Why Google Chose a Reactor Design Never Built Before

Why hyperscalers are choosing unproven reactors over proven large-scale nuclear plants

On March 4, 2026, the U.S. Nuclear Regulatory Commission (NRC) issued a construction permit for a reactor to be built in Kemmerer, Wyoming. It’s TerraPower’s sodium fast reactor¹ — the company founded by Bill Gates. It was the first commercial approval for a non-light-water reactor in over 40 years, and a few weeks later, ground was actually broken.

At the very same moment, the number of large-scale nuclear plants under construction in the U.S. stood at 0. This despite the fact that the design work is done, there’s a clear licensing pathway, and Westinghouse’s AP1000 has already had two units completed.

Reader, the companies with the most urgent power needs — Google, Amazon, Microsoft, the hyperscalers — are choosing a reactor design that has never once been built, over proven large-scale nuclear plants. This isn’t because they love risk. It’s because in the West, “proven technology” no longer means “technology you can build quickly.”

---

There’s Not a Single Large Reactor Under Construction in the US Right Now

Vogtle Units 3 and 4 in Georgia created this situation. Final construction costs hit roughly $36 billion, more than double the original estimate, and the project took 15 years to complete. Georgia Power’s share alone was $11 billion, and that burden got passed on as a 25% rate hike for ratepayers.

Because of this single project, US utilities have been putting off decisions on new large reactors. There’s no insurance product that protects against cost overruns. Every dollar over budget lands directly on the balance sheet, and utilities have to justify rate increases to regulators. So everyone’s standing there with the same posture: “I’d rather not go first.” There’s interest, there’s evaluation, there are letters of intent being written. But nobody’s making the final investment decision.

BloombergNEF has tallied up the cooperation announcements between data center operators and nuclear companies at 51GW. That’s more than 50 reactors’ worth of capacity. But most of these deals lack binding power purchase agreements², there’s no financing announcement attached, and formal applications haven’t even been filed with the NRC. That’s why BNEF barely factors this 51GW into its forecasts.

It’s not that the government is sitting on its hands, either. If anything, US policy support over the past year has been unprecedented. A wave of nuclear-related executive orders came down all at once, and the NRC has committed to a hard deadline: a decision within 18 months of application receipt. The Department of Energy rolled out loan commitments to back construction of 10 large reactors, and even set a goal of getting 10 reactors into construction by 2030.

And still, nothing’s moving. What’s missing right now isn’t government support—it’s demand willing to buy. No utility is raising its hand to say “we’ll build it.” The regulatory framework is ready, the money is ready, but the first buyer hasn’t shown up. If you’ve ever tried to sell a new product, this scene will look familiar. Without an early adopter, even the best terms stay on paper.

Cooperation announcements between data center operators and nuclear companies stand at 51GW. New large reactor construction starts in the US stand at 0. That gap is what the Western nuclear market actually looks like right now.

Big reactors take so long not because of regulation, but because of a bottleneck in pressure-vessel supply chains

So why do large reactors take so long to build? People usually blame regulation, but in my view the more decisive factor is that there are only a handful of places on Earth that can actually manufacture the parts.

A gigawatt-class reactor pressure vessel is an enormous forged component. The equipment capable of producing it is concentrated almost entirely in Japan and Korea, and total global capacity runs at only two or three units a year. There’s none of it inside the United States. The moment you try to build several reactors at once, the bottleneck isn’t the blueprint — it’s this forging press.

That’s why shrinking the reactor size widens the pool of suppliers. Many more companies can manufacture the pressure vessel for a 300–500MW SMR³. On top of that, while components are being stamped out in a factory, site grading can proceed in parallel on the ground. Shorter construction time means lower interest during construction⁴. This matters more than it sounds. Tie up trillions of won for over a decade, and interest alone can wreck a project’s economics.

What sets China apart isn’t technology — it’s this repetition. Building six units back-to-back on a single site means the same crew rolls straight from Unit 1 to Unit 2 to Unit 3. 90–95% of components are made domestically, interest rates and labor costs are low, and the licensing pathway is simple. The result: construction at roughly one-fifth the cost of the West, finished in 5–6 years.

That rhythm was on display again just last week. On 4 August 2026, China approved 8 new reactors in a single batch. Six of them use the domestic Hualong One⁵ design — the first batch under the 15th Five-Year Plan (2026–2030). As of the end of 2025, China has 59 reactors operating and 35 under construction totaling 41.9GW, holding the world’s largest under-construction capacity for 19 straight years. BNEF projects that by 2030, China’s nuclear capacity could reach 102GW, overtaking the United States.

This gap turns into a political problem in export markets. For a country like Saudi Arabia weighing whether to adopt nuclear power, American technology runs around $15,000 per kW, while Chinese proposals come in around $2,700. Once the gap exceeds fivefold, it stops being a technology choice and becomes a fiscal decision. If the West wants to persuade buyers to “use our reactors,” security arguments alone won’t cut it — it has to close the cost gap.

Still, China’s domestic advantage doesn’t automatically translate abroad. The only country where China has actually built a reactor is, more or less, Pakistan. It was pushed out of the UK, and Argentina’s deal keeps unraveling with every change of government. Being cheap at home and being cheap in someone else’s country are two different problems. That’s why China developed something separate: the Linglong One, a 100MW SMR. Inside China, where power demand is exploding, a 100MW reactor has no real use. It was designed from the start as an entry-level model aimed at markets like Africa and South Asia — places where the grid is too small to handle a large reactor.

The battle isn’t reactor performance — it’s locking in repeat orders

Every price advanced reactor companies quote today is a first-of-a-kind cost⁶. Of course it’s expensive. A price that can compete with gas or renewables only shows up around the 10th or 12th unit.

But to get to unit 10, you first need orders for 10 units. In other words, what decides this market isn’t reactor performance — it’s whether you’ve locked in a string of repeat orders. There are currently more than 50 SMR and advanced reactor developers. BNEF believes only two or three of them will manage to secure both repeat orders and capital at the same time. The rest won’t disappear because their technology is bad — they’ll disappear because they can’t rack up enough repetitions.

The fact that capturing repeat orders and capital is what matters shows up in how hyperscalers are choosing, too. Google is betting simultaneously on Kairos Power (based on TRISO fuel⁷) and Commonwealth Fusion (fusion). Amazon has put both equity and a power purchase agreement into X-energy’s high-temperature gas reactor. Neither company picked a single winner — they bought an options portfolio. That’s because the point when they’ll actually need the power is the mid-2030s, and nobody today knows which companies will still be standing by then.

Utilities, by contrast, can’t operate this way. They have a least-cost-supply obligation, and if something fails, they have to explain it to regulators as a rate hike. That’s why everyone is watching Ontario Power Generation build first with GE Hitachi’s 300MW design. If that succeeds, TVA and Duke Energy are set up to follow.

India has turned this principle into an outright procurement condition. It won’t buy any SMR that lacks a commercial operating track record within its own borders. This holds even though India has set a 100GW target by 2047, amended its Atomic Energy Act to open the door to private participation, and set an aggressive cost target of under $2,000/kW. The message is clear: India won’t serve as a testing ground for unproven technology. BNEF’s forecast puts India at just 13GW by 2036 — still a long way from the target.

The one thing the United States can reliably scale up right now isn’t new construction — it’s revival. Restarting Palisades, the Crane Clean Energy Center (formerly Three Mile Island), and Duane Arnold — three units in total — brings back 2.2GW, while some 30-odd uprates⁸ at existing plants add another 2.3GW by 2036. The fact that switching existing plants back on is faster than building new ones captures exactly where things stand right now.

Oswarld’s Lens

When I’m building a go-to-market strategy, one of the pieces of advice I give most often is “use what’s proven.” New technology tends to have a failure-recovery cost that dwarfs its adoption cost. But the nuclear market has created conditions where that advice simply doesn’t apply.

Going through this material again, what struck me is that technological maturity isn’t determined by the technology alone. Maturity comes from the supply chain that builds it, the organizations that have actually handled it, and the number of times it’s been built. On paper, AP1000 is a fully proven design. But the US effectively has no team that has built one on budget and on schedule — its only attempt took 15 years. China, meanwhile, has already repeated the same kind of build more than 30 times. Same technology, yet in the US it’s a risky venture and in China it’s routine construction. That’s the reason why.

In last Thursday’s issue, writing about ASML, I noted that handing over the complete blueprints still wouldn’t let you build the machine. Nuclear power follows exactly the same logic. The designs are public and the licensing pathway is open, but the 20 years of accumulated field judgment that fills the gap between them no longer exists in the West. Whether it’s semiconductors or nuclear plants, what now separates industrial competitiveness isn’t the blueprint — it’s how many times you’ve recently built the thing with your own hands.

I see the exact same pattern in AI deployment. If you pick a “proven” solution but no one inside your organization has actually operated it, then within that organization it isn’t proven technology at all. Conversely, an unfamiliar tool that a team has run two or three times is already mature technology for them. What matters when making an adoption decision isn’t the vendor’s reference list — it’s your own organization’s repetition count.

And I can’t leave out Korea. The bottleneck the West is facing is a shortage of places that can manufacture the components — and Korea sits on the supply side of exactly that bottleneck. Doosan Enerbility, a South Korean heavy-industry conglomerate, has signed a deal with X-Energy to secure core SMR materials in advance, and announced roughly ₩800,000,000,000 (~$590M) in equipment investment for a dedicated SMR factory. More than 50 companies are competing to sell reactors, but the number of companies capable of supplying parts and materials to all 50 of them is far smaller. My view is that the suppliers have better odds of survival than the developers. To be clear, this isn’t a call on any individual stock — it’s a read on industry structure.

Looking at the draft against the source, everything checks out well — no Hangul remains, numbers match (0, 10th), structure is preserved, and the tone/meaning align closely with the Korean original.

Closing

To sum up, three things.

First, large reactors are no longer the safe choice in the West. After Vogtle, utilities decided not to go first, and as a result, there are zero large-reactor construction starts underway in the US today.

Second, hyperscalers wanting speed are spreading their bets across multiple unproven reactor designs instead. They’re not locking in one design — they’re keeping several candidates open at once.

Third, the contest isn’t decided by reactor performance but by repeat orders. A company that can’t bring the cost of its first unit down to the cost of its tenth unit disappears, no matter how good its technology is.

If your organization is evaluating a new technology, I’d suggest asking this question before you ever look at a vendor’s reference list: “How many times has our team actually run this?” Whether a technology is mature isn’t determined by the technology itself — it’s determined by how much your organization has actually handled it.

What about you, Reader? If you’ve ever picked the “proven option” only to have it take even longer, I’d love to hear in the comments whether the bottleneck was the technology, the people, or the supply chain.

---

💬 Tell me in the comments about a time the “proven choice” turned out slower — and what the bottleneck was. I’ll factor it into the next issue. 📨 If you know a colleague wrestling with tech procurement or infrastructure investment decisions, pass this along.

---

The English draft matches the Korean source accurately with no distortions, omissions, or number mismatches. No corrections needed.

References & Further Reading

Primary sources

• Tom Rowlands-Rees, Stephanie Diaz, Chris Gadomski, “Nuclear Power Market Outlook 1H 2026: Approaches Diverge”, BloombergNEF Switched On, 2026. ··· This is the backbone of today’s piece. The key points are the near-total lack of actual progress against the 51GW of announcements, and the forecast that “only two or three out of 50 projects will survive.”
• “NRC Approves TerraPower Construction Permit”, ANS Nuclear Newswire, 2026.3. ··· This lays out the backstory of the first non-light-water reactor approval in some 40 years, and the review timeline. The fact that it took 2 years from application to approval is the starting point of this piece.
• “What Was Learned from Building New Nuclear Reactors?”, POWER Magazine. ··· The source for Vogtle Units 3 and 4’s final cost of $36 billion and 15-year construction timeline. It also covers the up-to-25% rate hikes.
• “8 new nuclear reactors given full approval”, China Daily, 2026.8.4. ··· This covers the first approved batch under China’s 15th Five-Year Plan. Keep in mind this is state media, but the figures of 59 reactors in operation and 35 under construction come from the China Nuclear Energy Association’s tally.

Background

• Jane Nakano, “China’s Nuclear Energy Priorities Under Its 15th Five-Year Plan”, CSIS, 2026.5. ··· A report that dissects the policy structure behind China’s speed. Start here if you want to dig deeper into the “repetition structure” logic in today’s piece.
• “China is building half of the world’s new nuclear power”, Global Energy Monitor. ··· Confirms, with data, the geographic concentration of capacity currently under construction.
• “India will need Rs 25 lakh crore to achieve 100 GW nuclear target by 2047”, Down To Earth. ··· A summary of the TERI report, which tests the feasibility of India’s target from a capital-requirement standpoint. Worth reading alongside BNEF’s more conservative forecast of 13GW by 2036.

Related past issues worth reading

• Issue 240: The challenge ASML faces in feeding decades of tacit field knowledge into AI
• Issue 157: China’s review of AI export controls and reports on domestic-chip data centers

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Sodium Fast Reactor: A reactor that uses liquid sodium instead of water to carry away heat. Sodium conducts heat better than water, allowing for smaller reactor designs, but it’s tricky to handle since it reacts violently with both air and water. ↩

2. Offtake agreement: A contract signed before a plant is built that locks in who will buy its electricity, and at what price. Banks won’t lend money without one, so this document is effectively what decides whether construction goes ahead at all. ↩

3. SMR (Small Modular Reactor): A reactor that works on the same principles as conventional designs but is scaled down to the 300-500MW class. The goal is to manufacture components in a factory and assemble them on site. ↩

4. Interest during construction (IDC): The financing costs that accumulate while a plant is being built. Since these costs keep piling up during a period when not a single kWh of electricity is being sold, longer construction timelines rapidly erode a project’s economics. ↩

5. Hualong One: A 1,000MW-class pressurized water reactor designed in-house by China. It’s also the flagship model for Chinese nuclear exports. ↩

6. First-of-a-kind cost (FOAK): The cost of building a given design for the first time. Repeated builds bring costs down through learning effects, and the point costs reach through that process is called the nth-of-a-kind (NOAK) cost. ↩

7. TRISO fuel: Fuel made of uranium particles wrapped in multiple layers of ceramic shielding. The shielding itself acts as a containment barrier, keeping radioactive material from leaking out even at high temperatures. ↩

8. Uprating: Modifying a reactor’s surrounding equipment—rather than building a new plant—to extract more electricity from the same reactor. It’s far faster and cheaper than new construction. ↩