AI's Power Problem Is Rewriting the Investment Case for Nuclear

AI capital spending added 1.1 percentage points to US GDP growth in the first half of 2025, according to JPMorgan. The next leg up of that cycle depends on something more pedestrian: whether the electrical grid can deliver the power required to run the data centres being built to absorb the spend. The current answer, in the regions where the largest projects are being sited, is no — at least not on the timeline the hyperscalers are working to.

That constraint is the shift. For two decades, nuclear's investment case has been an environmental argument rated against politically-determined energy policy. Over the last 18 months it has quietly become a private-sector procurement story. Meta has signed long-dated PPAs with Vistra, TerraPower and Oklo. Microsoft committed to restarting Three Mile Island Unit 1. Amazon Web Services acquired a colocated data centre campus at the Susquehanna nuclear plant directly from Talen Energy. The buyers of marginal nuclear capacity in the United States are no longer regulated utilities; they are corporate counterparties with multi-year compute commitments and the balance sheets to underwrite the capital cost.

Why Nuclear, Specifically

The constraint is duty cycle. Wind plants run at full output for an average of nine hours per day; solar for six. Nuclear runs at close to 24, and at the scale of generation a hyperscaler wants to colocate with a single GPU campus, no other low-carbon technology compares. Battery storage solves part of the intermittency problem but not the underlying generation gap. Natural gas is the substitute hyperscalers are actually using where nuclear is not available, which is bringing the carbon intensity of AI infrastructure into uncomfortable conversation with the same companies' net-zero commitments.

The policy backdrop is now lined up to support a build-out. The 33-country pledge at COP30 in Belém to triple global nuclear capacity by 2050 is the headline number. More relevant is the shift in financing posture: the European Industrial Alliance on Small Modular Reactors is structuring de-risking facilities for first-of-a-kind projects, and the UK has committed £2.5 billion of state capital behind the Rolls-Royce programme. Public capital absorbing the early-stage execution risk is what changes the asset class for private investors.

Where the Listed Equity Story Sits

The temptation is to play this directly through the SMR developers — Oklo, NuScale Power on the US side, Rolls-Royce on the UK side. We would caution against treating those names as a single thesis. NuScale's flagship Carbon Free Power Project was cancelled in 2023 on cost. The technology that will dominate the commercial SMR market in 2035 is not necessarily the one with the loudest pre-revenue narrative today.

The more defensible second-order exposure is the grid operators that will physically deliver the electricity nuclear capacity generates — and on which the AI buildout is also dependent regardless of whether the marginal megawatt is nuclear, gas or anything else. National Grid committed £35 billion of capex to UK transmission infrastructure in 2024 and is running ahead of plan on deployment. Its Q2 2026 update flagged £5 billion deployed in H1 of the financial year against an £11 billion full-year run rate. The UK government's AI Energy Council has earmarked a further £30 billion of investment into AI growth zones across the North-East of England.

The grid operators capture revenue from the AI infrastructure build whether or not the SMR programmes execute. The SMR developers only capture revenue if they do.

Our View

The AI-driven nuclear renaissance is real, but the listed-equity expression of it is not what the headlines suggest. We see the lower-volatility expression of the thesis — regulated grid capacity — as carrying meaningfully less execution risk than the SMR developer names while exposing investors to the same underlying demand curve. For the upside-skewed SMR exposure, Rolls-Royce is our preferred listed UK pure play; it sits on a manufacturing base and a state-backed export pipeline that the US-listed comparables cannot match.

Investors weighing the broader theme should size the SMR developer names against the multi-year cancellation track record in the sector, and treat the AI demand pull as a tailwind that operates on a different timeline to the technology rollout. The AI capacity problem is acute now. The first commercial UK SMR is targeted for the mid-2030s. Those two timelines will not meet without intermediate substitution.

Sources

Educational information only. This is not personal financial advice.