AI hardware: Slowing down? Not so fast.

A few weeks ago, a consumer app gave semiconductor stocks a fresh boost.
Meta’s new personal AI agent, Muse, rapidly climbed the US app rankings, and semiconductor stocks rallied as investors started to imagine what successful consumer adoption of agentic AI could mean for compute demand.

The timing was particularly interesting because, only a few days earlier, an apparently contradictory narrative had emerged. Leading AI researchers had started openly discussing whether the industry should slow the pace at which increasingly powerful frontier models are developed.

The contrast perfectly illustrates the noise investors are currently navigating. Rather than reacting to every headline, we think it is more useful to return to the underlying hardware thesis: what could actually break it and what might the market still be underestimating?

After a spectacular run for hardware, two risks dominate that debate today.

Recent calls from leading AI researchers to slow the development of ever-more-powerful models were initially interpreted as bad news for semiconductors. The logic seems straightforward: if the time between frontier-model generations lengthens, fewer enormous training runs should be required and therefore less compute.

But that is only one side of the equation. A greater emphasis on AI safety could itself become remarkably compute intensive, as models need to be tested, monitored, evaluated and potentially supervised by other AI systems. AI watching AI sounds slightly dystopian, but it also requires a lot of tokens, and therefore a lot of compute. At the same time, stricter safety requirements could favour the largest frontier labs, which are better able to absorb the fixed costs of testing, auditing and compliance, potentially raising barriers to entry.

Most importantly, a slower pace of progress at the absolute frontier could change where compute is being used rather than materially reduce how much is needed. Resources could shift from continuously training the next generation of models towards testing, deploying and running the models that already exist, supporting more inference, more agents and potentially much broader adoption.

This is probably the more important question. The scale of the investment cycle has become extraordinary, with hyperscaler and cloud capex forecast to exceed a record c.$1.2tn by 2027, compared with less than $200bn in 2023². At the same time, the financial debate is becoming more demanding: free-cash-flow profiles are under pressure, financing costs have increased and investors increasingly want evidence that hundreds of billions of dollars of investment can ultimately generate adequate returns.

The current pace of capex growth clearly cannot continue indefinitely. But there is an important difference between capex growth slowing and capex actually falling. Even a significant deceleration would leave absolute spending at historically high levels and continue to represent substantial incremental demand for the companies supplying the infrastructure. Importantly, we are still largely in the input phase of the AI cycle: infrastructure is being built before the full economic output of AI has materialised. This timing mismatch makes today’s capex look particularly heavy relative to revenues, but it may also be precisely what is needed to enable the next phase of adoption.

Competitive dynamics also make a sharp retrenchment more difficult. Cloud, search, advertising, productivity software and consumer ecosystems are all being reshaped by AI. Underinvesting may protect free cash flow today, but it also risks leaving a platform technologically behind tomorrow. This creates something close to an “AI or die” dynamic: hyperscalers will become more selective about where they spend, but the hurdle for materially cutting investment remains high.

Currently, the main brake on the cycle may increasingly be physical constraints rather than willingness to invest.

The risks are real, but they are only half of the story. At the same time, several developments suggest that the compute opportunity is broadening well beyond the original AI build-out.

The first phase of generative AI was relatively straightforward from a hardware perspective: train increasingly powerful models on enormous GPU clusters and then use those models to answer prompts. Agentic AI changes that equation, and Muse provides a useful illustration.

Instead of simply generating an answer, an agent can operate browsers, interact with applications and carry out a sequence of tasks on behalf of a user, potentially continuing to work in the background for minutes or even hours. GPUs and AI accelerators still perform the heavy inference work, but CPUs orchestrate the operating systems, browsers and potentially hundreds of smaller actions around them. Memory, networking, storage, power and cooling all benefit as those workloads become more complex.

In other words, agents broaden the AI hardware bill of materials. Muse is only an early example. But it gives the market something it has been waiting for: a tangible consumer use case for more inference, for longer, across a broader range of hardware.

Today, supply constraints are appearing across substrates, speciality materials, optical connectivity, power management, cooling and multiple parts of the semiconductor equipment chain.

Nvidia’s latest earnings provided a striking illustration. The company expects revenue to grow by around 70% next fiscal year, far above previous market expectations³. But revenue forecasts can rise with a few changes to a spreadsheet; the physical supply chain cannot adjust nearly as quickly. Factories take time to build, specialised components remain difficult to produce and qualification cycles can last months or years.

As one constraint is solved, another tends to emerge: more GPUs require more memory; more memory requires additional packaging capacity; larger clusters need faster networking; greater density creates new cooling and power requirements. The bottleneck keeps moving further along the value chain. A company does not need to manufacture the most glamorous AI chip to capture attractive economics.

One final distinction remains underappreciated. The market keeps asking whether hyperscalers will ultimately generate enough AI revenue to justify the hundreds of billions of dollars they are investing. Their spending happens today, while much of the monetisation comes later.

For many hardware companies, the equation is simpler: their customers are spending today. Orders translate into revenues, revenues into earnings and, for the companies we favour, earnings into free cash flow. Combined with bottlenecks and long lead times, this can give selected hardware companies greater earnings visibility than the companies financing the AI build-out.

Valuation matters just as much. Despite the strong rally in semiconductors, we continue to find opportunities where expected free-cash-flow growth remains attractive relative to the valuation paid. This contrasts with parts of the hyperscaler universe, where capital intensity is rising just as investors are becoming more demanding about the timing and profitability of future AI revenues.

None of this means buying hardware indiscriminately. Expectations are high, technological disruption remains a constant risk and capacity will eventually catch up with demand. Stock selection and valuation discipline therefore matter more, not less.

But the broader conclusion remains unchanged. The two major risks facing AI hardware today are real: the pace of frontier-model development could slow, and hyperscaler capex growth will eventually moderate. Yet neither necessarily implies a collapse in compute demand, while the next phase of AI could broaden that demand across more workloads, more components and more parts of the supply chain.
For hardware, we do not think we are there yet.