Open models are winning on usage, but it’s far less clear who’s getting paid for them. As TechCrunch AI reports, that’s the big question in the latest Equity episode, where Rebecca Bellan talks with Irene Solaiman, Hugging Face’s Chief Policy Officer and SVP of Strategy, about what startups should know before building on open models. TechCrunch’s summary poses the questions rather than giving her answers, so the analysis below is mine, based on recent data.
The usage shift is big
Enterprises, governments and startups want the same three things, according to TechCrunch: lower costs, more control and more flexibility. Recent numbers show how far that has gone:
- Open-weight models (trained models anyone can download and run) now handle 34% of enterprise AI token usage, up from 23% a year ago, according to a September Enterprise Technology Research survey of 200 respondents reported by Techstrong.ai.
- Respondents expect that share to hit 41% within 12 months. That’s their own forecast, not a measured trend, so treat it with some caution.
- 69% of organizations running these models in production cite cost savings as a reason.
Be careful comparing these figures, because the measures vary a lot. Menlo Ventures put open source at just 11% of enterprise LLM usage in 2025, down from 19% the year before.
Downloads don’t pay the bills
The same reporting found that open models still attract less enterprise spending than proprietary ones. So they’re taking usage faster than they’re taking revenue.
Think of an open model as a published recipe. The recipe is free, but someone still sells the ovens, runs the kitchen and owns the restaurant. My read on who’s best placed to make money:
- Infrastructure sellers: chipmakers and inference hosts get paid no matter which model wins.
- Model makers with something else to sell: Alibaba can give away Qwen because it also runs a cloud business.
- Application builders: companies with proprietary data and sticky workflows can keep their margin as model prices fall.
- Pure-play open labs: they have the hardest path, since their main product is free.
Attention is also highly concentrated. A summary of Hugging Face’s mid-2026 data says 1.5% of repositories account for 99.2% of downloads. A few models get nearly all the attention, and everyone else is fighting over scraps.
One caveat: Hugging Face hosts these models, so it benefits when open AI grows. Solaiman knows this field well, and her earlier work at OpenAI included GPT-2’s staged release, but weigh any open-source optimism with that incentive in mind.
America is behind on downloads
TechCrunch also asks whether American players can keep up with Chinese open models. On downloads, they’re currently behind. Hugging Face’s spring report gave Chinese models about 41% of downloads from February 2025 to February 2026, compared with 36.5% for US models. Alibaba’s Qwen family reportedly passed 3 billion downloads this summer.
Downloads are a noisy signal, since automated pulls and small variants can inflate them. They still matter, because the model developers try first often becomes the one they build on.
What I expect by 2028
These are my projections, not Solaiman’s:
- Mixed setups become normal. Open models take high-volume routine work, while closed models keep the hardest tasks.
- Profit keeps moving away from the model itself and toward serving, tooling and workflow.
- Model origin becomes a standard procurement question, especially for government and regulated buyers. That could split the market, even if Chinese models keep leading on raw downloads.
What to do now
- Price the whole stack. A free model still needs GPUs, engineers, monitoring and security work, and at low volume a closed API may cost less.
- Read the license. “Open” doesn’t always mean unrestricted commercial use.
- Test on your own workload. Public benchmarks won’t tell you how a model handles your documents and edge cases.
- Keep models swappable. Put a routing layer between your product and the model so switching takes days, not a rebuild.
- Decide your origin policy early. If you sell to government or regulated industries, find out what your customers will accept before you commit to a model.
- Build your moat above the model. If a rival can download the same weights tomorrow, your edge has to come from your data, distribution or workflow.
The free part of AI is getting bigger, and the paid part is moving up and down the stack around it. The full conversation with Solaiman is on TechCrunch’s Equity podcast.