CTOs may face cheaper AI capacity as Baillie Gifford warns of overheating

The prevailing wisdom suggests that massive investment in AI infrastructure is a rational response to insatiable demand. However, this view overlooks a lurking risk. Within 18 months, the current pace of AI infrastructure development will likely result in systemic overcapacity, requiring a fundamental reallocation of capital from hardware into specialized, high-value AI software.

Baillie Gifford narrows the bubble claim to infrastructure

According to Baillie Gifford’s [official research](https://www.bailliegifford.com/en/usa/institutional-investor/insights/ic-article/2025-q4-us-perspectives-are-we-in-an-ai-bubble-10059067), the US market is not being described as a broad bubble. The firm’s narrower claim, as summarized in the reporting packet, is that AI-driven infrastructure build-outs show signs of overheating while current market premiums remain anchored in superior profitability.

That framing matters because it separates the software adoption story from the physical capacity story: an enterprise can keep buying AI tools even if the market has overpaid for the infrastructure needed to run them.

The one core idea here is compute scarcity. In the present market, limited access to AI infrastructure gives suppliers pricing power and pushes enterprise buyers toward advance commitments.

Baillie Gifford’s warning implies a different future: if capacity is built faster than profitable workplace use cases mature, scarcity can become slack, and the buyer’s problem shifts from getting enough capacity to refusing the wrong kind of capacity. That is a procurement and finance problem before it is a model-performance problem.

The popular read mistakes scarcity for durable demand

The consensus reading Baillie Gifford is pushing against is that large AI infrastructure spending is simply the rational cost of meeting durable demand. In that version, high valuations follow from future growth, and the concentration of spending is a feature of the market’s best companies rather than a warning sign.

The weak point in that story is timing: infrastructure is committed before many corporate AI projects have shown whether they reduce labor needs, raise revenue, or survive renewal scrutiny.

That timing mismatch is where the mispriced risk sits.

If the build-out is partly driven by fear of missing out, the market can confuse booked capacity with productive capacity. The future-of-work consequence is not a clean wave of automation; it is a messier repricing of AI projects inside companies. A chief AI officer who could justify broad experimentation when compute felt scarce may face a different argument when capacity is plentiful: why is this particular tool, workflow, or internal model worth keeping when the underlying infrastructure is no longer the constrained asset?

The missing numbers are the story

Baillie Gifford’s public framing, as provided here, does not give the numbers executives would need to separate a justified build-out from overheating. It does not state utilization rates, contract lengths, customer concentration, application-level returns, or whether the most profitable demand is tied to one-off model training commitments or recurring inference usage.

The baseline is also unclear: overheating measured against what prior capacity cycle, on what class of infrastructure, and with what evidence that demand will fail to fill supply? The concrete limitation is blunt: this is a market argument, not a capacity audit.

That omission does not make the concern irrelevant. It changes how it should be used. A board should not read Baillie Gifford’s piece as proof that AI infrastructure will collapse; the source does not establish that. It should read it as a warning that profitability at the market level can mask a procurement trap at the company level, where executives buy long-duration access to capacity without knowing which workflows will still matter by renewal.

Overcapacity would move power from infrastructure sellers to software buyers The second-order effect is easy to miss because the public debate usually stops at chip and data center demand. If infrastructure overheats, the next pricing battle moves closer to the workplace: legal review systems, sales assistants, coding copilots, clinical documentation tools, factory-planning software, and internal knowledge systems will have to show that their value comes from domain fit rather than privileged access to scarce compute.

That favors niche AI software that can prove a specific workflow outcome over general infrastructure exposure.

For enterprise buyers, this would change the internal politics of AI budgets. The first phase rewarded teams that could secure access and move fast; a looser capacity market rewards teams that can prove retention, avoided work, or revenue contribution. That is why the labor implication is subtler than “AI replaces jobs.” The budget pressure lands first on AI program offices, platform teams, and vendors selling broad capability without a narrow owner inside the business.

The counter-read is concentration without collapse

The obvious objection is that Baillie Gifford’s own framing does not say the US market is in a broad bubble. Concentration can reflect genuine profitability, and the companies funding AI infrastructure may have stronger balance sheets and more distribution than past speculative builders.

If demand keeps compounding, today’s build-out may look less like overcapacity than preemption. That counter-read is serious because the source summary itself says market premiums are anchored in superior profitability, not merely narrative enthusiasm.

But that is also why the bubble label is less useful than the capacity question. A market can avoid a broad crash while still producing bad enterprise contracts, stranded internal projects, and vendor margin pressure. The falsifiable version of the concern is narrower: if capacity becomes easier to buy before workplace AI proves durable returns, spending shifts away from generic hardware exposure and toward software that owns a measurable business process.

The near-term signals sit inside budgets, not slogans Over the next 6 months, the useful signals will be less theatrical than market commentary. Watch whether large AI suppliers begin emphasizing utilization and customer returns rather than build-out scale; whether enterprise buyers shorten commitments or demand more flexibility in AI infrastructure contracts; whether vertical AI vendors talk less about model capability and more about gross margins after compute use; and whether internal AI programs survive renewal because a business unit owns the result, not because a central innovation team funded the experiment.

Those signals would show whether Baillie Gifford has identified a genuine mispricing in the compute layer or only a temporary anxiety in an otherwise profitable expansion.

Implications, hedged: if Baillie Gifford’s warning proves directionally right, the future-of-work story in 2026 will not be that companies stop using AI. It will be that the scarce asset moves from compute access to organizational proof — which team owns the workflow, which vendor can defend renewal, and which AI project can survive when capacity is no longer the excuse. That would make AI infrastructure overheating a labor and procurement story hiding inside a market note.