The supplier cut the price of intelligence twice, then bought the machine that etches it into silicon.
OpenAI cut its cheapest frontier model roughly eighty percent, Meta shipped a coding model built to undercut, and AMD bought a startup that burns model weights into chips. The floor keeps dropping on purpose.
While the consulting firms were learning to count the cost of AI, the companies that set that cost spent the week driving it down. On the thirtieth of July OpenAI cut the price of its cheapest frontier model by roughly eighty percent, dropping it into the low-cost tier occupied by open and Chinese models. Days later Meta shipped a coding model positioned explicitly on price against the incumbents, its leadership saying in as many words that the product would compete less on the frontier of capability than on cost. Then, on the sixth of August, AMD acquired Taalas, a startup that etches model weights directly into silicon, a bet that inference can be made dramatically cheaper by turning a model into a chip rather than running it on a general-purpose one.
Three moves, one direction. Each is a supplier deciding that the way to win is to make a unit of intelligence cost less, and the effect on the buyer is a model price that is not merely falling but falling on a schedule set by someone else. This is the fact that sits underneath the whole cost-discipline story the firms told this week. The reason token spend is so hard to govern is not only that workforces consume more of it, as Accenture correctly noted. It is that the unit price is a moving target controlled by the labs, and a governance practice tuned to this month's prices is stale by the next earnings call.
For a buyer, a falling model price is good news and a trap at the same time. It is good news because the deterministic work gets cheaper. It is a trap because it invites you to build your position on the model itself, the one component whose price and availability you do not control and whose replacement is now announced every few weeks. When the thing at the center of your architecture is being commoditized on purpose by the people who sell it, the center is the wrong place to plant.
The move is to plant one level up. A switch that routes each request to whichever model is cheapest and defensible this quarter turns the price war from a threat into a tailwind, because every cut the labs make flows straight to your bill without a renegotiation. And the deterministic procedure that rides above the switch does not care what a token costs at all, because it does not spend one. Own those two layers and the labs can cut prices every week; each cut is a gift, not a disruption.
The firms spent the week selling the discipline of watching a price they cannot control. The suppliers spent the same week proving how fast that price can move. Put the two together and the conclusion writes itself. Do not anchor the business on the model. Own the switch above it and the procedure above that, and let the supplier race the price to zero on your behalf.