Microsoft Eyes $600M Savings by Swapping OpenAI for Moonshot’s K3

Moonshot AI’s Kimi K3 dropped July 16 as the largest open-weight model yet, and it took about 72 hours to reignite a policy war that goes straight to the bottom line of every enterprise AI buyer. Dean W. Ball, OpenAI’s head of strategic futures and a former Trump White House AI adviser, called it “a very good model” whose performance probably isn’t just lifted from someone else’s work—but he also noted it’s “very token hungry” and perhaps not the bargain its $15 per million output tokens suggests, especially since maximum reasoning effort is the only serving mode right now.

What really lit the fuse was Ball’s prediction that the Trump administration will eventually settle on soft regulatory pressure against Chinese open-weight models. Not an outright ban—he called that “one of the dumber motifs in AI policy”—but agency guidance hinting at backdoors, enough to make regulated enterprises flinch without any formal prohibition. “It needn’t be that well justified,” he wrote, which David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, immediately flagged as either a confession or a forecast of regulatory capture, and unacceptable either way.

The commercial arithmetic is what makes this more than a D.C. shouting match. Open-weight models ate 29% of tokens through Vercel’s production gateway in June, up from roughly 11% in April, while accounting for under 4% of total spending. That compression is now hitting the hyperscalers directly: The Information reports Microsoft is adding K3 to Azure and actively evaluating whether it can replace OpenAI and Anthropic models for some Copilot features, with potential inference savings of up to $600 million. GitHub already made Moonshot’s Kimi K2.7 Code available in Copilot’s model picker on July 1, hosted on Azure, so the plumbing is already in place.

Microsoft hasn’t confirmed the $600 million figure or which features are under review, and an evaluation isn’t a deployment. But when your largest customer of both leading American frontier labs starts pricing the alternative, the competitive threat stops being theoretical. Axios reports that Commerce, the NSA, and the Office of the National Cyber Director were already weighing Entity List designations for Chinese AI labs last year, so Ball’s prediction may simply be catching up to what’s already in motion.