Palmisano Warns AI's Long-Cycle Bets Face a Fast-Moving Market

Former IBM chief executive Sam Palmisano told Bloomberg Markets that the AI boom rests on long-cycle investments in data centers, chips and energy, and that Big Tech may have to rethink those spending plans if adoption slows or cheaper Chinese open-source models gain traction.

The most awkward fact about the artificial intelligence buildout is a matter of arithmetic on time, not money. Concrete, transformers, substations and land take years to commit and years to pay back. The software running on top of them can be rewritten in a quarter — and undercut by a free download in a weekend.

That is the mismatch Sam Palmisano, the former chief executive of IBM, laid out in an interview with Bloomberg Markets. His argument: the AI boom is built on long-cycle bets in data centers, chips and energy, even as AI software changes at breakneck speed. If adoption slows, or if cheaper Chinese open-source models gain traction, Big Tech may have to rethink the spending plans behind today's race.

It is a warning delivered from inside the industry's own history. Palmisano ran a company that lived through the aftermath of one infrastructure boom and spent the better part of a decade reorganising around what was left of it.

Why the duration mismatch is the whole argument

Infrastructure investment has a specific economic character. You commit capital up front, you depreciate it over a long life, and you need a reasonably predictable stream of demand to justify the commitment on day one. Utilities work this way. So do railroads, fiber networks and, increasingly, AI data centers.

Software does not work this way at all. Its marginal cost is close to nothing, its release cycle is measured in weeks, and a better approach can render the previous one commercially irrelevant almost immediately. When an industry stacks fast-moving software economics on top of slow-moving infrastructure economics, the risk is not that the infrastructure fails. It is that the infrastructure is sized for a demand curve that the software no longer generates.

Palmisano's framing puts two specific triggers on that risk. The first is adoption: enterprises buying AI capacity more slowly, or in smaller increments, than the buildout assumes. The second is substitution: cheaper open-source models from China taking share from expensive frontier systems, compressing the price of inference and, with it, the revenue that is supposed to service all that steel and silicon.

Energy is the part that cannot be rescheduled

Of the three legs Palmisano names — data centers, chips and energy — energy is the least flexible. Chips can be reallocated between workloads or resold. A data center shell can be repurposed, slowly. But grid interconnections, generation capacity and long-term power contracts are negotiated years ahead and are difficult to unwind without cost.

That gives the energy leg an outsized role in any reassessment. A hyperscaler that decides to slow its pace of construction still has to deal with the power commitments it has already signed. Utilities, independent power producers and equipment suppliers have in turn been planning their own capital programs against those commitments. A change in tone from the buyers propagates outward through a supply chain that is itself long-cycle.

This is why Palmisano's caution is a broader-market issue rather than a technology-sector one. The AI trade has spread well beyond the model developers into power, industrial equipment, cooling, construction and the credit that finances all of it.

Markets are not pricing the catch

Nothing in the tape suggests investors have taken the warning to heart. At the last close before the interview aired, on Friday, 21 August 2026, the SPDR S&P 500 ETF Trust (NYSEARCA: SPY) finished at $765.72, up 0.41% on the day from a previous close of $762.60, having traded between $764.17 and $767.85.

The technology-heavy Invesco QQQ Trust (NASDAQ: QQQ), the vehicle most directly exposed to the AI complex, closed at $713.44, a gain of 0.35% from $710.93, with a session range of $709.20 to $715.67. The SPDR Dow Jones Industrial Average ETF Trust (NYSEARCA: DIA) was the strongest of the three, closing at $532.22, up 0.89% from $527.51.

Two things are worth noting there. One is that the Nasdaq proxy lagged both the broad market and the Dow on the day — a modest reversal of the leadership pattern that has defined the AI rally, though a single session proves nothing. The other is that all three closed in the upper part of their daily ranges. This is not a market bracing for a capex reset.

What a rethink would actually look like

Investors watching for evidence that Palmisano's scenario is playing out should be looking for changes in language before changes in numbers. Capital expenditure guidance rarely gets cut in one step. It gets qualified first: phrases about "pacing", "flexibility", "optionality on the build schedule", or a shift from committed capacity to leased capacity.

• Depreciation assumptions. Extending the useful life of AI servers flatters near-term earnings; shortening it signals that the hardware is obsoleting faster than planned. Either direction tells you something.
• Lease versus own. A move toward third-party capacity is a way of buying time without buying concrete.
• Power contract news. Renegotiated or deferred energy agreements would be the clearest early tell, because they are the hardest commitments to walk back quietly.
• Inference pricing. If the cost per token charged to enterprises keeps falling, the revenue side of the buildout thins even if usage grows.
• Open-source share. Enterprise adoption of freely available models — Chinese or otherwise — directly reduces the addressable market for paid frontier systems.

The historical rhyme, and its limits

The obvious comparison is to the telecom buildout of the late 1990s, when carriers laid vastly more capacity than near-term demand required. That episode is instructive but not identical. Fiber, once laid, is essentially permanent and its capacity can be unlocked with cheaper electronics. AI compute depreciates on a hardware cycle, and it consumes power continuously whether or not it is earning revenue. Idle fiber is cheap to own; idle GPUs are not.

That cuts both ways for the current boom. It means an overbuild is more expensive to carry than the last one. It also means capacity is retired rather than left dark, which can clear an overhang faster.

Palmisano is not calling a top. He is pointing at a structural feature of the trade that is easy to lose sight of while the numbers keep going up: the assets are slow, the market they serve is fast, and the two only have to fall out of step once for the spending plans behind them to require an explanation.

Key facts

• Nasdaq 100 proxy (NASDAQ: QQQ): $713.44 at the close, +0.35%, as of 21 Aug 2026 20:00 GMT
• S&P 500 proxy (NYSEARCA: SPY): $765.72 at the close, +0.41%, as of 21 Aug 2026 20:00 GMT
• Dow proxy (NYSEARCA: DIA): $532.22 at the close, +0.89%, as of 21 Aug 2026 20:00 GMT
• The thesis: Long-cycle bets in data centers, chips and energy versus fast-changing AI software

Frequently asked questions

What exactly did Sam Palmisano say about AI spending?

The former IBM chief executive told Bloomberg Markets that the AI boom is built on long-cycle bets in data centers, chips and energy, while the AI software running on that infrastructure changes at breakneck speed. He said that if adoption slows or cheaper Chinese open-source models gain traction, Big Tech may have to rethink the spending plans behind today's AI race.

What is a 'long-cycle bet' in this context?

It is capital committed years before it generates revenue and recovered over many more years. Data centers, semiconductor fabrication capacity and power generation all fit: land, permits, grid connections and construction are locked in well ahead of demand. The commitment cannot easily be reversed if the demand assumptions behind it change.

Why do cheaper Chinese open-source models matter to the spending case?

Freely available models compress the price enterprises are willing to pay for AI capability. If comparable performance can be obtained without paying for a frontier system, the revenue supporting the infrastructure buildout thins even if total AI usage rises. That undermines the payback assumptions used to justify the original capital commitments.

How did the major indices close before the interview aired?

At the last trade on Friday, 21 August 2026, the S&P 500 proxy SPY closed at $765.72, up 0.41%. The Nasdaq 100 proxy QQQ closed at $713.44, up 0.35%. The Dow proxy DIA was the strongest, closing at $532.22, up 0.89%. All three finished in the upper part of their daily ranges.

What early signs would show Big Tech is actually rethinking AI capex?

Language usually shifts before numbers do. Watch for qualified capital expenditure guidance, changes to the assumed useful life of AI servers in depreciation schedules, a shift from owned to leased data center capacity, and any renegotiation or deferral of long-term power purchase agreements — the hardest commitments to unwind quietly.

How does this compare with the late-1990s telecom buildout?

The parallel is imperfect. Fiber laid in that era was permanent and cheap to hold idle, and its capacity could later be unlocked with better electronics. AI compute depreciates on a hardware cycle and consumes power whether or not it is earning. An overbuild is costlier to carry, but capacity is retired rather than left dark.

Sources

• The AI Spending Spree Comes With a Catch — Bloomberg Markets

Photo: _ Whittington · Pexels Licence — source