The AI Investment Race: Why Real Demand Can Still Produce Fragility
This series examines the current AI infrastructure build-out as a competition among hyperscalers, AI laboratories, chipmakers, neoclouds, lenders, and investors. Its central question is not whether AI has value. It is how a real expansion in AI use can coexist with an investment boom that becomes unusually exposed to disappointment, financing stress, and contagion. The aim is to give the reader a practical mental model: separate evidence of demand from evidence of sustainable returns, then follow how competition and balance-sheet choices turn one into the other.
The subject is a race for AI infrastructure
AI services require physical capacity: advanced processors, data centres, power, cooling, networking, and the specialised systems that turn computation into a usable product. The firms building that capacity are not making decisions in isolation. They are competing for a small number of positions that may control the most valuable models, users, developers, and data. A coalition that arrives late may face higher switching costs and a smaller share of the market even if its technology is good. That possibility changes the meaning of investment. Capacity is not only a way to serve demand; it is also a way to prevent a rival from becoming the platform through which that demand is served.
The Bank for International Settlements working paper that anchors this analysis describes the scale in those terms. It reports that capital expenditure by major hyperscalers is set to exceed $700 billion in 2026 alone, while expectations for aggregate AI-infrastructure investment over the following years run into the trillions. Those figures are not presented as proof that the build-out is irrational. AI may generate enough productivity and revenue to justify a large expansion. They establish why the question has moved beyond the fortunes of a few technology companies: a reversal would involve specialised capital, lenders, suppliers, investors, and the firms whose valuations incorporate expectations of AI growth.
The financing relationships matter because the build-out is partly organised through partnerships rather than simple arms-length purchases. A hyperscaler may take an equity position in an AI lab while the lab commits to buying future compute. A chipmaker or neocloud may sit in the same web of supply and funding. Such arrangements can solve a genuine coordination problem. The builder wants confidence that capacity will be used; the lab wants confidence that capacity will be available. But the same links can make one firm’s disappointment relevant to another firm’s balance sheet.
That is the subject here: not “AI” as an undifferentiated technology, but a concentrated, financed race to build the infrastructure on which AI services depend. Financial fragility means that the system’s capacity to absorb a disappointment is weakened by the way the boom has been built. It does not mean that the technology is worthless, that every project is uneconomic, or that a bust is certain. It means that scale, timing, leverage, specialised assets, and cross-holdings can make the consequences of a bad draw larger than the original disappointment.
Three features make this exposure different from a simple story about a company’s sales slowing. First, infrastructure is lumpy and partly specialised. A processor, data-centre fit-out, power connection, or custom arrangement may be highly valuable inside the intended AI ecosystem and much less valuable if many builders retrench simultaneously. The asset can therefore lose value at the same moment that several potential buyers are trying to preserve their own cash. Second, the clocks do not match. A firm can sign a multiyear purchase commitment and borrow against expected future revenue before it knows how quickly users will adopt a service or how quickly prices will fall. Third, the relationships are concentrated. If the same backer finances more than one lab, or owns a stake in one lab while supplying another, an impairment in one position can reduce the resources available to support the others. These are transmission mechanisms, not assumptions that a downturn will automatically become a crisis.
The distinction also matters for how readers should interpret the word “race.” A race does not require irrational participants. Each firm may be responding sensibly to the possibility that being late is more costly than building excess capacity. The collective result can still be unstable because the firms do not fully value the revenue they deny to rivals, the resale losses they impose on one another in a common liquidation, or the collateral damage that travels through shared counterparties. The question is therefore about incentives and exposure, not about assigning motives to particular executives or declaring the technology a bubble in advance.
Demand is growing; revenue is a separate question
There is strong evidence that AI use is expanding. An independent study based on licensed OpenRouter data tracks realised requests rather than survey responses or company projections. Across more than 400 large language models, the dataset records 380.8 trillion tokens from January 2024 through April 2026. Weekly token consumption rises from 11.4 billion at the beginning of the sample to 15.6 trillion at the end. The dataset covers approximately 2% of current global monthly token consumption, so it is not a complete census, but its user–model–day structure provides unusually direct evidence that people and organisations are using more AI.
The study also finds that the composition of use is changing. Agentic requests—workflows in which a model plans, calls an external tool, takes an action, or coordinates with another agent—rise from a negligible share in 2024 to 52.2% of tokens by the latest full week in the sample. That sounds like an unambiguous argument for more infrastructure, but it is not an unambiguous argument for more revenue. Agentic token volume grows while dollars per agentic token fall, partly because prompt caching avoids repeated computation and providers route requests toward cheaper backing models. More computation can therefore be demanded at the same time that the price earned per unit of computation declines.
This distinction is essential for understanding an infrastructure race. Usage is a quantity. Revenue combines quantity with price. Sustainable investment returns must go further: they must cover operating costs, the cost of capital, depreciation, and the losses that occur if capacity cannot be redeployed at its original value. A fast-growing token series can support the case for expansion without establishing that every additional data centre will earn an adequate return.
Financial markets appear to recognise that AI exposure is both an opportunity and a risk. The OpenRouter study estimates firm-level exposure from the way stock returns move with a weekly factor built from AI tokens, dollars, and active users. In its value-weighted US portfolio, firms with high exposure outperform low-exposure firms by 64.1 basis points per week over the sample; the spread remains 55.9 basis points after Fama–French five-factor and momentum adjustments. The result also survives controls for technology stocks, semiconductors, AI-themed funds, industry, and broad search attention. The authors interpret this as a market price for AI exposure, not as a stable long-run return or a measure of causal productivity.
The breadth of that exposure is informative. It reaches beyond firms normally labelled as technology companies, into consumer-facing and capital-heavy industries. That makes the AI investment race economically consequential, but it does not settle whether the build-out is too large. A market can attach a premium to an uncertain growth exposure precisely because investors require compensation for holding it. Nor does a return spread from 2024–26 tell us what the premium will be after the technology matures. The evidence supports a narrower conclusion: AI expectations are already influencing a broad set of valuations, while the underlying demand and pricing process remain in motion.
The mechanism to keep in view
The core mechanism can be stated without starting with equations. Suppose several coalitions are trying to become dominant suppliers. Each one considers the revenue it can earn from additional capacity, but also the market share it can take from rivals. The private incentive is therefore larger than the sector-wide gain. What is a strategic victory for one coalition may be partly a transfer from another coalition rather than new output for society as a whole.
The BIS model formalises this as a winner-take-most contest with diminishing returns at the sector level. Aggregate compute raises AI revenue, but the revenue function is concave: as capacity expands, each additional unit contributes less to total sector output. A coalition nevertheless receives a contest payoff tied to its relative capacity and its early position. When it chooses investment, it does not fully subtract the business it steals from rivals. The planner, by contrast, counts the sector’s total output and the cost of capital. That gap is the first source of overbuilding.
The next source is timing. If early capacity attracts users, data, developers, and research talent, then a firm may commit before it knows how productive the technology will be. Waiting preserves the option to learn. Waiting may also surrender the position that makes the investment valuable. The race consequently shifts spending into the period when uncertainty is greatest. Debt and circular financing make that early commitment possible, but they also put more specialised capital at risk if the expected growth does not arrive.
The sequence is therefore more specific than “optimism causes a bubble.” Real demand makes expansion plausible. Competition can push the level of capacity above the sector’s efficient level. Early-mover advantages can pull the spending forward. Circular equity ties can coordinate demand and supply while connecting balance sheets. Debt can enlarge the build-out while making a synchronised retrenchment harder to absorb. If the disappointment is large enough, a thin resale market can turn a decline in expected revenue into a loss on the infrastructure itself. Cross-holdings can then transmit the stress to firms that were not the original source of the problem.
This map does not say that the final node must occur. It says that the links are economically plausible and can be studied separately. Demand evidence speaks most directly to the first node. The contest model explains the second and third. The bust and network extensions explain the later nodes. None of those forms of evidence should be substituted for another.
The unresolved question is why rational firms would choose a level of investment that leaves the sector collectively more exposed than a planner would choose. Answering it requires looking closely at the private return to winning the race, rather than treating aggregate demand as the only determinant of investment. That is where the next step begins.