Nuclear and emerging sources

Wind, solar and batteries can supply most of a data center's energy cheaply, but as Lesson 5 showed, matching a constant 24-hour load with clean energy in every hour requires firm resources that produce whenever needed. Since 2024, data center operators have become the leading private buyers of firm clean power: restarting retired nuclear plants, contracting the output of existing ones, funding uprates, ordering small modular reactors, and investing in enhanced geothermal, long-duration storage and even fusion. This lesson examines these sources, their costs and timelines, and how data center owners contract for them.

Why firm clean power

A firm resource can produce at its rated output whenever called upon, apart from planned and forced outages. Nuclear plants run at capacity factors above 90%, geothermal similarly, and hydro with storage can follow demand. The value of firm resources rises steeply as the target for hourly clean matching rises, because meeting the last 10 to 20% of hours with wind, solar and short-duration batteries requires large overbuilding.

import numpy as np

rng = np.random.default_rng(7)
H = 8760
h = np.arange(H) % 24
day = np.arange(H) // 24
season = 1 + 0.3 * np.cos(2 * np.pi * (day - 172) / 365)      # more sun in summer
clouds = np.repeat(rng.uniform(0.3, 1.0, 365), 24)
solar = np.clip(np.sin(np.pi * (h - 6) / 12), 0, None) * season * clouds
solar /= solar.mean()                                          # 1 MW average per unit
w = np.zeros(H)
for t in range(1, H):                                          # slowly varying wind
    w[t] = 0.97 * w[t - 1] + rng.normal(0, 0.25)
wind = np.clip(0.35 + 0.25 * w, 0, 1)
wind /= wind.mean()
load = np.full(H, 100.0)

# cost of each resource per MWh of its output (illustrative)
cost = {"solar": 50, "wind": 60, "firm": 110}

def score(s, wd, f):
    gen = s * solar + wd * wind + f
    matched = np.minimum(load, gen).sum()
    spend = (cost["solar"] * s * solar.sum() + cost["wind"] * wd * wind.sum()
             + cost["firm"] * f * H)
    return matched / load.sum(), spend / load.sum()

best = {}
for s in range(0, 401, 20):
    for wd in range(0, 401, 20):
        for f in range(0, 101, 10):
            c, k = score(s, wd, f)
            for target in (0.8, 0.9, 0.95, 0.99):
                key = (target, f > 0)
                if c >= target and (key not in best or k < best[key][0]):
                    best[key] = (k, s, wd, f)

for target in (0.8, 0.9, 0.95, 0.99):
    for firm in (False, True):
        tag = "with firm" if firm else "no firm  "
        if (target, firm) not in best:
            print(f"CFE {target:.0%} {tag}: not reached with up to 400 MW each")
            continue
        k, s, wd, f = best[(target, firm)]
        print(f"CFE {target:.0%} {tag}: ${k:4.0f}/MWh  solar {s:3d}, wind {wd:3d}, firm {f:3d} MW")

CFE 80% no firm  : $  80/MWh  solar  40, wind 100, firm   0 MW
CFE 80% with firm: $  77/MWh  solar  20, wind  20, firm  50 MW
CFE 90% no firm  : $ 182/MWh  solar 100, wind 220, firm   0 MW
CFE 90% with firm: $  99/MWh  solar   0, wind   0, firm  90 MW
CFE 95% no firm  : not reached with up to 400 MW each
CFE 95% with firm: $ 110/MWh  solar   0, wind   0, firm 100 MW
CFE 99% no firm  : not reached with up to 400 MW each
CFE 99% with firm: $ 110/MWh  solar   0, wind   0, firm 100 MW

In this simplified model (one year of synthetic hourly solar and wind, firm clean power at $110/MWh, and no credit for surplus energy), reaching 80% hourly matching with wind and solar alone costs about $80/MWh of load. Reaching 90% without firm power more than doubles the cost, because large amounts of wind and solar must be built to cover calm, cloudy periods, and 95% cannot be reached at all within the range searched. With a firm resource, even 99% costs about $110/MWh, far less than the wind and solar cost at 90%. Adding storage, selling surplus energy and using real weather data changes the numbers, but the pattern — rapidly rising costs without firm resources as targets approach 100% — appears in most studies.

Existing nuclear plants

The US has about 94 operating reactors supplying nearly 20% of its electricity and nearly half of its carbon-free electricity. Most received 20-year license renewals to 60 years, and several have been approved to operate to 80 years. Between 2013 and 2022, about a dozen reactors closed, mainly because cheap natural gas and subsidized renewables made them unprofitable in wholesale markets. Data center demand has reversed their economics: in 2024 and 2025, large technology companies signed long-term contracts for existing plants' output, including Amazon's 1.92 GW agreement with Talen Energy for power from the Susquehanna plant in Pennsylvania (restructured in 2025 from a co-location arrangement to a grid-delivered contract) and Meta's 20-year agreement for the Clinton plant in Illinois.

Ways to obtain power from existing nuclear plants include:

• Contracts for existing output: long-term PPAs that keep plants running and support license extensions; critics note that they may simply move existing clean power from other customers to data centers, raising prices for others, unless paired with new capacity.
• Uprates: modifications to increase a reactor's output by a few percent to over 10%, often funded by data center contracts. Constellation and others plan more than 1 GW of uprates across their fleets.
• Restarts: returning recently retired plants to service. Holtec has been restarting the 800 MW Palisades plant in Michigan, the first US restart of a plant that had entered decommissioning, after NRC approval in 2025. Constellation is restarting Three Mile Island unit 1 (renamed the Crane Clean Energy Center, 835 MW) under a 20-year agreement with Microsoft, targeted for 2027 to 2028, and NextEra plans to restart the Duane Arnold plant in Iowa for Google.

New large reactors

Vogtle units 3 and 4 in Georgia, the first new US reactors in decades (AP1000 design, 1,117 MW each), entered service in 2023 and 2024, years late and at a total cost of over $30 billion. Their experience shows the cost of first-of-a-kind construction, but also that later units are cheaper: unit 4 was built faster than unit 3. In 2025, the federal government set a goal of 400 GW of nuclear capacity by 2050, ordered the NRC to decide on license applications within 18 months, and announced an agreement with Westinghouse for at least $80 billion of new AP1000 reactors.

Small modular reactors

Small modular reactors (SMRs) are reactors of up to about 300 MW per unit, designed to be built from factory-made modules with simpler, passive safety systems. Their promise is lower construction risk, shorter schedules, and costs that fall as more identical units are built. Advanced designs use different coolants (sodium, molten salt, helium gas) and fuels, offering higher temperatures and new capabilities. Microreactors of a few MW to tens of MW target remote sites and specialized uses.

In the US, the NRC has been modernizing its licensing framework, and DOE's Reactor Pilot Program aimed for at least three advanced test reactors to reach criticality by July 4, 2026; three achieved zero-power criticality by that date. Most commercial SMRs serving data centers are targeted for the early to mid 2030s. Technology companies typically sign agreements to purchase power, invest in developers or fund early project development, sharing first-of-a-kind risk.

SMRs face challenges: first units are expensive (NuScale's first project in Utah was canceled in 2023 when estimated prices rose to $89/MWh even with subsidies), fuel for many advanced designs (high-assay low-enriched uranium, HALEU) is not yet produced at commercial scale outside Russia, supply chains and skilled workforces must be built, and licensing new technologies takes time. Data center owners should treat SMRs as a 2030s resource and plan bridging supply accordingly.

Contracting for nuclear power

Data centers access nuclear power through several structures:

• Grid-delivered PPAs: the plant sells into the grid, and the buyer receives energy attribute certificates and a financial hedge (as in a VPPA) or physical delivery through its utility.
• Co-location: the data center connects directly to the plant, reducing transmission use. This raises the cost-allocation questions discussed in Lesson 7, and regulators have scrutinized it closely.
• Utility tariffs: some utilities offer clean transition or firm clean tariffs letting a data center pay a premium for new nuclear or geothermal supply procured by the utility.
• Equity and development funding: investing in reactor developers or early project costs in exchange for rights to future output.

Contract terms for new nuclear typically include long durations (20 years or more), price escalation, risk sharing for construction delays and cost overruns, and provisions for license changes. Because first units carry high risk, buyers increasingly commit to fleets of several units to earn lower prices on later ones.

Geothermal and hydro

Conventional geothermal power uses naturally permeable hot rock and is limited to specific regions (in the US mainly California and Nevada). Enhanced geothermal systems (EGS) apply horizontal drilling and hydraulic stimulation, techniques from the oil and gas industry, to create permeable reservoirs in hot dry rock, greatly expanding the potential resource. Closed-loop systems circulate fluid through sealed wellbores without fractures.

Fervo Energy's Cape Station in Utah, using EGS, is expected to deliver first power in 2026 and about 100 MW by early 2027; in September 2026 Google signed a 396 MW agreement for its expansion, with an option for about 600 MW more. Google had earlier contracted geothermal from Fervo in Nevada and through a clean transition tariff with NV Energy, and Meta has contracted EGS in New Mexico. Geothermal retains its tax credits longer than wind and solar under the 2025 law.

Hydro supplies about 6% of US electricity, mostly in the Northwest and Southeast, and pumped storage provides most existing grid storage. New large dams are unlikely, but upgrades of existing dams, adding generation to non-powered dams, and imports of Canadian hydro through new lines (such as the Champlain Hudson Power Express to New York City) add firm clean supply.

Longer-term options

• Long-duration storage: technologies such as iron-air batteries (about 100 hours), compressed air, pumped hydro and thermal storage can shift wind and solar energy across days, partly substituting for firm generation.
• Gas with carbon capture and storage (CCS): capturing 90% or more of CO₂ from gas plants and storing it underground. Several projects are planned near data center markets, especially where geology suits storage (Alberta, Texas, Louisiana, North Dakota), but costs and capture rates at scale are still being proven.
• Hydrogen: hydrogen produced with clean electricity can fuel turbines or fuel cells, but conversion losses (about 60 to 70% of the original energy) make it expensive for power generation except for long-duration backup.
• Fusion: private fusion companies have signed agreements with data center operators, including Helion's agreement to supply Microsoft from about 2028 and Commonwealth Fusion Systems' agreement with Google for its ARC plant in Virginia in the early 2030s. No fusion device has yet produced net electricity, so these are high-risk, long-term options rather than planning assumptions.

Putting it together

A data center owner planning a decade ahead might combine:

• Near term (now to 2028): grid supply with renewables PPAs, existing nuclear or hydro contracts where available, on-site gas or fuel cells for bridging, and batteries for flexibility.
• Medium term (2028 to 2032): restarts, uprates, geothermal and gas with CCS, plus 24/7 matching programs with utilities.
• Long term (2030s): SMRs and new large reactors, long-duration storage, and possibly fusion, with on-site gas transitioning to backup or being replaced.

Each choice must weigh cost, timing, risk, emissions, community acceptance and regulatory approval. The owners that secured power early with diverse, flexible portfolios have been able to grow; those that waited for the grid alone often could not.

Lesson 11 compares these long-term options with other paths on cost, risk and timing, and Lesson 12 combines them into a staged power strategy.