BadHabitus

← Back to Blog

Blog · AI infrastructure and political economy

The AI Stranded Asset Problem

Data centers, overcapacity, and the bubble that may not leave an infrastructure legacy

Every bubble leaves something behind. The question is what kind of remains.

The 2008 financial crisis destroyed lives, neighborhoods, balance sheets, pensions, local governments, and trust. But the houses did not disappear. Subprime borrowers were evicted. Mortgages were written down, sold, restructured, securitized, foreclosed, and absorbed into other portfolios. The social damage was enormous, but the underlying asset could be reoccupied, rented, sold, renovated, or folded back into the housing market. The house remained legible as a house.

The dot-com bubble left a different kind of wreckage. Capital was burned, firms vanished, frauds were exposed, and vast telecom investments were stranded in the short run. But fiber optic cable remained fiber optic cable. Dark fiber could be lit later. The overbuild that helped destroy speculative valuations also became part of the physical basis for the next stage of the internet economy. A bubble misallocated capital, but part of that capital survived as infrastructure.

The AI bubble may be different. If it breaks, it may leave behind assets that are far harder to reincorporate into any durable production regime: specialized data centers, power contracts, cooling systems, GPU clusters, substations, gas plants, grid upgrades, water claims, tax abatements, and debt structures built around a projected demand for compute that may not arrive at the promised scale. The remains may not become the next fiber backbone. They may become a new class of stranded asset.

This is the infrastructure underside of the arguments I have been developing in earlier AI posts. In The AI Substitution Trap, the central problem was that AI is sold as augmentation while institutions use it to reorganize labor, judgment, and authority around substitution. In AI and the Crisis of the Wage Form, the issue was the separation between socially produced intelligence and privately captured income streams. In The Token Budget Problem, the emphasis was exhaustion: human attention and institutional purpose compressed into a technical economy of prompts, outputs, and limits. This post follows the same logic into land, power, water, hardware, and debt.

The AI industry is not only building models. It is building a physical regime around the assumption that model scaling, inference demand, enterprise adoption, and consumer use will continue expanding fast enough to justify staggering fixed investment. McKinsey estimated in 2025 that data centers could require $6.7 trillion in global capital spending by 2030, with $5.2 trillion of that tied to AI workloads. The International Energy Agency projects global data-center electricity consumption roughly doubling to around 945 TWh by 2030, with AI as a central driver. In the United States, Lawrence Berkeley National Laboratory estimated that data centers consumed about 4.4 percent of U.S. electricity in 2023 and could rise to between 6.7 and 12 percent by 2028.

These projections are usually presented as proof of inevitability. They should also be read as evidence of risk. The AI boom is no longer a story about software firms with high margins and little physical footprint. It is becoming a story about fixed capital, energy systems, public utilities, construction timelines, local opposition, financing loops, and assets whose value depends on a future intensity of use that has not yet been secured.

The industry already shows signs of mismatch between announced capacity and executable demand. Sightline Climate estimated that 30 to 50 percent of the announced 2026 U.S. data-center pipeline was unlikely to come online before the end of the year. Other reporting has pointed to delays caused by grid constraints, power-equipment shortages, local opposition, and construction bottlenecks. This is not simple overcapacity in the old sense. Some operators still face shortages of the right kind of compute, in the right location, with the right power contract. But scarcity and overbuild can exist together: scarcity in immediate usable capacity, overbuild in announced projects and leveraged plans designed for demand curves that may bend before the assets pay back.

The Colossus example is revealing when stated carefully. xAI announced a compute partnership in which Anthropic would use Colossus capacity. Reuters later reported that Elon Musk described the arrangement as a six-month lease with a 90-day termination provision, despite earlier expectations of a larger and longer-term deal. That does not prove that xAI has no use for its own compute. It shows something more structurally interesting: even firms built around the public claim of existential compute scarcity may need to monetize capacity through short-term leasing to rivals.

That is one of the classic signs of a speculative infrastructure regime. Capacity is built ahead of demand. Then demand has to be manufactured, subsidized, redirected, or rented into existence. The industry begins to feed itself: cloud providers buy GPUs; model companies buy compute; chipmakers invest in cloud firms; start-ups use investor money to buy infrastructure services from firms that depend on the start-ups’ projected demand. This is not the same as saying all demand is fake. It is saying that the boundary between demand and financing becomes blurred.

The deeper problem is that compute demand is not like housing demand, and AI data centers are not like fiber. Housing need is socially brutal but durable. People need somewhere to live. Fiber carried far more future uses than investors could price during the dot-com crash. AI data centers are more specialized. Their value depends on particular hardware cycles, model architectures, power densities, cooling requirements, chip supply chains, and workloads. If the economic center of AI moves toward smaller models, cheaper inference, edge deployment, open-source efficiency, or specialized chips with lower energy requirements, then much of today’s buildout may not become the base of a future general-purpose infrastructure. It may become the wrong infrastructure.

DeepSeek exposed this possibility. In early 2025, the Chinese AI company claimed it had built competitive models with lower-cost chips and less capital than U.S. rivals, triggering market doubts about the assumption that frontier AI requires ever-larger expenditures on chips, power, and data centers. The point is not that DeepSeek disproves the need for data centers. The point is that it punctures inevitability. If capable AI can be produced with fewer high-end chips, cheaper training runs, better algorithmic efficiency, smaller models, or more targeted inference, then the projected path from more AI to more giant data centers becomes less secure.

AI may survive the bubble in a leaner form while the infrastructure built for the maximalist version loses its rationale. That would not mean AI was fake. It would mean the infrastructure regime built around it was overfit to one imagined trajectory of AI development.

This is what makes the stranded asset problem severe. A data center is not simply a warehouse. It is a bundle of physical and contractual commitments: land, shell, racks, chips, networking equipment, cooling systems, backup generators, power purchase agreements, transmission upgrades, water infrastructure, tax abatements, local political deals, and depreciation schedules. The GPUs themselves depreciate rapidly and are optimized for particular workloads. The buildings and grid connections may be reusable in some cases, but not automatically, not without further capital expenditure, and not at valuations justified by the AI mania.

What else absorbs this capacity? Cars will not. Consumer electronics will not. Ordinary enterprise software will not suddenly need the same concentration of high-end AI accelerators. Scientific computing, drug discovery, simulation, weather modeling, rendering, and other high-performance workloads can use compute, but they are not obviously large enough to absorb every speculative AI campus built around the expectation of mass-model demand. Some assets will find second lives. The question is whether enough will, at prices high enough, and quickly enough, to validate the debt and infrastructure built around them.

The energy constraint sharpens the issue. Data centers are not cheap warehouses full of servers; they are claims on power systems. The asset is not only the data center. The asset is the power regime built around it. Gas plants, transmission lines, substations, transformers, cooling infrastructure, and utility investments may be justified by the promise of future AI demand. If that demand fails to arrive, arrives in a different place, or arrives in a less energy-intensive form, the costs do not disappear. They migrate. Ratepayers, municipalities, utilities, and state governments may inherit the bill.

Wisconsin offers an early warning. Wisconsin Watch reported that ratepayers already owe nearly $1 billion on stranded assets, mostly power plants closed earlier than expected, and raised the question of whether new energy projects built for data-center demand could create another stranded-asset cycle. Separate reporting asked who pays if data centers do not materialize or do not use as much energy as predicted. The question is not local. It is the ratepayer version of the bubble problem: when speculative infrastructure is built through regulated utilities and public incentives, private overprojection can become a public obligation.

The parallel with fossil fuels is not accidental. Oil rigs, pipelines, refineries, coal plants, and gas infrastructure become politically explosive when their future use becomes uncertain. Once capital has been sunk, firms and regions seek protection. They demand subsidies, favorable regulation, public guarantees, delayed transition, emergency designations, rate recovery, and national-security rationales. The stranded asset does not quietly accept its impairment. It organizes politically.

This is where the AI infrastructure boom connects to Political Capitalism and the Afterlife of Monopoly Capital. The more fragile the economics become, the more political the asset becomes. If a data center cannot justify itself through ordinary market demand, it can still seek justification through defense, national competitiveness, public procurement, emergency AI adoption, utility-rate structures, tax incentives, and state-backed infrastructure. The bubble then tries to save itself by becoming policy.

That is why the comparison with earlier bubbles is too comforting. The 2008 housing crisis left homes, but also dispossession. The dot-com crash left fiber, but also bankruptcies. The AI buildout may leave power-hungry machine rooms, obsolete accelerators, strained grids, water conflicts, tax abatements, and energy projects whose repayment depends on a demand curve partly manufactured by the boom itself. The social losses could be borne by workers, ratepayers, municipalities, and public institutions long after private valuations have adjusted.

The industry will answer that demand is still growing. It may be right in the narrow sense. AI use will continue. Firms will integrate AI into workflows. Some models will become infrastructural. Data centers will not vanish. But the issue is not whether AI continues to exist. The issue is whether the maximal buildout now underway corresponds to a durable production regime, or whether it reflects a speculative attempt to force that regime into existence through capital spending before the use case, revenue model, energy system, and public legitimacy have stabilized.

The AI bubble is not only a stock-market story. It is a bet on a physical future. It assumes that intelligence can be industrialized through more chips, more power, more water, more campuses, more debt, more inference, more subscriptions, more enterprise adoption, more state support, and more everyday dependence. If the bet succeeds, we get a new infrastructure of machine intelligence governed by a handful of firms and utilities. If it fails, we may get something stranger: an infrastructure built for an intelligence economy that never quite arrived at scale.

That would make the AI data center the emblem of the present: a monument to externalized intelligence without a social purpose capable of absorbing it. In All That Is Solid Melts Into Search, I argued that modern culture increasingly turns inherited forms into retrievable content. The data-center boom is the material version of the same process. Human knowledge, language, judgment, and attention are externalized into systems that require enormous fixed capital to operate. Then the systems demand that society reorganize itself around their continued use.

The question is not whether AI can do anything. It can. The question is whether the infrastructure being built in its name corresponds to real human purposes, or whether the purposes will have to be invented afterward to protect the infrastructure. When an industry must stimulate demand to justify the machines already built, the machine has ceased to be an instrument and has become a claimant.

Every bubble leaves something behind. The danger of this one is that what remains may not be a usable foundation for the next economy. It may be a landscape of stranded compute, stranded power, stranded debt, and stranded public capacity, defended by firms that will insist the future still requires the very assets the present no longer knows how to use.

References