From Enriched Uranium to Enriched Compute: The Paducah Conversion and the Quiet Centralization of Intelligence

The federal government wants to turn a Kentucky uranium enrichment plant into an AI data center. The report, first surfaced by Crypto Briefing, offers no operator, no budget, no timeline. Just the transformation. Paducah was once the quiet heart of the American nuclear complex, a gaseous diffusion facility that for six decades pulled thousands of megawatts from the Tennessee Valley grid to enrich uranium for weapons and reactors. The same electrical architecture that fed the Cold War now sits in line for a second career as an aggregate of machine intelligence. I spent three months in 2017 auditing the whitepapers of forty-two failed ICOs. I found that 85 percent of them lacked any sustainable value proposition beyond speculation, and I learned a lesson that has never left me: what a summary omits is usually the substance. This announcement is a summary. The real document — the liability ledger, the environmental baseline, the tenant agreement, the audit framework — has not been written. And that is where the story actually lives. Before we can evaluate this conversion, we need to understand what Paducah actually is. The Paducah Gaseous Diffusion Plant, located in McCracken County on the banks of the Ohio River in far western Kentucky, began operations in 1952 as part of the postwar expansion of American nuclear infrastructure. It was one of three gaseous diffusion enrichment sites — Oak Ridge, Paducah, Portsmouth — built to produce enriched uranium for the weapons stockpile and, later, for commercial nuclear fuel. Gaseous diffusion is an extraordinarily energy-intensive process: uranium hexafluoride gas is forced through thousands of porous barriers, and separating the desired fraction of U-235 requires an immense cascade of pumps and compressors. At its peak, Paducah consumed on the order of three thousand megawatts of electricity, enough to power a city of several million people. The site imported that power through massive substations and switchyards that remain in place. That history is also a history of waste. For decades, operations released trichloroethylene, perfluorinated compounds, and technetium-99 into the surrounding environment. Groundwater plumes beneath the site remain a live liability. The Department of Energy's Office of Environmental Management has been working on remediation since the Cold War ended, and the effort is measured in billions of dollars. The plant was listed on the National Priorities List — the Superfund list — in 1994. Paducah ceased enrichment in 2013, and its buildings and equipment have been in various states of decommissioning and surveillance ever since. Meanwhile, the context on the other side of the equation: artificial intelligence has outgrown the grid. Data center developers in the United States now face interconnection queues measured in years. Power procurement has replaced land as the binding constraint for training clusters. Industry projections for the largest AI campuses run toward one gigawatt or more, and the thermal demands of GPU clusters have made cooling a site-selection criterion as important as fiber access. The hyperscalers have responded by signing nuclear power agreements, buying into small modular reactor startups, and acquiring land adjacent to existing industrial power infrastructure. In that sense, a federal proposal to convert a decommissioned uranium plant into an AI data center is less surprising than it first appears. It is the logical conclusion of the electricity bottleneck. But conversion is a word with a long memory. Let me start with what the site genuinely offers, because dismissing this as political theater would be a mistake. Gaseous diffusion plants were built to move enormous quantities of electric power and heat. That is, almost word for word, the requirement list of a modern hyperscale AI data center. Power is the first inheritance. Paducah's substations and switchyards were engineered to receive and distribute electricity at a scale that few industrial parcels in the country can match. The site contains high-capacity feeders designed for continuous, heavy industrial load — the same load profile as a GPU training cluster. A typical hyperscale AI campus today draws one hundred to five hundred megawatts, with frontrunners pushing into the top of that range. The utility infrastructure required for such facilities — high-voltage transmission access, redundant substations, dedicated feeders — is exactly the electrical skeleton Paducah already possesses. A greenfield developer would wait years for interconnection. Paducah may already be interconnected at levels that would otherwise take a decade to procure. The practical value of that existing connection is not merely financial; it is temporal. In the race to train frontier models, a facility that can draw power this decade instead of next is a strategic asset before a single server is installed. The second inheritance is cooling. The site sits adjacent to the Ohio River and contains an extensive industrial water withdrawal and thermal rejection system built to dissipate the enormous heat load of the diffusion cascades. For GPU clusters, where cooling is often the second-largest operational cost after electricity, this existing thermal infrastructure is not a small convenience. Modern AI data centers increasingly use direct liquid cooling and immersion cooling. The plumbing, the pumping capacity, and the heat rejection systems at Paducah can all be repurposed with careful retrofits. The question is whether the original cooling water systems, designed for a different chemical process, can be rehabilitated to the purity and corrosion standards that precision electronics demand. That is an engineering hurdle, not an impossibility. The third inheritance is physical security. This is the element that press coverage consistently misses. Paducah was designed as part of a weapons-related supply chain. It has hardened perimeter security, layered access control, and a site layout engineered around containing both nuclear material and information about it. For an AI facility that may handle sensitive defense, intelligence, or research workloads, this is an extraordinary starting point. A federal AI data center with national security applications would need exactly this kind of high-assurance physical isolation. The high wall with a machine inside is not simply a metaphor; it is an architectural lineage. The same structures that kept enriched uranium from leaving the site could keep model weights from leaking out of it. Finally, there is land and structure. The site spans thousands of acres, with multiple large process buildings designed for heavy equipment, high ceilings, and dense electrical distribution. My own fieldwork gives me a reference point. In early 2021, I visited two legacy industrial sites in the American Midwest that had been converted into bitcoin mining facilities. Both were old manufacturing plants with oversized electrical service, and in both cases the conversion was materially faster and cheaper than a greenfield build. But those were twenty-megawatt operations. Paducah is a different order of magnitude — the difference between a neighborhood generator and a regional power station. The reuse logic is sound, but the scale carries consequences that those small mining operations never had to face. There is a reason uranium enrichment plants are not sold on the open market, and the reason is not sentimental. Paducah is not empty land. It is a contaminated industrial site that the federal government is obligated to remediate under Superfund law and under decades of consent orders with the Commonwealth of Kentucky. The contamination picture is well documented in public records. Groundwater beneath the western portion of the site contains trichloroethylene, a degreasing solvent that is a known carcinogen, in concentrations that have required active pump-and-treat remediation. There is also technetium-99, a mobile radioactive isotope with a half-life of more than two hundred thousand years, detected in groundwater migrating toward the Ohio River. The DOE has constructed barrier walls and extraction wells, but the efficacy of that effort is uncertain and contested. This is not a minor footnote. It is the economic and ethical center of gravity of the entire project. So here is the question that no announcement will answer: if this site becomes an AI data center, does the cleanup get faster or slower? There are two readings. The first is optimistic and genuinely plausible: the commercial value of the data center provides a revenue stream that partially funds remediation, and the presence of active operations forces a faster, more aggressive cleanup program. This is a form of brownfield redevelopment — the reuse of contaminated land in a way that introduces cash flow, responsible stewardship, and a constituency for environmental quality. It has worked elsewhere, though rarely at this scale. The second reading is darker. A commercial lease on contaminated federal land can become a mechanism for deferring cleanup. The DOE's long-standing struggle is that environmental remediation budgets are politically fragile. A data center lease, by contrast, produces revenue and jobs — two things that elected officials love. If the cleanup costs are refinanced through a lease structure that gives the operator control of the site, the remediation timeline may quietly stretch from decades to generations. The toxic ledger does not disappear when a server rack is installed. It is merely moved below a very expensive floor. I have spent most of my career looking at documents engineered to reassure. The 85 percent of ICOs that lacked sustainable value propositions were almost always the ones with the most polished executive summaries. There is a parallel here. Every announcement about an AI data center on a former uranium plant is an exercise in narrative compression. The full text — the environmental baseline, the cost-benefit analysis, the liability transfer structure — is what determines whether this project builds something of value or simply repackages decades of deferred responsibility. Don't confuse liquidity with loyalty, and do not confuse a lease payment with a cleanup. The environmental justice dimension compounds this. The Purchase area of far western Kentucky is a region of modest economic diversification, and the local community has lived with the risks and the stigma of the nuclear plant for seventy years. If the AI data center becomes a vehicle for good jobs and genuine remediation, the opportunity is real. If it becomes a way to keep the contamination in place while training military algorithms, the sacrifice zone narrative becomes literal. Communities that hosted the nuclear economy made real sacrifices; they deserve better than a rhetorical conversion that asks them to sacrifice again. Let me step back now from the weeds of the site and look at what this project signals at the level of the state. The federal government is not primarily a builder of data centers. Historically, the civilian federal footprint in compute has run through research institutions — national laboratories, university clusters funded by the National Science Foundation, and more recently cloud contracts with NASA, the National Institutes of Health, and the Department of Defense. The hyperscale sector has been almost entirely private. Microsoft, Amazon, Google, and Meta build and operate their own infrastructure. The government rents from them. This Paducah proposal, if advanced, breaks that pattern. It suggests that Washington is willing to become an owner-operator of AI compute infrastructure, using federal land and federal energy assets as the basis. That would reposition the state from a renter of the compute market to a landlord within it. This is not a neutral operational detail. It is a shift in the political economy of artificial intelligence, and it deserves far more examination than a brief news alert. The strategic logic is easy to trace. The United States has decided that AI capability is a matter of national security. The CHIPS and Science Act of 2022 committed tens of billions of dollars to domestic semiconductor manufacturing. Export controls on advanced AI chips to China, introduced in October 2022 and tightened repeatedly since, treat compute as a strategic commodity. Executive Order 14110 on safe, secure, and trustworthy AI, issued in October 2023, imposed reporting requirements on the training of large models. All of these measures share a common premise: computation, and AI compute in particular, is a strategic resource rivaling oil or uranium. In that frame, the conversion of a uranium enrichment plant into an AI data center is less an anomaly than the next logical step. The state that once controlled the nuclear fuel cycle is reaching for control of the compute cycle. There is also a geopolitical mirror. China's national East-West Computing project, launched with official priming in 2022, coordinates data centers across the country in an explicitly centralizing design. The European Union's AI Factories initiative, funded through EuroHPC, aims to provide sovereign AI compute to member states. Every major power is treating compute as public infrastructure. The United States is now signaling that it will participate in the same game. Whether it does so more transparently and accountably than its peers is the open question — and it is not an academic one. The difference between a sovereign compute reserve and a national surveillance infrastructure is entirely a matter of governance design, and governance design does not happen by accident. Here is where I have to raise my own flag. I have spent more than two decades watching this industry, and I have built a career on the belief that decentralization is an ethical imperative, not a technical preference. A federal government that owns a strategic AI data center is the precise inverse of my founding values. But the fact that something contradicts my values does not automatically make it wrong. The relevant question is not whether a sovereign AI facility is centralizing — it obviously is. The relevant question is whether that centralization is accountable, bounded, and transparent enough that its risks can be survived. That is a governance question, not an ideology question. In 2024, I worked with five traditional finance academics to draft a values-based investment framework for institutional allocators. One of our recurring findings was that institutions feared blockchain not for its technology but for its cultural opacity — they could not see who was accountable for what. The same principle applies a thousand times over to government AI infrastructure. Accountability is not a design feature that can be bolted on later; it is structural. If Paducah becomes a national AI facility, the governance of that facility — who decides what models run on it, who audits training data, who can verify compute logs — has to be established before the first rack is loaded, not after. There is one more layer I need to add. In 2026, I co-led a research effort with a group of AI scientists to design Ethical Oracles: smart contracts that enforce human-centric constraints in autonomous transactions. The central insight of that work was that transparency alone is insufficient; power must be structured so that it can be audited under adversarial conditions. A government AI data center built on a former nuclear site is the most adversarial condition I can conceive of. Without verifiable audit infrastructure — a public compute log, independent oversight, cryptographic attestation — the facility's output will be indistinguishable from a black box, and the public will rightly treat it as one. This is the section where I need to be honest about the cognitive dissonance within my own community. Much of Web3's value proposition runs directly against this project. Decentralized physical infrastructure networks — DePIN — exist precisely because centralized data centers have become a bottleneck. Distributed compute platforms and storage markets have spent years arguing that the future of computation is crowd-sourced, distributed, and censorship-resistant. A federal government converting a Cold War nuclear plant into a megawatt-scale AI fortress is the strongest possible counter-thesis. It says, in effect: the most consequential AI compute will not be distributed. It will be owned by states. The irony is layered. The same communities that celebrate decentralization as a moral absolute are often the ones that most loudly celebrate American technological supremacy. You cannot say in one breath that compute should be owned by no one, and in the next, that the United States must win the AI race against China with all available state assets. Those positions are in tension. The Paducah project forces that tension into daylight, and the blockchain community has not yet developed a mature answer to it. Let me also address a practical point that the crypto world should not miss. This project, if it materializes, will be a direct competitor for the energy and grid capacity that decentralized compute networks would have wanted. Every megawatt consumed by a sovereign AI data center in Kentucky is a megawatt that is not available to a GPU miner in Ohio or a distributed training node in Texas. The compute market is constrained by physical infrastructure, not by demand alone. When the state enters the market as a dominant consumer, it changes the price curve and the availability curve for everyone else. Distributed networks that assumed energy would remain abundant and cheap may find themselves priced out by a landlord that does not need to earn a market return. But there is an opening here, and it is one that the Web3 community is poorly positioned to exploit because it spends too much time talking to itself. The genuine need that Paducah exposes is verifiable computation. If the government builds a sovereign AI facility, the public has a legitimate interest in knowing what is being computed there. Zero-knowledge proofs — a technology I studied deeply during the quiet months after the 2022 market collapse — could be part of the solution. A facility that publishes cryptographic proofs that its compute was performed correctly, without exposing the underlying data, is a facility that can be audited without compromising security. That capability is something the blockchain community uniquely understands. It will require humility to offer it; the decentralized idealists offering audit tools to a federal data center is an inversion of the usual posture. But the alternative is to be spectators as the most important compute infrastructure of our era is built under zero public accountability. I am not arguing that the Department of Energy will run its AI workloads on decentralized networks. That would be nonsensical. I am arguing that the derivative technologies of decentralization — verifiable audit trails, cryptographic transparency, key custody standards, tamper-evident logging — are the only viable bridge between state-controlled compute and democratic accountability. The values I embedded in the Ethical Oracles project, the community work I did during the DeFi summer, and every long-form piece I have written about trustless social contracts translate directly to this fight, even when the infrastructure itself wears a centralized face. Let us move from philosophy to capital, because this is where most projects either flower or rot. The federal government's role in commercial data centers has historically been that of landlord or tenant, not developer. The typical pattern is a public-private partnership: the Department of Energy or the General Services Administration contributes land, power infrastructure, and expedited permitting; a private operator invests capital and runs the facility; the government promises anchor workloads in exchange for priority access. This is the structure most experienced insiders expect for Paducah, and the structure has real appeal. It allows the federal government to avoid the enormous capital expenditure of a hyperscale build — commonly several billion dollars for a campus of this size — while still securing strategic compute capacity. The economics hinge on one thing: the anchor tenant. A data center's financial model is a function of committed capacity, not potential demand. If a major cloud provider or a defense contractor signs an offtake agreement for a quarter of the facility, the financing closes and the construction timeline becomes real. If nobody commits, the project becomes a policy artifact — a press release with a power bill. My ICO audit instinct applies here without modification. The difference between a genuine infrastructure project and a speculative one is the existence of a tenant who has committed real money. A memorandum of understanding is not a revenue contract. Don't confuse liquidity with loyalty. There is also the question of the power arrangement. Paducah sits in the territory of the Tennessee Valley Authority, a federal power corporation that has long been among the largest electricity producers in the country. TVA is already engaged with nuclear energy and has explored small modular reactor deployments at sites across its footprint. It does not require extraordinary imagination to see TVA offering a bespoke industrial power contract for the Paducah site, possibly tied to an SMR development that provides baseload, carbon-free electricity specifically for AI compute. Such an arrangement would be the first of its kind at scale: a federal power utility, a federal nuclear heritage site, and a private operator unified under one financial structure. The listed companies that would benefit from this pattern are not difficult to enumerate. Engineering and construction contractors with nuclear heritage — firms like Bechtel, Fluor, and BWXT. Thermal management and cooling specialists like Vertiv and Modine. Grid equipment makers like GE Vernova. Reactor developers like NuScale and Kairos Power, if the SMR path is taken. And the Department of Energy's Loan Programs Office is actively seeking projects to finance under its clean energy mandate; a Paducah conversion, with its nuclear and grid components, could be eligible for loan guarantees that make the capital stack almost too cheap to refuse. These are not speculative stock tips. They are the direct routes of capital flow for any serious federal infrastructure conversion. But the timeline caveat is severe. Federal infrastructure megaprojects in the United States run over schedule at rates that would be scandalous in the private sector. The uranium enrichment complex itself is a legendary example of cost escalation. A project that commits today to an AI data center at Paducah will not deliver meaningful compute before the early 2030s, if ever, and the world of AI infrastructure in that decade will be materially different. Chip efficiency, cooling technology, and possibly the compute model itself will have shifted. The 2026-era assumptions embedded in this announcement are the greatest unacknowledged risk in the entire proposition. I should also flag a cynical reading, because years of reading project documents have taught me that political announcements are often not intended for the industry at all. An AI data center on an old Kentucky uranium plant is, above all, a powerful message to a region and a nation: the federal government is doing something about AI, something about jobs, something about the energy transition. Whether the project reaches first concrete is almost secondary to the message it sends. That is not necessarily fatal — policy signaling has genuine value — but investors and community members should discount promotional infrastructure accordingly. The final layer of the technical story is symbolic, and I do not think we can understand this project without it. Artificial intelligence and nuclear power are converging across the American landscape. Microsoft's agreement to restart the Three Mile Island reactor for AI workloads was a shock to the industry precisely because it normalized the idea that data centers could purchase their own nuclear plants. Google has contracted with Kairos Power for small modular reactors. Amazon has invested in X-energy and signed agreements for nuclear-powered data centers. Utilities across the country are exploring whether on-site nuclear generation can anchor grid-scale compute. The logic is straightforward: AI needs zero-carbon, baseload, high-density power, and nuclear is the only source that provides all three without the intermittency of solar and wind. Paducah would take this convergence to its natural conclusion: a former nuclear fuel cycle site where the nuclear infrastructure itself becomes the power backbone for machine intelligence. The symbolism is thick. The Department of Energy would effectively be saying that the atomic nucleus, which once enriched fissile material for the Cold War, now enriches machine intelligence for the economic contests of the next generation. Whether the site hosts SMRs or draws from TVA's broader fleet, the project cannot escape its nuclear genealogy. There is a danger here, and it is the same danger I saw in the ICO boom: the conflation of historical narrative with operational reality. Uranium heritage does not equal electricity deliverability. The site's existing power infrastructure is at the end of its useful life. The switchyards may be serviceable, but the original transformers and feed gear are approaching or exceeding fifty years of age. Repowering Paducah will require new equipment, new grid agreements, and new voltage conversion infrastructure. The nuclear past creates the illusion of a ready-made power connection. In most cases, that illusion is false. The electrical architecture of a gaseous diffusion plant was optimized for steady, enormous, round-the-clock loads. An AI training facility has a similar load shape, which helps, but the equipment cannot be assumed to be safe or efficient simply because it is historic. And I have not even mentioned the human dimension of energy transition. The workers who staffed Paducah at its peak numbered in the thousands. They held unionized, highly skilled jobs with a defined industrial culture. Data center operations, by contrast, employ far fewer people per megawatt, and the skill profile is entirely different: network engineering, thermal management, security, and increasingly AI operations. The premise that an AI data center replaces the lost jobs of the nuclear era is economically questionable. It converts a site that once employed a community into a site that employs a small technical elite. That is a legitimate improvement over an empty and contaminated site, but it is not the revival that the rhetoric implies. Communities stripped of good jobs are not healed by the arrival of two hundred well-paid technicians. They still need education, healthcare, and the dignity of broadly shared work. Data centers do not provide that broad-based dignity, and I would be failing in honesty if I pretended otherwise. Here is the claim that I believe most analysts will miss, so let me state it plainly: the Paducah conversion is not primarily an AI project. It is an environmental remediation financing vehicle wearing an AI costume. The Department of Energy has a budget problem that transcends any particular administration. It is obligated, by law and by consent decree, to clean up sites it no longer uses. Paducah's contamination will cost billions more no matter what happens. An AI data center narrative converts an endless liability stream into a hopeful asset story, unlocking private capital, local political champions, and budget line items that a pure remediation program could never command. The AI framing is the fundraising pitch. The cleanup is the underlying debt. This reading has consequences. If the environment is the real purpose, then the project's evaluation metrics are inverted. The success of an AI data center is measured in compute online and jobs created. The success of a remediation program is measured in cubic feet of contaminated soil removed and plume migration stopped. These metrics can conflict. The pressure to get the data center operational — ribbon cuttings, tenant announcements, political victories — will inevitably compete with the slower, less glamorous work of making the site safe. When a cost overrun occurs on the cleanup side, who eats it? When the operator demands larger portions of the site, who is protecting the public from vapor intrusion risks? The second contrarian point is about obsolescence. Federal infrastructure projects move on a timescale of a decade or more. Environmental review, remediation, procurement, construction — optimistic timelines say 2032, realistic timelines say 2035 or beyond. The compute technology of that era will look nothing like the compute technology of 2026. GPU architectures, power densities, and cooling designs are all in rapid flux. A facility that begins construction with today's specifications may be architecturally obsolete before it can accept its first production workload. Hyperscalers accept this risk because they build fast and iterate. The federal government will not, and the gap between those two speeds is where the project's economic viability goes to die. Paducah is a test. Not of whether the United States can build AI data centers — it can, and it already has. The test is whether a democratic society can repurpose its most toxic industrial inheritance without repeating the ethical failures that produced it. The announcement hides the ledger: the contaminated groundwater, the deferred cleanup, the missing tenant, the absent oversight. But the ledger writes itself eventually. The chain remembers what the market forgets. So does the soil. If the conversion happens with verifiable remediation, genuine community partnership, and cryptographic accountability for the compute inside those walls, it could be the most important infrastructure project of the decade. If it happens as a costume for deferred responsibility, it will be remembered as another way to move poison to the margins. Infrastructure is ideology made physical. The question is not whether we can build it. The question is whether we can build it honest.