Five Years of CPU Demand: Why the AI Infrastructure Boom Is Outpacing Its Own Evidence
Lisa Su, AMD's chief executive, told audiences at a recent industry event that the market for central processing units will grow massively over the next five years, driven largely by enterprise adoption of AI infrastructure. The claim is not isolated. Across the semiconductor sector — from AMD to Nvidia to the component suppliers below them in the value chain — the message from manufacturers to investors has been consistent: demand for computing hardware will sustain elevated growth through the end of this decade.
Simultaneously, a joint Gallup and National Bureau of Economic Research survey found that 89 percent of business leaders reported no measurable impact from AI on their company's labor productivity in the past three years. When pressed on the question of whether AI had improved output per worker, most said it had not. The disconnect between what the hardware sector is selling and what buyers are experiencing has become one of the defining paradoxes of the current tech cycle.
The productivity gap in plain numbers
The Gallup-NBER findings are not a fringe result. The survey drew from a broad sample of executives across sectors and firm sizes. Its framing was direct: in the past three years, has AI meaningfully changed how much your workforce produces per hour, per dollar, or per output unit? For nearly nine in ten respondents, the answer was no.
The result sits in tension with the scale of capital expenditure flowing into AI. Corporate AI spending globally has reached into the hundreds of billions of dollars annually. Hardware procurement — the CPUs, GPUs, and network infrastructure that underpin AI workloads — represents the largest single line item in those budgets. If the productivity gains were materialising at scale, it would be showing up in these surveys.
There are several plausible explanations for why the gap exists. The most straightforward is measurement: companies may be spending on AI infrastructure while lacking the internal analytics to attribute output changes to the technology. AI often displaces routine tasks rather than amplifying the output of remaining workers — a dynamic that can hold revenue flat while altering the composition of work. In that scenario, the productivity signal takes longer to appear in aggregate statistics.
A second possibility is implementation lag. AI tools have been commercially available at scale for fewer than three years. Enterprises that began deployment in 2023 or 2024 may simply not have reached the point where workflow redesign and tool integration generate measurable efficiency gains. The Gallup finding captures a moment in a process, not its conclusion.
A third explanation — the one the hardware sector has an economic interest in suppressing — is that AI is not delivering the productivity gains it promised. The semiconductor companies selling the infrastructure understand that their revenue growth depends on the premise that AI will eventually transform how work is done. That premise has not yet been confirmed by the buyers AMD and its competitors are relying on.
What the hardware cycle is actually signalling
AMD's five-year demand forecast is not based on corporate sentiment surveys. It is based on order books, channel data, and the inventory patterns of hyperscale cloud providers and enterprise buyers. CPU sales — the category Su explicitly named — are a proxy for infrastructure buildout: to run AI workloads at scale, companies need server hardware, networking equipment, and the semiconductor components that tie them together. AMD's position at the centre of that supply chain gives its projections a specificity that analyst commentary often lacks.
The signal is consistent across the sector. TSMC has expanded advanced packaging capacity. Memory manufacturers are scaling high-bandwidth DRAM to meet demand from AI training clusters. The entire supply chain, from raw silicon to finished systems, is running at utilisation levels that suggest demand is not theoretical.
Yet infrastructure sales are not the same as productivity returns. A company that buys $500 million in computing hardware to run AI applications has made an investment in the preconditions for productivity. Whether those preconditions materialise into actual output gains depends on factors the semiconductor manufacturers cannot control: the quality of AI models deployed, the readiness of internal processes to absorb them, and whether the competitive environment rewards efficiency gains or simply rewards the adoption of new technology as a signalling mechanism.
The structural frame: a bet, not a proof
The conversation happening inside corporate boardrooms right now is essentially a wager. Business leaders are spending heavily on AI infrastructure in the expectation that the technology will eventually generate measurable returns. The hardware manufacturers are selling into that expectation and pricing it into their growth forecasts. Neither side has the proof the other needs.
This is not unusual in technology cycles. The productivity impact of enterprise computing took most of a decade to appear in aggregate US economic data, long after the hardware had been deployed. The internet transformed business models over a longer horizon than early investors expected. AI may follow a similar path — the gains real but slower to materialise than the spending would suggest.
The risk is asymmetric. If AI does deliver transformative productivity gains, the companies that built infrastructure early will hold structural advantages: better-trained models, more mature operational processes, and a deeper understanding of where the technology actually creates value. If it does not — if the returns plateau or the efficiency gains are offset by new categories of cost — the hardware manufacturers will have captured their revenue while the buyers absorb the impairment.
The government figures on improper payments in fiscal year 2024 offer a secondary data point. Across 68 federal programs, agencies reported $162 billion in improper payments — a figure that reflects not AI failure but the broader challenge of ensuring that large-scale capital deployment produces intended outcomes. When spending runs ahead of measurement and oversight, inefficiency becomes structural.
What comes next
The five-year CPU growth horizon AMD has articulated will be tested by what happens in years two and three. The survey data on AI productivity is not the final word — it captures a specific moment in a rollout process that has years to run. But it is the most direct signal available from the people actually deploying AI at scale: the corporate leaders whose capital expenditure is funding the semiconductor cycle.
Those leaders are buying hardware while reporting that AI has not changed their productivity. That is not a contradiction — it may be a rational response to a technology whose returns are real but deferred. It may also be a market dynamic that the hardware manufacturers are well aware of, and that explains why their growth projections, while bullish, are carefully hedged around the pace of enterprise adoption rather than the pace of productivity return.
The Epoch Times reported on 22 May 2026 that House Republicans withdrew a scheduled vote on the Iran War Powers Resolution, an episode that underscores the extent to which AI policy and geopolitical competition are now entangled at the legislative level. The infrastructure AMD is selling feeds directly into that competitive environment — and the productivity question will determine whether the investment is a strategic advantage or a very large sunk cost.
The hardware is moving. The gains are not yet visible. The bet, for now, remains open.
This desk watched how the wire framed AMD's demand forecast versus the Gallup productivity data — they ran as separate stories; this piece argues for reading them together as a single market signal.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://t.me/nikkeiasia/15162