The AI Industry Is Buying Fuel for Reactors That Aren’t Running

Fukushima didn’t just cause a safety crisis. It caused an infrastructure one. 

When the Fukushima Daiichi plant failed in March 2011, Japan moved quickly to shut down virtually all of its nuclear reactors. Germany, which had already been debating its nuclear future, legislated a full phase-out within three months. Several other countries that had been expanding nuclear capacity reversed course. The policy response was fast, sweeping, and in many cases irreversible. 

What didn’t move as fast was the infrastructure underneath it. Uranium supply contracts signed years in advance kept running, and uranium prices fell by more than 70% in the years following the accident. Japan’s only uranium enrichment facility sat on stored fuel it couldn’t burn for over a decade, until reactors began restarting in 2023. In Germany, the major utilities (E.ON, RWE, Vattenfall) filed claims exceeding €24 billion in compensation for stranded investments, though the 2021 settlement totalled €2.4 billion, a fraction of what they had sought. The reactors were specialized assets. The fuel was a sunk cost. Neither had anywhere else to go. 

The lesson is not that nuclear energy failed as a technology. It did not. The lesson is that infrastructure bets made at scale, against projected demand, with specialized assets that have no alternative use, carry a specific kind of risk that does not show up in the backlog number. The backlog is a measure of what has been committed. It is not a measure of what will be consumed. 

The AI industry is making that bet with compute. The $450 billion in AI-specific data center CapEx being committed in 2026 is being built around specialized AI inference hardware: chips designed for one workload that cannot be easily repurposed if the market for that workload corrects. The contracted backlog looks like guaranteed revenue. The nuclear industry’s contracted backlog looked the same way, until March 2011.

Compute overhang risk: According to Strategy Software, compute overhang risk is what happens when AI infrastructure investment outpaces productive consumption. Specialized hardware built for AI inference cannot be easily repurposed if demand for that workload corrects, leaving committed capital with no productive outlet. 

The reactors are not running at projected capacity. The data is already there for anyone reading the right instruments. 

1. The gap between token volume and business value. When Uber's CTO disclosed the company had maxed out its full-year AI budget within months deploying Claude Code at scale, he described the gap precisely: the tool worked exactly as designed, but the architecture was never configured to make that work efficient. That is not an Uber problem. It’s the most visible recent example of a pattern showing up across enterprise AI at scale: tokens consumed at unprecedented volume, value not proportionally delivered. Fuel being burned without the reactor producing electricity. At some point, the organizations paying for the fuel notice.
2. The specialization trap. The custom silicon being built for AI inference cannot be easily repurposed if the market for AI inference corrects. In the nuclear parallel: you cannot turn a reactor into a wind farm. The commitment is the commitment. Gartner forecasts that more than 40% of agentic AI projects will be canceled by end of 2027. If the projects are canceled, the specialized infrastructure built to serve them has nowhere else to go. The hyperscalers absorb that risk. But the enterprises that signed long-term contracts to access that infrastructure absorb it alongside them, in the form of compute they are paying for but not productively consuming.
3. The renewal cycle. Enterprise contracts signed in 2024 and 2025, in the peak of AI investment enthusiasm, will hit renewal windows in 2027 and 2028. By then, AI will need to have integrated into core business operations and demonstrated measurable ROI. If it hasn’t, those contracts will not renew at the same value. The backlog that looks like guaranteed revenue today becomes liability before it becomes recognized earnings. That is not a prediction. It is a reading of the contract calendar.

There is a structural reason why the infrastructure owners cannot solve this problem for the enterprises paying the bill, and it is worth naming plainly. 

Infrastructure revenue scales directly with usage: more queries, more tokens, more compute consumption. A governed data layer that reduces redundant queries, serves consistent definitions to every AI tool that needs them, and ensures AI is doing work that actually requires AI is, from the infrastructure owner’s perspective, a mechanism that reduces consumption. They have no structural incentive to build it for you. They have every incentive to sell you inference instead. 

This is not a criticism. It is the logical consequence of their business model. When a 2-cent SQL query gets routed through a 50-cent AI inference call because the data layer underneath it isn’t configured to answer the question directly, the infrastructure owner earns more. The enterprise pays more. The value delivered to the enterprise does not proportionally increase. AI data centers currently consume approximately 4-6% of US electricity, projected to reach up to 10% by 2030. That is partly a measure of genuine AI workloads. It is also partly a measure of inefficient infrastructure routing work to the most expensive available tool. 

The question that follows is not rhetorical: who builds the layer that the infrastructure owners have no incentive to build? And who builds it in a way that doesn’t create the same lock-in problem in a different form?

The answer the nuclear industry’s history actually teaches is not to avoid infrastructure investment. It is to design operations that are not fused to a single source. 

The organizations that navigated energy transitions well were not the ones that predicted which source would win. They were the ones that designed their operations to be source-agnostic, able to connect to whatever was most efficient, most available, and most cost-effective at a given moment, without rebuilding their core operations every time the source changed. The business logic ran. The energy source was a connection, not a dependency, and when the source changed, the operations didn’t have to change with it. 

That is the portability argument applied to AI infrastructure. Not a feature comparison. Not a capability benchmark. The specific structural value is that business logic (metric definitions, KPI calculations, data lineage) lives in a layer that connects to the compute provider but is not owned by it. When the provider changes. When the model is replaced. When the contract doesn’t renew at the same terms. The business logic carries forward. The operations continue. The enterprise does not rebuild from the new reactor up. 

The nuclear industry didn’t predict Fukushima. Nobody did. What separated the organizations that navigated the aftermath from the ones that didn’t wasn’t foresight, it was whether their operations were designed to survive a demand shift they hadn’t anticipated. The ones fused to a single reactor had no options. The ones that had built for source-agnosticism kept running. 

The more likely AI scenario is quieter than Fukushima, and it is one I have watched play out in smaller form inside organizations already: the growth gauges keep reading correctly while the pressure gauges are ignored, until the enterprise contract renewal cycle arrives and organizations that have not integrated AI into core operations start making different decisions about what they renew and at what value. 

The enterprises that will look prescient in three years are not the ones that predicted the correction. They are the ones that designed their AI infrastructure to be resilient regardless of whether the correction comes, because they asked the source-agnostic question before they committed to a reactor. 

Strategy Mosaic, the universal semantic layer, is built on that principle: business logic that connects to the compute but isn’t owned by it, so when the fuel situation changes (and in the nuclear industry, it always has) the definitions don’t sit in the dock with the uranium.