The AI-Gets-Cheaper Assumption Is the Weak Joint

The ROI model behind most enterprise AI programs has a soft joint nobody stress-tests: it assumes the per-unit cost of inference keeps falling, forever, on a smooth curve. That assumption is doing a lot of load-bearing work, and it's the first thing I'd pull on.

The physical layer is pushing the other way from several directions at once. A senior TSMC executive, in a rare interview, "does not rule out price rises" on the chips everything runs on. Read that as what it is. A dominant supplier with real pricing power signaling pricing power, not a committed schedule. But you don't build a three-year business case on a vendor's reassurance, and "the costs will come down" was always a vendor reassurance.

Then Seattle, home to two of the biggest cloud operators, passes a year-long moratorium on new AI datacenters. Maybe that's local politics and stays local. Maybe it's the first municipality to say the energy draw isn't worth it, and others copy the homework. I genuinely don't know which, and that uncertainty is exactly the point. Constrained supply doesn't get cheaper.

"AI gets cheaper every year" is a forecast wearing the costume of a fact.

The comforting story is that Wall Street will sort this out once the big labs go public, that the market will force the cost structure into the open. Public markets spent years failing to read cloud unit economics after those IPOs. There's no reason to think they'll price AI infrastructure cost honestly out of the gate either.

So the discipline is the same one that always separates a real case from a hopeful one. "Three days to fifteen minutes" survives a chip price hike. "We're transforming with AI" doesn't survive anything. Run your vendor economics at flat or rising infrastructure cost. If the case only works on the discount, it isn't a case.