The Frontier Got Crowded
Last week the AI conversation was about the invoice. This week it was about the supplier.
Kimi K3 shipped. Qwen followed a day later. Three shows landed on the same story inside five days, and none of them treated it as a capability story.
Sriram Krishnan, just out of his tenure as the White House senior policy advisor on AI, gave the clearest version on a16z. A few months ago the frontier labs looked like they were pulling permanently away from everyone else. That is over. "If you are in the token consumption business, which I am, it's been a great time." What follows is a margin story, not a benchmark story. More choice puts pricing pressure on the frontier labs, they have to match it, and "it's probably going to erode into their gross margin." The value slides down the stack to whoever has power and GPUs.
Then the question worth sitting with: "Where is the real moat, if you're a frontier lab? Is it in the intelligence, or is it in the harness?" He leaned toward the harness. The intelligence a notch behind the absolute frontier may already be a commodity.
Lin Qiao, who runs Fireworks and sits exactly where that value is sliding, put numbers on it for 20VC. She expects a 10x reduction in token cost over three years, driving 100x usage. Asked whether OpenAI and Anthropic are overvalued if most enterprise workflows can run on open models, she did not dodge: "I think people start to realize it." Her stated motive was not price. "What I don't want to see is there's only one company owns intelligence."
Clément Delangue has the receipts for that being a business rather than a philosophy. Hugging Face crossed $100 million in annual recurring revenue selling around open source. Asked whether frontier labs are being unfairly competed with by distillation, he refused the framing. "If you suck, you suck, with or without distillation." Then the real point: "We're heading towards a world where a few companies are completely dominating, concentrating all power, all capabilities, all wealth, and that's much more dangerous than maybe losing a couple of billion dollars of revenue."
The awkward part of the week is that the leading open models are not American. Krishnan, who spent his government tenure on exactly this, said so plainly, and described a security researcher who moved work onto Kimi because the American model kept refusing him. "A very weird spot to be." His closing argument was unsentimental. "If you are providing a product of value, capitalism will find a way to make the supply chain work for you." Chips, neoclouds, data centers, gas turbines. The stack reorganizes around whoever is delivering.
Mentions of open source and open weights ran 5.5% of business podcast episodes in July 2025. This month, 10.0%. That did not double because more people learned the term. It doubled because it became a procurement decision.
Sources: Sriram Krishnan (formerly The White House), a16z, "Sriram Krishnan on Open Source AI's Biggest Week Yet," Jul 24, 2026. Lin Qiao (Fireworks AI), The Twenty Minute VC, "Are OpenAI & Anthropic Overvalued? The Open-Source AI Reality," Jul 20, 2026. Clément Delangue (Hugging Face), a16z, "Hugging Face's CEO on Open Source AI, Model Routing, and the Future of Competition," Jul 20, 2026.
A rapid run of open-weight model releases is putting pricing pressure on the frontier labs, pushing value down the stack toward inference providers and raising the question of whether the moat is the model or the harness.