AI investment is real, but the compute it buys will not arrive until 2028 or 2029

A structural gap separates AI investment from AI capacity, and the gap will not close for years. Patrick O’Shaughnessy, host of Invest Like the Best, puts the timeline plainly: all the money that companies are putting in today manifests as compute in 2028 and 2029, not now. The implication is direct. The shortage that the industry describes as a present-tense problem will remain a present-tense problem well past the point when most planners expect relief.

The statement is worth dwelling on because the public narrative around AI infrastructure runs in the opposite direction. Announcements of capital expenditure are treated, routinely, as announcements of capacity. A company commits substantial sums to data center construction, and the implicit reading is that supply pressure will ease within a recognizable horizon. O’Shaughnessy’s observation cuts against that reading entirely. Capital committed today and compute available for use are separated by a multi-year production and construction pipeline. Conflating the two is not a minor forecasting error. It is a category error about how physical infrastructure works.

The lag is not mysterious in its origins. Data centers require permitting, grid interconnection, and physical construction, each stage measured in years rather than quarters. None of that compresses to match the urgency that AI demand timelines imply. The gap between announcement and operation is baked into the physical process, not a consequence of inefficiency that better management could eliminate.

So all today when we say there's not enough compute, it's not all the money that the companies are putting in today >> manifest in comput. No, it all manifests in compute in 2028 and 2029. Patrick O'Shaughnessy

The demand side of the equation does not wait for the supply side to catch up. Organizations building products on top of large language models, researchers running training runs, and enterprises deploying inference at scale are all competing for a pool of usable compute that, by O’Shaughnessy’s account, will not reach adequate scale until late 2028 or into 2029. In the interim, the shortage is the operating environment. It is not a temporary condition on its way to being resolved within the planning cycle of most organizations currently making infrastructure decisions.

What makes O’Shaughnessy’s framing useful is its specificity. He is not arguing that the industry is under-investing. The investment, on his account, is real and substantial. The point is that investment and available capacity are separated by a lag long enough to matter enormously for anyone whose decisions need to be made now. A company deciding whether to build proprietary infrastructure, contract for cloud capacity, or wait for market rates to fall is making that decision in an environment where the relief that capital investment implies is still years away.

The call is falsifiable, which is part of what makes it worth registering. If usable compute reaches adequate scale materially before late 2028, the timeline is wrong, and organizations that planned around it will have been too conservative. But the physical constraints O’Shaughnessy points to do not bend easily to optimism. Construction timelines, grid interconnection queues, and chip production cycles are not variables that respond to demand signals the way software production does. They respond to prior decisions made years earlier, which is precisely his point. The compute that will exist in 2026 and 2027 was, in most cases, already determined by decisions made before the current wave of AI investment accelerated. The compute that today’s investment produces lands in 2028 and 2029. Organizations that have not modeled for that reality are working from a plan that the calendar will eventually correct for them, on its own schedule and not theirs.