MIT Technology Review is reporting on an assessment by Jessica Wachter, a finance professor at the University of Pennsylvania's Wharton School, who has been examining the economic stakes surrounding the current wave of artificial intelligence investment. The piece frames the analysis around a set of business and technical uncertainties so numerous that Wachter reportedly began her work by anchoring on what she describes as a "remarkable fact" that stands apart from the contested terrain.
The framing alone tells a story worth unpacking. When an economist at one of the country's most respected business schools finds it necessary to begin an AI analysis not with projections but with a search for stable ground, it reflects something important about where the industry actually stands. The scale of capital being committed to AI infrastructure — data centers, chips, energy capacity, talent — has reached a point where the word "trillion" is no longer rhetorical. It describes real allocation decisions being made by real institutions, including some of the largest companies in the world. And those decisions are being made under conditions of genuine uncertainty about whether the returns will materialize in any form resembling the investments made to produce them.
This is not a new dynamic in technology history, but the speed and concentration of the current cycle sets it apart in important ways. The dot-com era saw enormous capital destruction, but it was spread across thousands of companies and took years to build to a peak. The current AI build-out is happening faster, with far more capital concentrated in fewer hands. A small number of hyperscalers — the major cloud providers and a handful of frontier AI labs — are absorbing the bulk of the spending. That concentration means the consequences of being wrong are not distributed widely across a speculative market but are instead nested inside institutions that also underpin large portions of the broader economy.
There is also an asymmetry in who bears which risks. The companies spending most aggressively on AI infrastructure are, in many cases, doing so from positions of significant existing revenue and cash generation. For them, the downside of overbuilding is a drag on returns, not an existential threat. But the ecosystem that has grown up around the AI boom — the startups, the specialized hardware vendors, the consultancies, the enterprise software companies repositioning themselves as AI businesses — is far more exposed. If enterprise adoption of AI tools proves slower or more complicated than the current narrative suggests, or if the productivity gains prove difficult to measure and therefore difficult to justify in procurement decisions, the secondary tier of the industry faces a much harder reckoning.
What Wachter's framing surfaces, at least as MIT Technology Review presents it, is the intellectual challenge of even assessing the situation clearly. The uncertainties are not just about market timing or competitive dynamics. They are technical: whether current model architectures will continue to improve in economically useful ways, whether the energy and compute requirements will become prohibitive, whether the alignment between what large language models do well and what businesses actually need is as close as the sales cycle implies. Each of those questions sits upstream of the financial ones, which means that traditional tools for evaluating investment risk may not be well calibrated to the problem.
The likely consequence of this analytical difficulty is that a wide range of actors — investors, regulators, corporate boards, policymakers — will continue making large decisions on the basis of assumptions they cannot fully validate. That is not unusual in technology transitions, but it tends to produce corrections that feel sudden even when the warning signs were visible in retrospect. The honest reading of the moment is that there is enough genuine capability in current AI systems to sustain the narrative for some time, which means the cycle could extend further before stress becomes visible in the numbers.
For enterprise customers, the near-term consequence is pressure to deploy AI tools at a pace that may outrun their ability to evaluate what those tools are actually delivering. For the labor market, the uncertainty about AI's economic impact makes it genuinely difficult to plan around — which may itself be a kind of harm, separate from whatever the eventual outcome turns out to be.
The things worth watching in the period ahead are fairly specific. Whether major technology companies begin adjusting their capital expenditure guidance for AI infrastructure will be one of the clearest signals of whether internal confidence is holding. Productivity data, slow and imperfect as it is, will eventually need to show something. And the behavior of enterprise software renewal cycles — whether companies that deployed AI tools in the last two years are expanding, maintaining, or quietly contracting those deployments — will tell a more grounded story than any projection model can. Wachter's instinct to start from what is not in question was a sound one. There is not much else to start from.
