Open Source AI Is the Last Fair Fight — Here's Why I'm Betting on It

I've been thinking about this for a while, and I'm going to say it plainly: the capital game in AI is over. The incumbents won. If you're a founder or a builder trying to compete at the frontier model layer with closed weights, you're playing on a tilted board and the tilt is not getting better.

In Q1 2026, OpenAI raised $122 billion. Anthropic raised $30 billion. xAI raised $20 billion. Together, those three companies captured 67.3% of all AI venture capital that quarter. The remaining $83.5 billion was split across 1,543 other deals. Three companies against 1,543. Let that land for a second. I'm not saying this to be dramatic. I'm saying it because the math matters. When capital concentrates that fast, it doesn't just buy compute. It buys regulatory access, it buys government relationships, and it shapes the rules of the road.

Which brings me to something that happened this week. GPT-5.6, which OpenAI is branding as Sol, Terra, and Luna, launched publicly on July 9, 2026. That's today. But before any of us could touch it, there was a 13-day window where access was limited to partners quote vetted by the government. Thirteen days where the U.S. government helped decide who got early access to a commercial AI product. I'm not here to debate the policy rationale. I'm here to point out what it means structurally: we now have a world where model releases can be gated by government fiat. That's a new kind of moat, and it doesn't belong to you or me.

Anthropic's Fable 5 and Mythos 5 got pulled globally for 19 days under U.S. export controls before being restored on July 1st. Nineteen days where a commercially available AI product simply vanished from the market. If you had built your product on top of those models, your product was gone too. Not because of anything you did. Because of a government decision you had no input on and no recourse against.

I've said before that open source AI will do to Anthropic and OpenAI what the open web did to AOL. I still believe that. But I want to be more specific about why, because I think the AOL comparison undersells what's actually happening right now at the capability level.

DeepSeek ships under MIT. Qwen ships under Apache 2.0. Mistral Large 3 ships under Apache 2.0. GLM-5.1 ships under MIT. These are not hobbyist models. These are frontier-grade systems you can run commercially, modify freely, and deploy without paying a royalty to anyone. And crucially, no government can pull them. Once the weights are out, they're out. There's no kill switch. There's no export control that reaches into your server rack and deletes the model you already downloaded.

A few months ago, a model called LongCat-2.0 was quietly benchmarked on SWE-bench Pro under a disguised name. It beat GPT-5.5. For two months. Nobody noticed. That's how close the gap is. The gap from closed frontier to open-source frontier has shrunk from more than 12 months to roughly 3 months, and it keeps shrinking. Anyone who tells you open-source models are a tier below closed models hasn't looked at the benchmarks lately.

Here's my actual thesis, the thing I keep coming back to: open-source AI is the last part of the AI stack where a scrappy founder can genuinely compete on equal footing. Not because open source is morally superior. I'm not making a moral argument. I'm making a structural one. The capital layer is captured. The government-access layer is captured. Inference at massive scale is expensive and the big players have structural cost advantages. But the model weights themselves, the actual intelligence, are public. MIT-licensed weights are yours. Nobody can revoke your access. Nobody can change your pricing. Nobody can pull your model on a government request. That's not a small thing. That might be the only thing left.

I think about founders right now building on closed APIs and I feel genuine concern. Not because the APIs are bad today, but because the last 90 days have shown us exactly how that dependency plays out when policy decisions land. You're not building on a platform. You're renting access to intelligence, and the landlord can change the terms or disappear the property with no notice.

So this is my bet. Not a hedge. A bet. Open-source AI models, the ones with permissive licenses and public weights, are where I'm spending my time, my thinking, and my conviction. Because this is the one part of the stack where the incumbents' structural advantages don't fully apply. The money doesn't buy the moat here. The government relationship doesn't buy the moat here. The weights are out, and they're yours. That's the last fair fight in AI, and I intend to show up for it.