Join my new list, CaseBliz

AMD up 185% this year. Nvidia worth $5 trillion. Meanwhile, 56% of CEOs say AI has done nothing for their bottom line. Something in these numbers doesn't quite add up.

On 3 September, Nvidia confirmed it would acquire Hugging Face for $12.93 billion. The price tag hides a detail worth sitting with: that exact number, 12,930,300,000, is not an accident. The digits 129303 are the decimal encoding of Unicode point U+1F917. The 🤗 emoji. Jensen Huang paid thirteen billion dollars and slipped a joke into the wire transfer.

The comedy aside, the deal is serious. Hugging Face hosts more than 18 million developers and 3 million models, and Nvidia has already released more than 500 open models on the platform. What Nvidia is buying isn't just a repository. It's the distribution layer where most of the world's open-source AI development actually happens.

The stated rationale is open access. Huang's blog post promised the platform would stay neutral, chip-agnostic, open to everyone. But by owning the hub, Nvidia gains real-time intelligence on which model architectures and workloads are gaining traction before any competitor does. That's not philanthropy. That's the best market research operation in the history of semiconductors.

Hugging Face turned down a $500 million investment from Nvidia last year over concerns about ceding influence to a single dominant backer. Those concerns haven't changed. The ownership structure just has. Regulators in the EU, US, and UK will almost certainly want a close look at whether the company that makes the chips also controlling where most open models live is a concentration problem or a convenience.

That's what the five largest US cloud and AI infrastructure companies plan to spend on capital projects in 2026, nearly double last year's total. It's the number behind AMD hitting a $1 trillion market cap for the first time on 21 September, up 185% this year. When half a trillion dollars is chasing the same supply of chips, even the second-biggest chip supplier becomes worth ten figures.

Chip stocks are at record highs. Infrastructure spending is doubling. And yet according to PwC's survey of 4,454 CEOs across 95 countries, 56% say AI has delivered zero meaningful financial benefit. Not disappointing results. Zero. Only 12% report gains on both cost and revenue.

Those two facts aren't contradictions. They're the same fact viewed from different altitudes. The companies building AI infrastructure are doing very well. The companies buying and deploying it mostly aren't. Gartner expects more than 40% of agentic AI projects to be cancelled by end of 2027. That's not a forecast about technology failing. It's a forecast about enterprises discovering they bought a solution before they had a problem worth solving.

The optimistic read is that the ROI always lags the infrastructure. Roads get built before traffic appears. The less optimistic read: the last time this many companies reported this little return while spending this much was 1999. The dot-com crash didn't invalidate the internet. But it did ruin a lot of people who assumed infrastructure spending was proof of demand.

Re-reading Carlota Perez on technological surges and financial bubbles. She'd find this week's numbers extremely familiar. Not necessarily reassuring.