6 billion in 2020 to over $9 billion by 2023, for a smaller output. SMR cost targets should be treated as aspirational until a first commercial plant demonstrates them. NuScale did not invalidate SMRs — it validates scepticism about headline projections.
Is geothermal available everywhere?
Conventional hydrothermal geothermal requires natural hot water resources near the surface. EGS opens approximately 25% of global land area — but viable resources are still concentrated in the western United States. Northern Virginia’s data centre cluster faces a geographic mismatch with current EGS technology, though that constraint may ease over time.
What does “behind-the-meter” power mean for data centres?
Behind-the-meter (BTM) is generation sited on or adjacent to the campus, bypassing grid interconnection queues running 5–10 years in the US. Gas turbines are the primary BTM option today; SMRs are designed with BTM deployment in mind for future campuses.
How does CCS work with gas turbines?
Only amine-based post-combustion capture is commercially available today — chemical solvents remove CO2 from exhaust gases. The efficiency penalty is 15–25% of net output, and CCS adds $30–60/MWh to LCOE, hence gas+CCS at $80–120+/MWh versus $60–100+ unabated.
What is the AI bubble risk and why does it affect energy infrastructure decisions?
The AI bubble risk is the possibility that current investment levels are unsustainable relative to near-term returns. A 20-year nuclear PPA is a firm financial obligation regardless of demand outcomes — if the build-out slows, you pay for capacity you do not need.
How does wind compare to solar for AI data centre applications?
Wind achieves higher capacity factors (35–40%) than solar (20–25%). Offshore wind reaches 45–55% but faces permitting uncertainty in the US under the current administration. Both require BESS or backup generation for firm-power applications.
What is the role of PPAs in securing AI data centre power?
A PPA (typically 10–25 years) specifies a fixed or indexed electricity price, giving generators the revenue certainty to finance capital-intensive projects and buyers price certainty. Google’s 25-year PPA with NextEra enabled the Duane Arnold restart; Fervo Energy’s offtake agreement with Google works on the same principle. The trade-off: a long-term demand commitment that creates AI bubble risk exposure.
How long does it take to connect new power to the US grid?
The interconnection queue runs 5–10 years for most US projects. Over 2,700 GW of projects were waiting as of 2024 — more than double existing US installed capacity. The IEA estimates nearly 20% of planned data centre projects globally could face grid connection challenges. This affects all grid-tied sources, not just renewables.
