Sequoia just bet $1 billion that nuclear reactors belong inside AI data centers

Power, not compute, is now the binding constraint on AI scaling, and Sequoia Capital is betting $1 billion that the only capital pool fast enough to close the energy gap is venture equity. Valar Atomics' deal also includes a $200 million credit line from Erebor and other banks 1. Sequoia partner Shaun Maguire is joining Valar's board 1. The previous $2 billion valuation was set when Valar raised $450 million across equity and debt 2.

The bet rests on a structural mismatch that no GPU improvement can resolve.

Why utility timelines cannot serve AI's demand curve

Goldman Sachs Research projects US data center power demand climbing from 31 gigawatts in 2025 to 66 gigawatts by 2027, a rate that doubles consumption roughly every two years 3. The US Government Accountability Office has found that planning, licensing, and building a nuclear power plant takes 10 to 12 years, with four years attributable to the Nuclear Regulatory Commission's licensing process alone and more than six years for construction 4.

Five demand doublings fit inside a single permitting window. In the time it takes to approve and build one traditional nuclear plant, AI infrastructure demand would multiply more than thirtyfold at current growth rates. No utility monopoly or federal loan program is structured to plan capacity for a load that grows by orders of magnitude between the application filing and the plant coming online. This is the gap Sequoia is betting private capital can close. Utility planning cycles operate on decades. Venture capital operates on quarters.

Even the NRC acknowledges the timeline problem. Chairman Ho Nieh told the Deseret News editorial board that the commission is working to shorten federal licensing review times to under 18 months 5. Eighteen months for review, plus construction time, against a demand curve that doubles every two years, leaves almost no margin.

What Valar has actually demonstrated

The company's own development timeline shows real compression. Valar spent two years completing its first reactor core, called NOVA, but only seven months taking its next design, the Ward 250, to criticality 1. The Ward 250, a Generation IV high-temperature gas reactor cooled by pressurized helium, went critical on June 18, 2026, at Valar's site in Orangeville, Utah 5.

On July 1, 2026, Valar activated the reactor on stage and used it to power an Nvidia Blackwell system, marking what the company calls the first time a startup-built reactor has generated electricity 6. At the same event, Valar and Nvidia announced a partnership to develop a waterless 30-megawatt AI factory in Utah 7. Nvidia recently debuted its DSX AI factory design, which uses a direct liquid cooling system that eliminates power-hungry water chillers 5.

But the scale gap is enormous. During the July demonstration, the reactor ran at 37 percent of its rated capacity and produced roughly 100 kilowatts of thermal energy, enough to power a single desktop computer 6. The proposed Nvidia AI factory would require 30 megawatts. Goldman Sachs projects the total US data center fleet will need 66 gigawatts by 2027 3. Valar says this funding will allow it to begin manufacturing fleets of reactors 1. That transition from one test reactor to a production line is the entire risk of the investment.

Valar is not alone in this race. The Department of Energy says two other firms, Deployable Energy's Unity and Antares Nuclear's Mark-0, have also achieved criticality 6. Antares raised $470 million, and X-energy raised $1 billion through its IPO 1. Amazon, Google, Microsoft, Nvidia, and Oracle have all placed bets on nuclear technologies since 2024 6.

What breaks first if they are wrong

The question is not whether nuclear fission can power AI. The Ward 250 proved it can, at least at desktop scale. The question is whether factory-built reactors can scale from 100 kilowatts to the gigawatts that AI infrastructure demands, on a timeline measured in quarters rather than decades. Three things break if Sequoia and Valar are wrong about how fast that can happen:

- Grid reliability. At least 75 data center projects were delayed or cancelled in the first quarter of 2026, and 7 in 10 Americans say they do not want a data center in their community 6. If on-site nuclear does not materialize fast enough, the strain on regional power grids intensifies and public opposition hardens.

- AI economics. Goldman Sachs' projected 66 gigawatts of US data center demand by 2027 already strains the existing grid 3. Power scarcity becomes compute scarcity, and compute scarcity raises the cost of every AI-dependent business model.

- The venture model itself. Valar says that with each reactor built, its development cycle, or what the company calls its tick rate, will become smaller 1. That is the software-iteration thesis applied to nuclear hardware. Software compresses because marginal units cost nearly nothing to replicate. Nuclear reactors involve enriched uranium, precision manufacturing, safety certification, and physics. If the cost curve does not bend the way software cost curves bend, the $6 billion valuation rests on an analogy that has never been tested at industrial scale.

Sequoia's $1 billion is a bet that the venture capital playbook can be transplanted from software into the most heavily regulated, capital-intensive industry on Earth. Whether it works determines not only who powers the next decade of AI, but whether the AI boom itself remains energy-viable.