Coinbase's Downloaded Weights Outrun the Entity List
The Weights Are Home
Coinbase switched its engineers over to two Chinese open-weight AI models this month, and the switch says more about deployment economics than about any benchmark score, the argument this desk has made for over a year: procurement decisions, not lab benchmarks, now decide who wins AI adoption. Coinbase CEO Brian Armstrong disclosed the exchange now defaults engineers to GLM 5.2, built by Zhipu AI, and Kimi K2.7, built by Moonshot AI. Internal AI spend is down roughly 50 percent, even as engineers used the tools more, not less. The mechanism behind that saving is the real story. Coinbase is not dialing a server in China every time an engineer asks a question. It downloaded the model weights, the actual trained parameters that make GLM 5.2 and Kimi K2.7 run, once, and now operates them on its own machines. No query ever crosses a border, because there is no border left to cross. A routing system built on top of that setup pushed Coinbase's cache hit rate, the share of repeat queries answered from memory instead of a fresh, costly inference call, from 5 percent to 60 percent. That is an infrastructure decision, not a model-quality one, and it is the one the Entity List's sales restriction on Zhipu was never written to cover, because it stops US firms from selling to Zhipu, not from running weights a company already owns.
A Ban With No Target
Moonshot released Kimi K3, a 2.8 trillion parameter model, the largest open-weight release to date. Axios reported the Trump administration is reviving Entity List designations, procurement bans, and liability rules for Chinese AI models. White House science adviser Michael Kratsios has accused Zhipu and Moonshot of running industrial-scale campaigns to distill the coding and reasoning performance of US models into their own. For that claim to justify the restrictions being drafted, the distilling would have to require an ongoing connection back to the lab being copied. It does not: a lab distills performance once, at release, and every use after is just running a file someone already owns. AI adviser David Sacks countered that a ban would entrench a duopoly and cede the open-weight contest outright. One detail outlasts both arguments. Zhipu AI has been on the Entity List since January 2025 for its role in Chinese military AI development, a rule that stops US firms from selling to Zhipu. Coinbase is running Zhipu's own weights anyway, because the Entity List governs sales, not a file a company already owns.
Coinbase is a regulated public company that already finished this fight before Washington opened it. The next question is not whether the coming Entity List rule survives the Kratsios-Sacks split inside the administration. It is how many other regulated US firms, banks and brokerages, are quietly running the same download-once playbook, sitting on Chinese model weights right now, before the revived Entity List rule Axios reported is even finalized.
Sources
- Coinbase's shift to GLM 5.2 and Kimi, cache-rate and spend detail (the-decoder.com)
- Kimi K3 launch, 2.8T parameters, open-weight release (venturebeat.com)
- Trump administration weighing Chinese AI model restrictions, Kratsios/Sacks split (fastcompany.com)
- Administration reviving ban push after Kimi K3, enforcement difficulty of open weights (tomshardware.com)