The Question Being Asked Out Loud
Amazon, Google, Microsoft, and Meta have collectively committed hundreds of billions of dollars to AI infrastructure expansion. Data centres are being built at a pace that has driven demand for power, cooling systems, copper wiring, and server hardware to levels that have materially shifted market dynamics in each of those component categories. The physical infrastructure buildout of AI is real, large, and generating genuine revenue for the companies that supply it.
The question that the semiconductor selloff of July 2026 is surfacing — and that investors are now asking loudly enough to move markets — is different: when do the hyperscalers start generating a return on those investments that justifies the capital deployed?
Broadcom's Q3 guidance of USD 16 billion in AI chip sales, which fell below the USD 17.2 billion analyst consensus and was not accompanied by an upgrade to full-year forecasts, was the proximate trigger for this week's selloff. But the question it raised is structural rather than quarterly. If the leading suppliers of AI infrastructure cannot consistently exceed the market's already elevated expectations, the investment thesis that has driven semiconductor valuations to extraordinary multiples requires reexamination. The broader concern is whether the economic value created by AI will be captured by the infrastructure layer — GPU manufacturers, memory suppliers, cooling system companies — or whether it will accrue primarily to the businesses that deploy AI to serve their customers.
The Infrastructure Vs Applications Debate
The internet precedent is instructive, even if imperfect. Cisco and Nortel built the physical infrastructure of the internet and watched their valuations collapse when the market concluded that the winners of the internet era would be the applications companies, not the network equipment vendors. The applications companies — Google, Amazon, Facebook, and eventually Apple — went on to generate returns that dwarfed what the infrastructure providers ever achieved. AI may follow a similar pattern: enormous productivity gains for the users of AI, concentrated in applications rather than infrastructure, with the infrastructure suppliers ultimately priced at commodity-level margins once the initial scarcity premium dissipates.
Federal Reserve Chairman Kevin Warsh's observation that AI presents "AI-driven inflation risk" — flagged to his colleagues as a concern even as he sees no current wage pressure — points to a different dynamic that supports the investment case. AI is already visible in energy demand, in data centre construction activity, in semiconductor purchasing patterns. Its deflationary productivity effects, by contrast, have been more modest and more concentrated than the most optimistic projections anticipated. The Fed is seeing the cost side of AI infrastructure long before it is seeing the productivity dividend.
For enterprise buyers, the adoption calculus is shifting. The first generation of AI adoption was characterised by fear of falling behind — companies bought AI tools and capabilities defensively, to avoid being disrupted by competitors who moved faster. The second generation of AI adoption, which the market is now entering, will be characterised by much harder questions about which AI applications are generating measurable business value and at what cost per unit of output improvement.
Where AI ROI Is Proven
The industries where AI ROI has been most clearly demonstrated — pharmaceutical drug discovery, software development, customer service automation, fraud detection in financial services — are also the industries where adoption is now accelerating regardless of macroeconomic conditions. The industries where AI ROI is least clear — complex strategic analysis, creative production at quality thresholds that matter commercially, regulatory and compliance decision-making — are the ones where experimental spending will slow most visibly in the second half of 2026.
What happens when the ROI answer disappoints is not that AI investment stops. It is that investment becomes more selective, more disciplined, and more concentrated in applications where value is empirically established. For the semiconductor and infrastructure supply chain, that selectivity means a market defined by application-specific demand rather than the broad buy-everything approach that characterised the AI spending boom through mid-2026. That distinction matters for every company whose revenue forecast assumes the AI capex cycle will continue indefinitely at current rates.
The supercycle in AI infrastructure spending is not over. What is changing is the composition of demand within it. The broad platform spending that characterised 2023 through mid-2026 — buying everything, deploying widely, optimising later — is giving way to a more disciplined approach that favours specific applications with measurable returns over speculative capability building. For AI vendors, that shift means revenue quality improves even if headline growth moderates. For the semiconductor and infrastructure supply chain, it means a market defined by application-specific demand rather than the buy-everything approach that created the extraordinary valuations the market is now correcting from.
What a More Disciplined Market Looks Like
The distinction between AI that generates measurable operational improvement and AI that serves primarily as a signalling device — communicating to investors, customers, and employees that a company is modern and technology-forward — is one that the more disciplined second generation of AI investment will force companies to make. The first generation of AI adoption was forgiving of this distinction. The second will not be.
For enterprise software vendors building AI-enhanced products, the shift toward ROI-focused adoption is actually a clarifying development. The customers who were not generating returns from AI implementations were also the customers most likely to churn, reduce usage, or complain about pricing. A customer base that has gone through the discipline of identifying where AI generates value — and is buying specifically for those applications — is a more stable and ultimately more valuable revenue base than one driven by FOMO purchasing.
For the broader economy, the question of whether AI delivers its promised productivity dividend at a scale that shows up in macroeconomic data remains open. Federal Reserve Chairman Kevin Warsh's observation about AI-driven inflation risk — the investment boom is visible and measurable, the productivity payoff less so — captures the current position accurately. The technology is real, the applications are multiplying, and the evidence that it is delivering broad-based productivity growth at the scale required to justify the infrastructure investment remains, for now, more promise than proof.