The sharp rally in AI stocks finally looks like it’s catching its breath. But dig a little deeper and the actual infrastructure story hasn’t lost an ounce of momentum. A new Jefferies report, titled “AI Fatigue and Rotation,” makes the case that money is quietly moving out of pricey AI momentum names and into cheaper value plays. Even so, the report is clear that global spending on AI infrastructure shows no signs of slowing down. The long-term winners, it argues, will keep being the companies building the actual backbone of AI rather than the ones racing to monetise it. The takeaway, in Jefferies’ own words: the AI trade is pausing, not ending.
The momentum is cooling off
Investors have started throwing around the phrase “AI fatigue” after two years of jaw-dropping gains across semiconductor and AI-linked stocks. South Korea makes the point pretty vividly. The Kospi has pulled back hard from its June highs, and leveraged ETFs tied to SK Hynix and Samsung Electronics have taken a beating too.
Jefferies isn’t alarmed by this. A correction after such a long run, it says, is healthy.
What it does warn against is mistaking this rotation for the end of the AI investment cycle altogether.
As the report puts it, so long as the AI capex arms race keeps going, the real beneficiaries will stay the “picks and shovels” companies — the ones supplying chips, servers, networking gear, memory, power infrastructure and data centres, rather than the tech giants actually spending all that money.
Who’s really winning here
There is a striking gap between how AI infrastructure suppliers have performed versus the hyperscalers themselves. Since early 2023, the big four hyperscalers — Alphabet, Amazon, Microsoft and Meta — have gained roughly 180%. Not bad at all. But the three dominant DRAM memory makers, Micron, SK Hynix and Samsung Electronics, have rocketed up about 760% in the same window. That gap alone tells you where the market thinks the real money is being made right now.
Jefferies also flags something worth sitting with: even as demand for AI compute keeps climbing, nobody really knows yet which hyperscaler, if any, will turn these massive investments into genuinely attractive returns. The report goes as far as saying there’s a real chance none of them manage to successfully monetise their AI capex.
The trillion-dollar spending race
Here’s where the numbers get genuinely staggering. Jefferies estimates the four big hyperscalers will collectively spend around $700 billion on AI-related capex in 2026 alone. Next year, that figure could climb past $800 billion. Throw in Oracle, OpenAI, Anthropic and the growing pack of AI cloud providers, and total AI infrastructure investment could blow past $1 trillion by 2027.
To put that in perspective:
- That’s roughly 3% of US GDP
- About 22% of all non-residential fixed investment in the US
- Close to a third of the total pre-tax profits generated by US non-financial companies
Numbers like these explain why demand for semiconductors, data centres, networking hardware, cooling systems and power infrastructure isn’t going anywhere anytime soon.
Earnings season will be the real test
The next big checkpoint for this whole AI narrative arrives with hyperscaler earnings later this month.
Jefferies points out that Alphabet, Microsoft, Amazon and Meta have all raised their AI spending guidance significantly. Microsoft is now looking at roughly $190 billion, while Amazon has held its guidance at around $200 billion. Alphabet and Meta have bumped up their capex plans too.
The question everyone’s really asking: will investors keep rewarding this spending even without matching earnings growth to show for it?
Some cracks are starting to show
Jefferies remains bullish on AI infrastructure overall, but it does not shy away from flagging risk either.
For one, hyperscalers have ramped up debt issuance considerably to fund all this AI buildout — they’ve already raised $169 billion through bond markets this year alone.
There is also a quieter concern building in the background: off-balance-sheet commitments. Citing Moody’s, the report notes that the five biggest US hyperscalers have piled up $662 billion in future data centre lease commitments that haven’t even started yet, pushing total future lease obligations close to $969 billion.
That’s a lot of financial weight sitting behind this boom.
A new, more selective phase
Jefferies does not think the AI story is finished. What it thinks is happening is a shift toward a more selective market. Pure momentum-driven gains in AI stocks may get harder to justify going forward. But companies actually building the physical infrastructure that powers AI still enjoy strong visibility, especially as global investment keeps accelerating.
For India, which is pouring money into semiconductor manufacturing, AI-ready data centres, cloud infrastructure and digital public infrastructure, this report reinforces something that’s been true for a while now: even if AI software valuations swing around, demand for the underlying infrastructure that makes AI possible looks set to stay resilient for years.
