Don't simply invest in AI companies. Invest in the economic infrastructure and productivity transformation AI creates — because those are not the same trade, and conflating them is where this cycle's real risk sits.
US information technology investment, as a share of GDP, has risen to its highest level since 2001. That is not a neutral comparison. 2001 is the dot-com peak — the last time American businesses poured capital into a technology this fast, this early, on the belief that the productivity payoff would justify the spend. The IMF flagged this figure itself in its latest World Economic Outlook, in the same breath as its growth upgrade. That pairing is the whole story of this letter: the investment is real, the growth contribution is real, and the IMF is naming the 2001 comparison on purpose.
The IMF's January 2026 outlook raised global growth to 3.3% for the year, crediting resilient activity and heavy AI-linked investment, particularly in North America and Asia. It did not stop there. In the same report, the Fund named the downside explicitly: "a recalibration of expectations regarding AI-driven productivity gains might lead to a significant drop in investment and a correction in equity markets, which have been increasingly influenced by a few large technology companies." The upside is named with equal specificity: "the accelerated and widespread adoption of AI could significantly improve productivity and stimulate more vigorous medium-term growth." This is not hedging. It is the IMF placing two genuinely different futures on the table and declining to pick one — which is itself the most honest signal available right now, and the reason this letter treats the question as genuinely open rather than a call to make either way.
A dedicated IMF scenario-planning exercise, run with roughly fifty participants across IMF departments and outside AI experts, sharpened the mechanism further: AI adoption can amplify investment booms and market concentration, creating new transition risks even when long-term productivity prospects improve. That distinction matters more than almost anything else in this debate. The risk this letter is tracking is not "AI turns out to be overhyped." It is that AI can be a genuinely transformative technology over ten years and still produce a genuinely damaging correction over the next two, because the investment cycle and the productivity delivery timeline do not have to move at the same speed.
The IMF's own research adds a second, more technical amplifier worth naming plainly: AI itself is now embedded in the trading infrastructure that would have to absorb any correction. Machine-learning models generate high-frequency signals, and IMF analysis has found that AI-based funds rebalance meaningfully faster than traditional strategies — a dynamic that helps liquidity and price discovery in calm markets, and amplifies swings specifically when many models respond to the same signal at the same time. A sentiment shift on AI-driven productivity expectations would be exactly the kind of shared signal that triggers correlated rebalancing across AI-driven funds simultaneously. The mechanism that makes the boom possible is the same mechanism that could make its unwind sharper than prior cycles.
The IMF also names the specific physical constraints that could cap the productivity payoff regardless of how good the frontier models get: energy and grid capacity, data-center infrastructure, and the simple persistence of tasks that still require physical presence. These bottlenecks can limit economy-wide productivity gains "even under rapid improvements at the frontier" — meaning the model can keep getting smarter while the economy's ability to actually deploy that intelligence at scale lags behind, for reasons that have nothing to do with the technology itself.
This is the distinction this letter is built around: AI as a technology and AI as an economic transformation are not the same trade. The technology can keep improving on every published benchmark while the economic transformation stalls on grid capacity, permitting delays, chip supply, and the unglamorous work of actually integrating new tools into existing workflows. Data centers, chips, electricity generation and transmission, and the robotics now moving from research labs into real factories and warehouses are the physical infrastructure of this cycle — assets whose value depends on utilization and cash flow, not on sentiment about the next model release. A GPU cluster earns its return by running workloads at a real price, continuously, for years. That return is measurable in a way that "AI exposure" as a narrative is not.
Which businesses actually capture the productivity gains is a separate, harder question from which businesses are building the infrastructure — and it is the question that will separate this cycle's actual winners from its participants. The 2001 comparison is instructive here too: the infrastructure build-out of that era was mostly real and mostly used eventually, but the equity value assigned to unproven business models ahead of that usage was not, and the correction fell hardest on the companies where the story was ahead of the cash flow.
The IMF is not predicting a crash. It is naming, in the same paragraph, that AI could drive the strongest medium-term growth in years or trigger a correction severe enough to weaken household wealth — and that the difference between those two outcomes is not the technology, but whether the productivity actually shows up on the timeline the market has already priced in.
Not the AI narrative. Cash flow and utilization, specifically: whether data center and chip capacity is running at high utilization and earning a real, sustained price for compute, not merely being built; whether the businesses deploying AI internally are showing it in measurable productivity metrics — output per employee, cycle time, error rates — rather than in earnings-call language about AI strategy; and whether capital expenditure is being funded by real operating cash flow or by debt issued against a productivity payoff that has not yet arrived. The IMF's own 2001 reference point is precise about which of those failed last time: the infrastructure got built and eventually got used, but the capital structure financing the equity story in front of it did not survive the gap between spending and delivery.
This resolves in a way that is checkable well before any single earnings season settles it. If aggregate measures of business productivity growth accelerate over the next several quarters in step with the current pace of AI capital expenditure, the optimistic IMF scenario is being validated in real time. If capital expenditure keeps accelerating while productivity statistics stay flat — the actual pattern the IMF is naming as its risk case — that gap is the signal to watch, not any single stock's valuation.
US IT investment at its highest share of GDP since 2001 is a real, measured fact, not a narrative. The IMF has named both directions this can resolve in the same report, without picking one — which is the most credible position available given the evidence. The distinction that actually matters for capital allocation is between AI as a technology, which keeps improving regardless, and AI as an economic transformation, which depends on grid capacity, chip supply, and real productivity showing up in the numbers on a timeline markets have already priced in. Invest in the infrastructure and the productivity transformation. The AI narrative alone is not an investment thesis.