The Return Doesn’t Ship With the Chip

The number everyone quotes

Ask whether AI is being adopted and the answer looks settled. McKinsey’s 2026 State of AI survey found nearly nine in ten organizations using AI regularly in at least one business function. That’s the number that gets screenshotted.

It’s also the wrong number to anchor on. Usage is a growth figure. It tells you the tools arrived. It doesn’t tell you whether anyone is getting paid for them.

The same survey carries the operator’s tell. Companies attributing 5% or more of EBIT to AI, the group McKinsey calls high performers, make up about 6% of respondents, unchanged from a year earlier. PwC’s 2026 survey of more than 4,400 chief executives lands in the same place: 56% have seen no significant financial benefit so far, and only 12% report both higher revenue and lower costs.

Read those side by side and you get a technology that has been deployed far faster than it has been absorbed.

We’ve run this experiment before

Electricity is the clean precedent. In his 1990 paper “The Dynamo and the Computer,” economic historian Paul David showed that the productivity gains from electric power arrived decades after the technology was available. Early adopters swapped a central steam engine for a central electric motor and kept everything else: the overhead line shafts, the multistory layout, the workflow built around a single power source.

The gains came later, when factories were redesigned around small motors at each machine. Single-story plants. Material flowing in a line. Jobs reorganized around the new layout. That took a generation of managers and engineers who had learned to think in the new terms.

Brynjolfsson, Rock and Syverson gave the pattern a name: the productivity J-curve. General-purpose technologies require large, mostly invisible investments in process redesign, training and organizational know-how. While those investments are being made, measured productivity can look flat or even worse, because the costs show up and the output doesn’t yet. The returns come on the back side of the J.

AI is following the same shape. The compute is the dynamo. The courses that haven’t been written, the roles that haven’t been redesigned, the managers who haven’t learned to run an AI-augmented team: those are the line shafts still hanging from the ceiling.

The rung inside the customer

I’ve spent this year tracing how the buildout’s constraint keeps relocating. It moved off the chip and into capital, packaging, memory, power, water and finally consent, each rung answering less to money than the one before.

There’s a rung below all of them, and it isn’t at the datacenter. It’s inside the customer.

Call it absorption: an organization’s capacity to change how work gets done fast enough to use what it has bought. No capex reaches it. No toolmaker ships it. It moves at the speed of training, process redesign and management attention, which is to say slowly, one company at a time.

The data says it’s binding. Kyndryl’s 2026 People Readiness Report found only 23% of leaders say their workforce is ready for AI, down six points from 2025. The small group it calls Pacesetters, about 9% of organizations, have done three things: redesigned roles around AI, run change management so people understand the new operating model, and built workforce readiness. Accenture’s research found that while roughly a fifth of companies are overhauling workflows, fewer than 10% are rethinking the jobs themselves.

Both firms sell transformation services, so discount the framing accordingly. The direction still matches the independent numbers from McKinsey and PwC.

Where the leverage sits

It’s tempting to treat low adoption as a forecast to argue about: either the payoff is coming or it isn’t. That debate misses where the leverage is.

The low number describes where most organizations are today. Which way it breaks depends on what each organization does next. Companies that do the training, redesign the roles and manage the change will land on the back side of the J-curve. Companies that buy licenses and wait won’t. Both outcomes will happen, in different places, at the same time.

So I don’t think “will AI pay off?” is the useful question. The market doesn’t deliver the return on its own schedule. Each company earns it, or doesn’t, by deciding whether to do the work.

The other side deserves its due. BCG’s 2026 Applied AI Index, released this month, says almost half of companies now generate some value from AI, and puts the leading 7.5% at roughly 2.5 times the shareholder return. That’s real progress. The distance between “some value” and “value that moves EBIT” is the absorption rung expressed as a single statistic.

Why the buildout should care

Here’s where it connects back to the datacenter.

The buildout runs on two clocks. The financing clock: accelerators depreciate over roughly five to six years, and they’re being financed at the highest long-term rates in more than two decades. The absorption clock: process redesign and retraining measured in years per organization, across millions of organizations.

If the absorption clock runs longer than the financing clock, the first generation of hardware may be retired before the J-curve pays it back. The hyperscalers can carry that gap on cash flow. A leveraged developer financing a five-year asset may not make it to the payoff.

There’s a hopeful version of the same math. If absorption matures in the early 2030s, it arrives just as the first large wave of installed accelerators comes up for refresh. The hardware bought then goes into organizations that finally know how to use it. In that case the buildout’s real return doesn’t show up on the first generation of chips. It shows up on the second.

That makes the most important variable for datacenter demand five years out something you won’t find at the fab, the substation or the zoning board. It sits in the training budget and the operating model of the average company buying AI.

The close

You don’t remove a bottleneck. You relocate it. The buildout has pushed the constraint all the way through the physical world, and it has now landed where money moves slowest: inside how organizations learn.

The compute is shipping on schedule. The return doesn’t ship with it.

Look at the organizations you know. Which ones are doing the work, and which are waiting for it to arrive?