Somebody Still Runs the Racks
A legal AI company stopped renting its intelligence this week and started owning it. Harvey announced its first in-house model, post-trained on the open-weight Kimi K3, and the verdict on This Week in Startups arrived inside a minute: "Open source is going to win it all."
Jason Calacanis has been making that call for three years, and defines winning narrowly. Not benchmarks. Token share. "The majority of tokens, the overwhelming majority of tokens, in corporate America. will not be on the frontier models, it will be inside those enterprises." His estimate of what Harvey had been paying for inference before, offered as a guess and labeled one: "10000000 a month with open AI."
The rebuttal came with arithmetic. Ray Rike and Peter Buchanan spent an episode on the category error underneath the celebration: open weights are not open source. Traditional open source scales at zero marginal cost. Weights do not, because "these open weight models don't get the same scaling advantage of traditional open source software," and every additional customer still buys chips, power and capacity. So the margin the closed labs collect does not evaporate when the weights get published. "it doesn't disappear when the model becomes open weight. So it just moves to whoever hosts it."
Their evidence is a split inside one Chinese lab's numbers. Per their reading of Z.AI, on-premise deployments, where the customer absorbs its own hosting, run roughly 49 to 50% gross margin. The same models served by API on third-party cloud run about 19%. Identical weights. The variable is who pays the electricity. Which is how they land on the hyperscalers winning regardless of who wins the argument: "somebody needs to host and run these son of a guns."
Martin Casado reconciled both camps in a single sentence, and flagged it as a guess too. "the big labs will probably dollar weighted, get 80% of the market going forward because that's historically what we've seen for large incumbents, but I think token weighted 60% will be long tail and open source." Calacanis can win the token count and the labs keep the revenue. Those are not competing claims.
The price data points the same way. Silicon Data's Steve Hou, who is building a reference price for GPU rental, says rates have been rising rather than collapsing, and that the market may be wrong about how fast the hardware becomes worthless: "these things may not depreciate as quickly as people thought and may have actually a longer lifespan as people have previously budgeted for." Hosting is not a commodity either, since hyperscalers hold "a persistent significant premium, maybe 2 to 3 times the cost of a new cloud" for a bundled service.
The cleanest version came from a much smaller company. Willow's Allan Guo made his dictation product free because he could see what was coming: "rather than, um, getting eaten by this commoditization, we want to commoditize it first."
Sources: Jason Calacanis and Allan Guo (Willow), This Week in Startups, "Open source is going to win it all: Harvey proves it | E2328," Aug 21, 2026. Ray Rike (Benchmarkit) and Peter Buchanan (New Plan), Metrics that Measure Up, "Can U.S. Frontier AI Labs Survive a Price War with Open-Weight Models?," Aug 18, 2026. Martin Casado (Andreessen Horowitz), a16z, "Martin Casado on Where the Value Is Going in AI," Aug 22, 2026. Steve Hou (Silicon Data), Equity, "AI has a GPU pricing problem. Silicon Data wants to fix it.," Aug 19, 2026.
Harvey's move to an open-weight base model was read as proof open source wins, but open weights lack open source's zero marginal cost, so the margin migrates from the labs to whoever pays to host and power the models.