Wall Street calls it demand. It isn't.

The largest pure-play data-center operator set IPO terms this week at a valuation near $100 billion — bigger than most of the tenants renting its space. That number matters because capital is still pouring into the physical layer of AI faster than the revenue meant to justify it. Watch the debt covenants in the filing, not the headline: much of the buildout is funded with borrowed money, and the interest bill assumes demand keeps compounding.

Allocation for Nvidia's next accelerator is already spoken for through mid-2027, per supplier checks — the chips are sold before they ship. But the real constraint has quietly moved from silicon to power: several announced data centers can't secure grid connections fast enough to plug the chips in. Utilities, not fabs, are now the bottleneck — and that's a far harder problem to throw money at.

A startup barely a year old raised $2 billion at a $12 billion valuation — with no public product and a revenue run-rate built largely on other AI companies' spending. A raise like this would have taken a decade before 2023. The real signal isn't the valuation; it's how far the funding cycle has compressed, and how much of this market's revenue is one AI company paying another.

The latest Fed minutes softened the language on inflation and hardened it on labor-market risk — the clearest hint yet that another cut is coming. Futures repriced within minutes, pushing rate-cut odds for the next meeting sharply higher. It matters for this whole issue: cheaper money is part of why equities and crypto both caught a bid, and why the AI buildout keeps finding financing.

A spot crypto ETF crossed $50 billion in assets, and the flows are increasingly institutional rather than retail — pensions and advisors allocating through the same wrappers they use for stocks. Crypto's plumbing (custody, ETFs, prime brokers) now looks a lot like traditional finance's. Which is exactly why the next drawdown is likely to behave like a risk-asset selloff, not a hedge.

A Big Tech name agreed to buy an AI-video lab for a reported $6 billion — a price that used to buy whole public companies, now paid mostly for talent and model weights. It's the acqui-hire playbook at industrial scale. The tell: the market is pricing AI capability like infrastructure you have to own, not software you can license — and the giants would rather buy the team than compete with it.

Earnings split the screen this week: a big bank beat on the back of higher-for-longer rates, while a hardware name warned on margins and guided down. Same macro, opposite outcomes. The lesson is that the AI trade isn't one trade — the gap between who benefits and who gets squeezed is widening, and index-level moves are hiding it.