Meta’s Llama releases were supposed to democratise AI. And in a narrow technical sense, they did. Models you can run locally on consumer hardware are now genuinely competitive on standard benchmarks with systems that cost hundreds of millions of dollars to train. That’s a real shift, and it happened faster than most people in the industry expected.

But something odd has followed from it. The gap closing hasn’t weakened the proprietary labs - it’s weakened the mid-tier. The companies that positioned themselves between fully open and frontier-closed are the ones getting quietly squeezed.

The Middle Is Gone

Two years ago, there was a defensible market position for an AI company that offered a fine-tuned or specialised model - something better than open-source but cheaper than OpenAI. That gap existed because running your own model required serious infrastructure and expertise. It was a real differentiator to handle that friction for customers.

That friction is largely gone now. Tools like Ollama have made local model deployment genuinely accessible. Cloud inference costs have fallen sharply. The customer who once paid a mid-tier provider a monthly fee to access a wrapped model now has real alternatives on both ends - free and open at one extreme, frontier reasoning capability at the other.

The middle position required the open-source ceiling to stay low. It didn’t.

Who Actually Benefits

Meta benefits, obviously - Llama releases generate enormous goodwill and research talent without requiring Meta to win the end-user product war directly. Mistral benefits in reputation even as its commercial leverage shrinks. The enterprise infrastructure companies - the ones selling GPUs, managing inference at scale, building the plumbing - benefit regardless of which models win.

The companies that don’t benefit are the ones that built products on top of models they don’t control, with no proprietary data layer, no hardware angle, and no defensible distribution advantage. There are a lot of those companies.

The Benchmark Problem

None of this is cleanly visible in benchmark comparisons, which is part of why the narrative has stayed optimistic for so long. Benchmarks measure task performance. They don’t measure the commercial logic of a business that depends on a performance gap that’s actively narrowing.

By the time a gap shows up in a company’s revenue, the benchmark charts have already moved on. The numbers looked fine right up until they didn’t.