Everyone read Kimi K3 the same way: a Chinese lab had won a lap of the benchmark race, and the American lead was gone. But a striking number of the sharpest reactions this week argued that the benchmark was never the scoreboard that mattered, and that the West is now measuring, and regulating, the wrong contest entirely. Read together, three of them describe a race that looks nothing like the one on the leaderboards.
The first argument is about weight, in the literal sense. Kimi K3 is "open," but the American Enterprise Institute pointed out that it is not exactly downloadable to your laptop. Just to hold its 2.8 trillion parameters in memory and run them at usable quality takes somewhere between 1.4 and 2 terabytes of high-bandwidth memory, the equivalent of a rack of Nvidia's top accelerators or dozens of Mac Studios, a setup costing hundreds of thousands of dollars before you pay for power. Free weights, in other words, push the real cost onto whoever has to serve them. Every organisation that cannot afford its own server rack still reaches K3 the way it reaches ChatGPT: through someone else's data centre, over someone else's API. By that logic the frontier has become a commons, and the contest that remains is industrial, high-bandwidth memory, advanced packaging, data-centre construction, and grids with power to spare. It is a competition in which export controls on the machines China cannot yet build still bite, even as controls on the models plainly do not.
The second argument is about the response. Writing in Forbes, the economist Christian Catalini dismissed the frontier labs' reflexive answer, the accusation that China simply stole its way to parity, as a "cope" that no longer explains the progress. He quoted a bluntly cynical policy recipe making the rounds: do not try to ban open source, just direct every agency to issue soft law that manufactures fear, a bulletin here warning of "backdoors" in Chinese models, enough regulatory risk there that every cautious enterprise quietly backs away. Catalini's counter is that the only real answer to Chinese open weights is American open weights, and that one arrived the same week: Mira Murati's Thinking Machines Lab released Inkling, its first open model, a 975-billion-parameter system trained on 45 trillion tokens and offered for others to fine-tune. Openness, on this view, is not a gift to China but the strategy that historically let America win the internet.
The third argument is about who regulators think they are watching. In a market analysis, the site Truth on the Market noted that China is "missing from the case file." Antitrust authorities in Washington, Brussels and Brasilia have built their AI theories around a supposed Western oligopoly, Google, Amazon, Microsoft, Meta, Apple, plus OpenAI and Anthropic, while Chinese models climbed from roughly 1 percent of the global market in late 2024 to about 15 percent a year later. Alibaba's Qwen family alone has passed 700 million downloads, overtaking Meta's Llama. Microsoft is weighing a fine-tuned DeepSeek model on Azure to cut Copilot costs; one San Francisco firm switched from Anthropic to DeepSeek and says it saved millions. "You don't need God to write your email," its founder shrugged. A regime that polices one side of a race while the other competes freely, the piece warned, may simply hand market share to the players beyond its reach.
The three arguments come from different corners, a defence-policy hawk, a pro-competition economist, a free-market law-and-economics blog, and they do not agree on what America should do. But they converge on what it should stop doing, which is treating a benchmark chart as the final score. On that scoreboard the race is close and getting closer. On the ones that may actually decide it, who can serve intelligence at scale, whose open models set the default, and whether regulators are even looking at the right companies, the game has barely been scored at all.