Coinbase Cuts AI Spending by Nearly Half, Shifts to Chinese Open-Source Models to Control Costs
Taylor Wilson
Coinbase CEO Brian Armstrong disclosed that the company slashed AI spending by nearly half while AI usage kept growing — the key move was switching default models to Chinese open-source alternatives, signaling that Chinese open-source AI is becoming a real cost lever for US companies.
How much did they save — and how?
Coinbase switched its default AI models to Chinese open-weight models and cut AI spending by nearly half, even as usage continued to rise.
This means → the savings came not from using less AI, but from using cheaper models — volume up, bill down.
In plain terms = same workload, different tool, invoice cut in half.
Which Chinese models are they using?
Armstrong named two models already in trial: GLM 5.2 from Zhipu AI's Z.ai, and Kimi 2.7 from Beijing-based Moonshot.
Both are open-weight models — the parameters are public, so a company can self-host instead of paying per API call.
This reflects a broader shift: as US frontier AI services get more expensive, Chinese open-source models are moving from benchmarking curiosities to actual production alternatives inside American firms.
What else did they do beyond switching models?
Armstrong also introduced an internal mechanism: making each engineer's actual token consumption visible across the company.
In plain terms = use as much AI as you want, but everyone sees your bill — and the more you spend, the higher the output bar the company sets for you.
This means → Coinbase's cost playbook is a two-lever approach — model selection plus cost transparency — not a blunt budget cut.
What does this mean for the market?
This case shows that cutting AI spending and expanding AI usage are not contradictory — the key variables are model choice and cost visibility.
Whether Chinese open-source models can replicate this path at more US companies will be a critical test of their commercial penetration.
This reflects a shift in AI competition: the next phase is not just "whose model is strongest" but "whose model is cheapest and easiest for enterprises to actually deploy."
Content is for reference only, not financial advice.