New Nuclear, Old Reactors

The reactors powering AI were mostly built in the last century. New nuclear is a 2030s energy story with a 2020s capital story, and states, more than start-ups, will decide who builds it.

Written by: George Patin | Head of Research, London Venture Capital Network

With the AI buildup running full steam ahead, one story we've been hearing about is the nuclear-for-AI lens. Some are calling it New Nuclear.

In practice, this is really a story in two parts. By October, US technology companies had contracted roughly 8.5 GW of output from existing nuclear plants, through long-term purchase agreements, restarts and uprates. Over the same period, the number of new reactors under construction to serve a data centre reached one: Kairos Power's 50 MW. Hermes 2, in Oak Ridge, Tennessee, due to start in 2030. Taken together, this comes out to a ratio of about 170 to 1.

The popular version of the thesis is that nuclear will power the AI race. That is half right. Except, it isn't the new SMRs powering the scaleup just yet, but older refurbished reactors. Existing reactors have become the most contested assets in American power, and AI contracts are now financing their uprates and even their restarts. Small modular reactors, microreactors and fusion are, for now, a 2030s story, but the investment and R&D in this space today is what will shape that not so very distant landscape.

New Nuclear and the Energy-Compute Loop

In April, another LVCN piece argued that the binding constraint on AI had moved from chips to power. That very much remains the case. The four largest hyperscalers have deployed roughly $720-745bn of capital expenditure in 2026. Buying electricity is evidently no longer sufficient; they are now turning to financing new construction, taking equity stakes in reactor startups or directly paying for utility upgrades.

Nuclear has emerged as one of the answers to this constraint. On the face of it, it's a rather neat solution: a highly concentrated energy source that can be purpose-built to cover all the needs of your installation without straining the existing grid. In the best case, you can even sell surplus energy back to the grid. In practice, though, a mismatch in reactor & data centre construction timelines, higher interest rates and construction backlogs are difficult to waive away.

One particularly interesting development is that AI is actually accelerating nuclear construction directly. The largest single award in the US Genesis Mission's first round, $60m over three years, went to using AI to deliver nuclear facilities faster and cheaper. Idaho National Laboratory is working with Nvidia on reactor deployment. Atomic Canyon has trained models on more than 50 million pages of the regulator's document archive. Licensing and safety-case work happens to be one of the largest bottlenecks in the industry, and it lends itself nicely to AI acceleration. However fast the applications themselves can be completed, though, we are yet to see the Nuclear Regulatory Commission granting permits any faster.

For now, that means we are seeing less new nuclear so much as old, decommissioned reactors being dragged back from beyond their concrete grave.

Behind that ration is, in many ways, a timing mismatch. New Nuclear is a new story. Many of the new, colossal data centres have either just been built in the past 5 years, or are still under construction. Even SMRs, touted as smaller and modular – it's in the name – take a good 5-6 years to set up. With the sole exception of China, conventional reactors take double that.

The response to this mismatch is a "bridge, then destination" model. In October, Blue Energy applied to connect a site in Texas that starts with gas and adds nuclear later. Its partner, Crusoe, has filed the first part of a construction-permit application for up to 1.5 GW of GE Vernova BWRX-300s at the Port of Victoria. Fermi America's Amarillo campus has the same shape: gas first, four large reactors later, and so far no binding tenant. The risk in every one of these is that the bridge becomes the destination. Once the gas plant and its grid connection are built and paid for, the nuclear phase has to be justified separately on its own price per MW terms – and those may not hold as favourably.

The Energy Scale-Up: What Is Actually in Play

Stepping back from AI, the scale of the energy buildout is clear. In 2025 the world added 692 GW of renewable capacity, 511 GW of that solar. It connected four new reactors with a combined 4.4 GW: two in China, one in India, one in Russia. That isn't entire a fair comparison, as a nuclear plant runs about 90% of the year compared to 15–20% with solar – the sun after all isn't always shining, especially in winter. Adjusted for that, though, solar alone still added something like 20 times the average output of new nuclear.

The picture with cost per megawatt is slightly better, but remains along similar lines. Lazard's 2026 levelised-cost study puts new onshore wind and utility-scale solar at roughly $37–99 per megawatt-hour without subsidy, combined-cycle gas at $51–129, and new nuclear at $175–255. This once again doesn't factor in uptime; a data centre after all runs day in, day out, sun or moon. Pricing that in narrows the gap but does not close it. A study for data-centre developers costed an off-grid site running 90% on solar and batteries at about $109/MWh, and one running 44% solar with gas back-up at about $93. Firmed solar still undercuts new nuclear, and it is far faster to construct.

Within nuclear the hierarchy is just as stark, and some of its helps to explain why the funding is flowing where it is. The cheapest new nuclear megawatt is by far uprates for old reactors. Google's October agreement with Constellation underwrites $4.3bn of upgrades that add 890 MW across 11 reactors by 2032, about $4,800 per kilowatt. The first new BWRX-300 in the G7, at Ontario's Darlington site, is budgeted at CAD 6.1bn for 300 MW, roughly four times that per kilowatt. Then there is what the military is willing to pay.

Small modular reactors were supposed to bend this curve. The idea is sound: build the same unit many times in a factory and let learning drive the cost down. I suspect this will eventually happen; one of the challenges with nuclear has always been how few reactors get built, limiting that learning effect. It depends on repeat orders for a single design, though. So far the interesting news story is from Europe. Rolls-Royce SMR has now been selected in the United Kingdom (Wylfa), the Czech Republic (Temelín, where the environmental assessment was filed this month) and Sweden (Ringhals, chosen over GE Vernova's design in June). The United States has a dozen designs competing for a handful of anchor customers, and almost no repeat orders. That is good for option value and bad for learning rates.

The Landscape: Reactors, Microreactors and Fusion

There's nuclear and then there's nuclear. While fission and fusion are of course totally different beasts, even within fission reactors that are a couple distinct classes.

Grid-scale SMRs

The first wave of Western SMRs is often being built by utilities and provinces, not just start-ups. Ontario's Darlington unit has been under construction since May 2025, with service around the end of 2030. The Tennessee Valley Authority received its construction permit for a BWRX-300 at Clinch River on 29 September, about fourteen months after applying. TerraPower has been building its Natrium plant in Wyoming since April, backed by about $2bn of federal cost-share. Kairos raised $643m this week for Hermes 2: $368m from private investors including Google's venture arm, DCVC and Samsung C&T, and $275m from the Department of Energy. NuScale, which holds the only fully certified US SMR design, still has no binding power contract for its 6 GW framework with TVA.

Microreactors

Microreactors are another interesting story. They're not quite SMRs: the main distinction is that they aren't modular or pre-fabbed, instead constructed in a similar fashion to classic nuclear, with a range of upgrades. Five test reactors reached criticality under federal programmes between June and August: Antares, Valar, Deployable Energy, Aalo and Oklo. Valar went from build to critical in seven months, outside a national laboratory. All five ran at zero or kilowatt-scale power. A data-centre unit is 10–300 MW, and the distance between the two is where earlier nuclear revivals died: fuel, qualification, factory scale-up, commercial licensing and first-of-a-kind cost.

Where, then, is the first commercial customer? Why, the US Army of course. Its Janus programme will pay up to $2.2bn, against milestones, to five vendors to build, own and operate reactors on bases, the first by September 2028. Radiant says its share is up to $750m for fifteen 1 MW units, roughly $50,000 per kilowatt of contract value. Military needs are needless to say rather different from commercial priorities, valuing resilience especially highly – and that is where some pull may come from. We will come back to this bit.

The capital: private up, public down

Since July, nine nuclear and fusion companies have raised about $4.5bn between them, roughly what the whole sector raised in the first seven months of the year. Public markets have gone the other way. Oklo and NuScale, the two most-watched listed developers, were each down about 55% for the year by mid-October, and both fell on the days the 10-year Treasury yield rose. That yield closed at 5.31% on 5 October, against 4.18% at the start of the year. Deep Fission's June IPO was cut from a planned ~$150m to $40m at the bottom of its range. Holders redeemed about $108m from newcleo's SPAC before it listed in September.

This can read this two ways. Private investors may be pricing what public markets cannot see yet, namely hyperscaler and military anchors that bring real cash flow. Or late-stage private marks have not caught up with a sector whose public comparables have halved under a higher cost of capital. Pre-revenue nuclear is close to the most rate-sensitive equity there is, because nearly all its value sits in cash flows from the 2030s. This is something we can only judge in retrospect, based on how future raises will go throughout 2026-2027.

Fusion

We don't know whether fusion is 30 years away, but it certainly won't land tomorrow. Most of the timelines announced appear to have been over-optimistic, and have since slipped. Commonwealth Fusion Systems, with about $4bn raised, now targets first plasma for SPARC in 2027, after 2025 and then 2026. Helion has softened its language for its Polaris prototype from producing net electricity to demonstrating electricity from fusion, and an executive has said its Orion plant may not reach full power until 2029–30. Zap Energy has added a fission business to earn revenue sooner. The money has kept coming regardless. Fusion raised a record $4.48bn in the year to July, about what fission start-ups raised in the first seven months of 2026.

That is of course a question of risk-reward: fusion is a bet on colossal energy generation capacity, far above our current technology. Fusion also isn't necessarily VC funding at all, at least not in the later funding rounds. CFS's July round came from pension funds, sovereign funds and industrial partners. Proxima's came with RWE, Google and German state support. In short, fusion is a fundamentally different bet on far longer timescales, and the picture there is as much to do with sovereignty and energy independence as it is with just energy prices or generation.

The investable near layer is the fusion supply chain. High-temperature superconducting magnets are one example: Tokamak Energy won the £70m magnet contract for the UK's STEP prototype. Medical isotopes are another: Astral Systems, in Bristol, raised £23m to bring isotopes to market by early 2027. Lasers, targets and pulsed power make up the rest. The same logic applies in fission, where the clearest gap is fuel. Russian supply of the high-assay fuel most advanced designs need ends in 2028. US enrichment is being rebuilt with $2.7bn of federal task orders to Centrus, General Matter and Orano, and Standard Nuclear raised $140m to make the fuel pellets. This is the one part of the nuclear stack where capital is clearly behind the need.

Who Is Buying?

As we've noted above, different layers of the nuclear stack have fundamentally different customers. There are three key ones.

Hyperscalers are the first buyer, and they behave like project financiers. October alone brought three deals: Google's 890 MW of uprates with Constellation, Amazon's 690 MW from Calvert Cliffs in Maryland, and Oracle's 125–250 MW subscription to Wisconsin's Point Beach. Against existing plants these are firm contracts. Against new reactors they are mostly frameworks and options, with roughly 6 to 8 developers behind most of them.

The state is the second, and for anything genuinely new it is the first. Nearly every Western first-of-a-kind reactor is anchored by a utility, a province, a government loan or the armed forces. The sums have grown pretty large.

The third buyer is the most interesting over the long run: states choosing technologies. What a state wants out of its energy stack goes way beyond price per megawatt: density, land, grid strength, industrial base, security, redundancy and probably a good half-a-dozen other factors shape what states decide to finance. So it is with nuclear. This is one of those cases where the landscape is pretty different by country.

The pattern isn't just "nuclear versus renewables". China builds the most of both. The United States leans on gas and sweats the fleet it already has, while funding a venture-backed field of new designs for the 2030s. Europe has chosen a utility-led route around one standard design. Russia treats reactors as an export product and bundles state finance, which, in the absence of any World Bank lending on new reactors, likely makes it the default vendor for much of the Global South this decade.

Britain sits at an odd angle. It has some of the most advanced pieces in Europe: the first Rolls-Royce SMR site, Sizewell C, a national fusion prototype and its magnet supplier, and a credible fusion start-up layer from Tokamak Energy to First Light and Astral. It also has industrial power prices that CNBC put at around four times American levels in May, a paused Stargate UK, and a Scottish freeze on new hyperscale approvals. As things stand, Britain is likely to host the supply chain and continue leading on the R&D without anywhere near the compute scaleup race the US is currently involved in. On that note, the overwhelming majority of top-shelf data centres are in the US or China, with some scattered across Europe. As far as the energy-compute thesis goes, this is likely where new nuclear will land.

Near Horizon

There is clearly a lot of resurgent interest in nuclear energy. Progress in nuclear hasn't just stood still: reactor designs have been evolving at a much faster pace than they are being constructed at. As far as the near future is concerned, there are a couple things to watch.

The first is whether the pace & scale of funding matches public valuations. With the 10-year above 5%, a sector valued almost entirely on 2030s cash flows is quite exposed. The first flat or down late-stage round, or the first pulled listing, might give the market something of a shakeup. The second is whether frameworks translate contracts. NuScale reportedly hopes to sign with TVA this year. The Department of Energy is due to name the partners for its AP1000 loans. In that respect committed funding and/or closed contracts are more interesting than optionality. The third is first-of-a-kind delivery. Darlington, Kemmerer and Hermes 2 all aim for power around 2030, and whether they arrive on time and near budget will set the cost of capital for everything that follows. This is also something worth watching more broadly: as much as price comparisons are a very real factor, there are some things we simply don't know about the new designs. Until they are deployed more extensively, we won't know what the picture is about their capabilities; prices may well fall within competitive levels in the first wave of new deployments.

For investors, the useful question is the same one this series asked of autonomous labs: funding notwithstanding, what can reach the market before the big outcomes arrive? On that test, fission & fusion are in fact reasonably robust. On their current timelines, though, they are more project finance and state procurement that happen to have VC cap tables. Where the venture shaped money is is often in the supply chains, from components & manufacturing to licensing and engineering software –– plus the designs with the potential to become industry-standard.

The existing nuclear fleet has turned out to be one of the best-positioned assets of the AI build-out, and the case for firm, dense, low-emission power in places without much sun or land is not going away. The AI race probably won't be led by nuclear just yet; solar and other renewables are outmatching any new SMR fleets in that department. For the time being, much of the pull towards nuclear is coming from states and other forms of public procurement, with different rationales. As reactors advance and improve, though, we may well see wider deployment and a broader ecosystem serving nuclear that will be worth revisiting.