5 min read

Rent, again.

The narrative about LLM “alchemy in the hands of a few” is smoke and mirrors. European labs just shipped frontier-approaching models across speech, image, and language in two weeks ¹. So if the models aren’t differentiated, what are the big labs actually selling?

That’s easy — the oldest trick in the big-tech playbook: build a platform around the model and charge rent. Claude Code, ChatGPT integrations, proprietary agent harnesses, curated plugin ecosystems. We’ll have OpenAI worlds and Anthropic worlds, the same way we had AWS and GCP worlds. The models are the commodity. The platform is the moat. The platform is where the lock-in lives.

This isn’t EU vs US. It’s big tech vs open source — again.

The 2010s cloud cycle ended with the long tail paying rent to a handful of platforms. Open source lost that cycle because building the alternative was expensive. Distributed databases, identity layers, storage engines — that took teams of PhDs and years of engineering. Open source couldn’t out-build the platforms.

The lock-in target this time is middleware — model routers, tool connectors, agent harnesses, integration glue. The middleware is plumbing, not frontier research. A small team that learns how to make the best use of AI can build what a platform team built in 2015. The economics flipped. Open source doesn’t need to out-build the platforms. It needs to out-organize them. That’s a different fight, and it’s winnable.

I ran 12+ models through the same open-source harness while building Pflegebericht, a production healthcare app. Open weight, European, American, Chinese. The differences were marginal. €15/month on tokens total. That €15 wasn’t an achievement — it was proof the model was never the bottleneck. Databricks’ benchmark on a multi-million-line codebase reached the same conclusion: “Open models […] are now able to handle even the highest level of task difficulty.” ². The value is in the layer above the model. That’s where the fight is.

The protocols are being defined right now — MCP for agent-to-tool ³, A2A for agent-to-agent ⁴, AG-UI for agent-to-user ⁵. The Agentic AI Foundation, under the Linux Foundation, now hosts 10 projects — MCP, Agent Skills, Open Responses — uniting Anthropic, OpenAI, and Hugging Face around vendor-neutral governance ⁶. The pieces exist. The consolidation hasn’t happened yet. Some are calling this the “Kubernetes moment” for open-weight AI ⁷. That’s the right instinct — but Kubernetes took years to consolidate from competing orchestrators. The question is whether open source organizes before the platforms capture the standards.

And here’s the tell: the open-weight frontier is narrowing. Mistral’s Voxtral line went closed after three Apache 2.0 generations ⁸. Mistral Large 4’s weights are still “Coming soon” on Hugging Face, three weeks past the preview — and reports suggest the license may abandon Apache 2.0 for a $20M revenue gate, the same “free-until-you-succeed” model Meta, DeepSeek, and Moonshot already adopted ⁹. Alibaba is reportedly planning a revenue share on the next Qwen ¹⁰. The labs know the models aren’t the moat. The moat is the terms. And the terms are closing.

If producing LLMs isn’t in the hands of the few — and it isn’t — then the next frontier isn’t open weight. It’s open source. The whole pipeline — data, training, tooling, integration. That’s where the effort goes. The window is open. It’s narrower than it was three weeks ago.

I’ve been through two hype cycles on this. I think the model isn’t the moat and the economics just flipped. I might be an idealist — but defeatism isn’t a counter-argument. If you’ve got a better one, bring it.

Footnotes

Published on October 2026