I think there are two different arguments being mixed together here: "there is an AI investment bubble" and "AI companies are all going to collapse". The first is entirely arguable. The second isn't supported by the current figures.
There is unquestionably a huge amount of money going into AI infrastructure. NVIDIA alone reported $215.9bn revenue for FY2026, up 65% year-on-year, with data-centre revenue of $62.3bn in its latest quarter. This isn't a company selling a product nobody wants.
OpenAI is also hardly operating on the basis of "one prompt = one loss". It raised $122bn in committed capital at an $852bn valuation in March. That does not prove it is profitable — it isn't — but it rather undermines the claim that it has "no route" to financial viability.
Anthropic has the same issue: frontier AI is extremely expensive, so profitability is a legitimate question, but "therefore it has no viable business model" doesn't follow.
Oracle is an even better example of why this needs nuance.
Oracle's FY2026 revenue was about $67.4bn, and its remaining performance obligations — essentially contracted future revenue — reached $638bn. Its cloud infrastructure revenue was growing extremely rapidly.
That doesn't mean Oracle is risk-free. Quite the opposite: it is spending enormous sums building AI infrastructure, and the financing requirements are substantial.
But going from:
"Oracle has taken on significant AI infrastructure risk"
to:
"Oracle will likely fail"
is a very large leap.
CoreWeave is probably the best example of the genuine risk.
It reported $2.58bn revenue in Q2 2026, more than double the previous year, and its contracted backlog was about $104bn. At the same time, it made a $626m net loss and has extremely heavy capital and financing requirements.
So yes — there absolutely IS an AI infrastructure bubble risk.
If AI demand suddenly fell, companies like CoreWeave could get hurt very badly because they have spent huge amounts of money on specialised infrastructure on the assumption that customers will keep buying computing capacity.
But that's very different from saying:
"Nobody believes these data centres will ever be built."
People clearly do believe the demand exists — there are already billions of dollars of actual revenue and hundreds of billions of dollars of contracted commitments.
The sensible bearish argument is therefore:
What happens if the expected future AI revenue doesn't justify the enormous capital expenditure being made today?
That is a serious question.
It could mean:
- data-centre projects delayed or cancelled;
- GPU prices falling;
- AI-company valuations collapsing;
- infrastructure companies suffering losses;
- investors writing down assets;
- banks and private lenders becoming more cautious;
- layoffs in AI and related construction/infrastructure;
- semiconductor demand falling;
- and potentially cheaper RAM, SSDs and GPUs after the supply cycle adjusts.
But none of that requires "the entire tech industry collapses".
Microsoft, Amazon, Google and Meta are enormous diversified companies with huge existing businesses and cash flows. An AI investment disaster could cause enormous losses and write-downs without causing those companies to disappear.
And Meta "going under" because of AI spending is, frankly, an extraordinary claim. It would require an enormous deterioration in its existing advertising business as well as failure of its AI investments.
The same applies to NVIDIA.
NVIDIA hasn't somehow forgotten how to make graphics cards because of AI. It has discovered that the same underlying GPU architecture is extremely valuable for AI and other accelerated-computing workloads. Its $215.9bn FY2026 revenue tells you how successful that transition has been.
And on the RAM/SSD point: AI infrastructure absolutely does consume enormous quantities of memory and storage, so an AI investment slowdown could eventually relieve some pressure. But semiconductor prices are cyclical anyway, so it isn't simply "AI disappears → RAM becomes cheap".
The labour point is actually much more interesting.
There are companies that have cut employees on the assumption that AI could replace parts of their jobs and subsequently discovered that human expertise was still required. That's a genuine phenomenon.
But again, it doesn't prove AI is useless. It proves that automating individual tasks is not the same as successfully automating an entire job or organisation.
And finally, the "you had to use AI to formulate your reply" argument isn't actually an argument.
If I use Excel to analyse a spreadsheet, you don't disprove the spreadsheet by saying "you used Excel".
If I use a database to follow 500 interconnected pieces of information, you don't disprove the conclusion by saying "you used a database".
Likewise, if someone uses AI to help analyse a complicated chain of economic consequences, the relevant question is whether the analysis and evidence are correct, not whether they typed it themselves.
So I'd put it this way:
There is a perfectly respectable argument that AI investment has become excessive and that a major correction could be coming.
There is not currently a respectable evidential basis for saying OpenAI has no route to viability, Oracle is likely to fail, CoreWeave will abandon its infrastructure, Meta is likely to collapse, or the entire technology industry will go under.
The technology can be real and the investment bubble can be real at the same time.
That's actually what happened with the internet.
The internet didn't turn out to be fake because the dot-com bubble burst. A lot of investors simply paid far too much for the future they expected.