Open-weight models do the tricks, the dollar pockets the money

An open-weight model publishes the weights obtained through training so that other developers can fine-tune it, while a frontier model keeps its design closed and charges for the finished product. China's Kimi K3 and Qwen, and America's NVIDIA and Meta, choosing open weights is less about US–China rivalry than an ecosystem differentiation strategy by latecomers that haven't reached the top tier; the evidence offered is that Qwen, with 2 billion downloads on Hugging Face in the first half of this year, far outpaced Meta (200 million) and Google (400 million). On the other hand, the 165 million agent payments on Coinbase's x402 protocol are not purchases by people but small fees for API calls between AIs, totaling only $50 million, and 99% of them are made in USDC. The heart of it is the paradox that developers use Chinese open-weight models while the payments flow in US-dollar stablecoins, and, together with CME's listing of GPU rental-price futures, the trend of compute turning into a financial asset is also worth watching. Figures such as 51% of top developers being Chinese are remarks without a primary source and need to be checked.

This is a note: my own write-up after reading several sources. It has no source link; events checked against a source are kept as records.