AI is attracting more capital than almost any technology in history. Governments are racing to announce national strategies. Hyperscalers are committing hundreds of billions of dollars to data centres. Venture funds are being raised at record pace. Every headline reads like a victory lap.
And yet: money flooding into a technology sector at this speed and scale is not a straightforward signal of health. It is also — in every prior technology cycle — the moment when things start to go wrong.
Both of those things are true at once. And they cannot coexist indefinitely.
The Reading
The Assumed Story: AI Investment as Proof of Arrival
The dominant narrative is seductive and not entirely wrong. Artificial intelligence is generating genuine commercial value — in drug discovery, software development, customer service automation, and logistics optimisation. OpenAI, Anthropic, and a widening roster of startups have demonstrated that large language models and related systems can do economically useful things that could not be automated before. Investors have noticed. So have governments, pension funds, and sovereign wealth vehicles.
The result is a capital commitment that, in nominal terms, may have no peacetime precedent in a single technology category over such a compressed period. Data centre construction, GPU procurement, energy infrastructure, and model training costs are all surging simultaneously. Hyperscalers — Microsoft, Google, Amazon, Meta — have collectively telegraphed capital expenditure plans running into the hundreds of billions of dollars over the next few years. In most industries, that level of simultaneous investment by multiple major players would simply be called a boom.
In tech history, it has sometimes been called something else entirely.
The Overlooked Angle: Capital Cycles Don’t Care About the Technology’s Merit
The uncomfortable insight — one that the sheer noise of the AI moment tends to drown out — is that a technology being genuinely transformative and a technology investment cycle being dangerously overextended are not mutually exclusive. They have, in fact, repeatedly coincided.
The internet was genuinely transformative. It also produced one of the largest capital misallocations in modern history. Fibre-optic infrastructure built at enormous cost during the late 1990s sat dark for years before demand caught up. The companies that eventually won — Amazon, Google — were survivors of a wave that destroyed most of what it funded.
The pattern is not that the underlying technology fails. It is that the rate of capital deployment outpaces the rate at which the technology can generate returns — and then the correction comes, often violently, regardless of the technology’s long-run merit.
There are structural reasons to think AI may be entering that zone. The cost to train and serve frontier models is enormous and, for most applications, still rising faster than the revenue those applications generate. AI token costs remain a significant constraint on profitable deployment, even as inference efficiency has improved. Enterprise adoption is real but uneven; the majority of AI pilots have not yet translated into the kind of at-scale, measurable ROI that would justify current infrastructure valuations on a discounted cash flow basis.
What makes the current moment structurally distinct from the dot-com era is that the primary capital commitments are coming not from public markets and retail investors but from large, balance-sheet-rich technology incumbents. This means the immediate bust scenario — a sudden collapse of funding and mass startup failure — is less likely than in 2000. But it also means that when the ROI reckoning arrives, it will land on the income statements of some of the most systematically important companies in the global economy, with consequences for equity markets, pension allocations, and technology labour markets that a purely VC-funded bust would not have triggered.
The Evidence: Where the Warning Signs Are Accumulating
The warning signs are not hidden. They are hiding in plain sight inside the optimistic framing that surrounds them.
Consider the widespread acknowledgement that most AI spending at the enterprise level is not yet generating clear returns. Surveys of enterprise technology leaders consistently find that pilots proliferate while production deployments at meaningful scale remain rare. The gap between AI experimentation budgets and AI revenue attribution is wide and, in many organisations, widening.
Consider also the energy constraint. AI data centres are consuming electricity, water, and land at a rate that is already straining grids and prompting regulatory scrutiny in multiple jurisdictions. The infrastructure required to sustain current AI ambitions is not a software investment that can be written off; it is physical, long-lived, and exposed to demand risk if the commercial thesis for AI services does not materialise at the projected scale.
And consider the strategic logic driving the investment itself. Much of the hyperscaler capex commitment appears driven less by a clear return calculation and more by a fear of being left behind — what economists sometimes call a coordination problem or an arms race dynamic. When the primary justification for a capital expenditure is “our competitors are spending it,” that is a signal about competitive anxiety, not commercial confidence.
Prominent investors have begun raising these concerns publicly. Michael Burry, the investor who famously shorted the 2008 housing market, has positioned against AI-adjacent equities — a contrarian stance that attracted significant attention precisely because the bull case appears so self-evident to most market participants.
What This Changes: The Politics of AI Capital
The investment surge is not just an economic story. It is reshaping political economies in ways that could lock in current trajectories regardless of whether the financial case holds up.
Governments that have staked industrial policy credibility on AI — including the United States, the United Kingdom, France, India, and others — now have strong political incentives to sustain the narrative of AI as a national priority. Export controls on AI chips and related hardware represent sovereign attempts to weaponise the capital cycle in favour of domestic players, adding a geopolitical dimension that makes an orderly slowdown politically difficult to engineer even if it were economically desirable.
There is also a labour market dimension. Senior policymakers are already reconfiguring education and workforce policy around AI assumptions that have not yet been validated at the economic scale being assumed. If the investment cycle corrects before those workforce transitions are complete, the social cost will be borne by workers and graduates — not by the hyperscalers that drove the capital deployment.
The Strongest Counterargument
The most credible pushback against the warning-sign thesis comes not from the AI industry’s promotional apparatus but from serious economic historians and technology analysts who point to a different historical precedent: the electrification of American industry in the early twentieth century.
In that case, the productivity gains from electricity took roughly two decades to show up clearly in economic statistics — not because the technology was failing, but because firms had to reorganise their entire production processes to extract value from it. The upfront capital investment looked disproportionate to short-run returns for years. Observers at the time who called it a misallocation were eventually proved wrong by a productivity surge that reshaped the global economy.
The argument, made by economists including MIT’s Daron Acemoglu and others in the broader AI economics literature, is that AI may require a similarly long gestation before returns arrive — and that judging the investment cycle by current ROI metrics is simply applying the wrong time horizon.
This is a genuinely strong objection. It is also not a complete answer. The electrification analogy assumes that the technology’s eventual productivity contribution will be large enough and broad enough to justify current capital levels — an assumption that remains contested. It also does not address the arms-race dynamic, which can produce overinvestment even in genuinely transformative technologies. The question is not whether AI will eventually matter. It is whether this specific investment cycle, at this specific pace, is calibrated to reality or to competitive fear.
The Prediction
Within the next 18 to 24 months, at least one major hyperscaler will announce a material reduction in its AI infrastructure capex guidance — not because AI has failed, but because the gap between investment and attributable revenue will become too wide to defend to shareholders. That announcement will be the catalyst for a broader re-pricing of AI-adjacent equities and a more sober public conversation about what AI is actually worth, commercially, right now. You will know this prediction is wrong if, by that point, enterprise AI deployments have scaled fast enough to produce clear, auditable productivity gains at the sector level — the electrification scenario, arriving on an unusually compressed timeline.
