At the beginning of the year the hyperscalers were all announcing huge capex spending, reaching nearly $700bn, up about 36% on last year. Most of this expenditure is tied to AI infrastructure. The market took a rather sceptical stance.
Big numbers always sound impressive, but will we really see the spending? In December Anthropic CEO Dario Amodei told The New York Times DealBook Summit: “There are some players who are ‘YOLO-ing’, who pull the risk dial too far, and I’m very concerned.” He was speaking directly about concerns of overspending on AI infrastructure.
For clarity on the issue, the market awaited results from the Taiwan Semiconductor Manufacturing Company (TSMC), which manufactures most of the chips being used in AI infrastructure (about nine out of 10 advanced AI accelerators are made by TSMC).
It came: in the first quarter, TSMC’s revenue was up 35% year on year. In March there was an acceleration of 45.2% in year-on-year revenue growth.
So the spending is happening. Even the cautious Amodei is now splurging, with $50bn committed to AI infrastructure in Texas and New York.
The next big question is whether the demand is there. That’s easy: yes, it is. Everybody is scrambling for more. Even part-time users (like me) are experiencing more throttling. Over the past few weeks I’ve seen new usage constraints clearly designed to get people to spread usage out over time rather than all at once. I have also been seeing time-out errors — also a new phenomenon.
In mid-2024 Microsoft CFO Amy Hood decided it was spending too much on new AI data centres and needed to scale back. At the time, that decision was applauded as sensible. But now Microsoft is scrambling as it has an Azure backlog of $80bn and the new problem is power and infrastructure, not chips.
This brings us to the real question: are clients spending enough to make the numbers make sense?
Anthropic is certainly seeing huge revenue growth. The group did its first $100m of revenue in 2023. Then in 2024 it got to $1bn and last year it hit $10bn. The initial 2026 forecast was upped to $18bn and now reports suggest it could be $30bn.
It’s all about use case — whether it’s vibe-coding the next software-as-a-service app or asking Claude how your stock portfolio did overnight. Users need to see real-world benefits. I have a simple red flag I am watching out for: the first nontech CEO of a large company who says at a results announcement that they’re cutting back on AI spend as they’re not seeing real benefits.
For now, though, all the pieces still fit into place. Sure, we may have a bubble, but we also have real spending supporting real infrastructure.