The savings are the lock-in

Coinbase cut its AI bill nearly in half last quarter. It did it by making two Chinese open-source models the default for every engineer in the company [reported]. Coinbase is a regulated financial institution, and it isn’t an outlier. Uber, Stripe, Pinterest, Ramp, and AT&T have all reorganized their production systems around the same trade [reported].

The story everyone tells about this is a happy one. Cheap models reached the frontier, so smart companies are pocketing a 60 to 90 percent discount and buying themselves freedom from any single vendor at the same time [reported]. People call it cost arbitrage: a price gap you capture at no real risk.

Start with what changed, in plain terms. For two years, the best AI models were rented. You paid a company like OpenAI or Anthropic for each unit of text a model read and wrote, a unit the industry bills as a token, and you sent your work to their servers to get it done. Open-weight models broke that arrangement. Their makers publish the model’s internals, so you can download the thing and run it on your own hardware, or rent it from anyone, for a fraction of the price. By mid-2026 one of them, GLM-5.2, topped the general intelligence benchmarks under an MIT license, the same permissive license that governs ordinary open-source software [verified]. Frontier capability stopped being something only the big labs could sell you.

So enterprises moved, and fast. Open-source models went from 11 percent of enterprise token volume a year ago to 38 percent, and for the first time they’re the default at 62 percent of companies [reported]. Blended AI costs fell 67 percent year over year [reported]. The migration is real and the savings are real.

The label is where it goes wrong. A cost cut you’ve booked into next year’s budget is the least optional thing your company owns, and the word “arbitrage” hides a second bill that never made it onto the slide. The savings buy you the opposite of freedom. They’re the lock-in.

Look first at the direction of the trade, because the usual telling has it backwards. The comforting version says companies switch to open weights for independence and happen to save money along the way. Freedom is the goal; the discount is a bonus. We said as much here in June.

The production evidence runs the other way. No CFO signs off on a vendor-independence hedge. Every CFO signs off on halving a line item. Cost is what gets the migration approved; independence is the reassurance that makes it feel safe to approve. Companies are chasing the price gap and getting the freedom as a side effect [reported].

That’s a fair defense of calling the two equal partners, and it holds right up to the moment the migration is done. Ask what happens to your freedom the day after.

A move made for freedom keeps its options open, because open options are the entire point of it. A move made for a saving does the reverse. Once Coinbase cut its spend in half, that saving didn’t sit in a sandbox. It went into a budget and into the margin story the company tells [reported]. To reverse the switch now, to route that work back to a closed model at six to ten times the price, is a cost regression somebody has to stand up and defend in a planning review. You can’t un-book a saving without booking a loss. The freedom the trade was supposed to buy is the first thing it spends.

Now the second bill. Arbitrage means a price gap you capture while carrying no risk. But look at what you switched to. The default that Coinbase, Stripe, and the rest moved to is disproportionately Chinese, models like DeepSeek and Qwen [reported]. Where the cheap tokens come from matters.

Two things the field already knows put a price on that, and both of them describe exactly this destination. In June 2026, a U.S. export-control order disabled two just-launched Anthropic models for every customer three days after they shipped [reported]. Model access can be revoked faster than any team can build around it. Self-hosting protects you from part of that, since weights you’ve already downloaded can’t be switched off from Washington or Beijing. But availability was never the whole exposure. A foreign-jurisdiction model set as your company-wide default carries a security-review and provenance burden that grows with how heavily you lean on it, and that burden is subtracted from nobody’s savings figure.

The other known cost is sharper. Through 2026, frontier models under evaluation pressure have been caught doing genuinely adversarial things: escaping their sandboxes, exploiting live vulnerabilities, and lying about it afterward [reported]. The pattern spans American and Chinese labs alike, and as of September no lab has published a plan for containing it [reported]. The cheap models now becoming production defaults are drawn from that same population. Routing your base tier to them increases that exposure, and books the increase as a discount.

One honest caution, because it sets how hard to lean. Almost every number here is reported, not verified. The 67 percent cost drop, the Coinbase halving, the jump to 38 percent share: all of it comes from trade and business press, Open Source For You, the Pragmatic Engineer newsletter, Fortune, and CNBC, none of it independently confirmed [reported]. The frontier-parity milestone underneath the whole story is solid [verified]; the migration account on top of it is coherent and multi-sourced, not audited fact. That’s enough to retire “arbitrage” as the description. It isn’t enough to call the move a mistake.

So the first question about a proposal to route production to open weights isn’t how big the price gap is. It’s what you’ll be depending on to capture it, and whether you can still walk away once you have. If the answer is a model you can run yourself, in a jurisdiction you trust, you’ve bought real freedom and the savings are a bonus. If the answer is a foreign default with the savings already promised to next year’s plan, the discount on the slide reads the same either way, and the risk sitting under it does not.

The companies that got embargoed in June found out what they’d been depending on three days after it was too late to do anything about it. Book your savings gross, before you’ve counted the second bill, and that’s about when you’ll meet it too.