Open-weight AI is booming, but the money sits in hosting, tuning and governance

The model is free; running it is not

Open models have moved into the mainstream, according to TechCrunch, with enterprises, governments and startups adopting them for lower costs, greater control and more flexibility. That shift frames an episode of TechCrunch's Equity podcast in which Rebecca Bellan speaks with Irene Solaiman, Chief Policy Officer and SVP of Strategy at Hugging Face, about what startups should weigh before building on open models — from tooling and infrastructure choices to the trade-offs shaping the field.

An analysis published on dev.to, originally written for a newsletter called Agentic Affairs, offers a blunt thesis: the download was never the product. The weights cost nothing, but everything needed to turn them into software a company can actually operate is where invoices get issued.

Open-weight is not the same as open-source

Most models marketed as open ship only their weights, the dev.to post argues — not the training data, not the recipe, not a way to reproduce the model. Several of the largest carry bespoke licences that impose revenue thresholds on managed services. The practical advice is to check licence terms rather than vendor messaging before committing.

Downloads concentrate on a few — increasingly Chinese ones

The dev.to analysis cites concentration at the top of the ecosystem: the 200 most-downloaded models, a hundredth of a percent of the roughly 2 million models on Hugging Face, account for close to half of all downloads. Developers who want something they can download and run increasingly start with DeepSeek or Alibaba's Qwen, the post argues, while American labs release open models voluntarily and irregularly. Steady release rhythm, not one-off announcements, is what earns developer attention.

The layers where revenue actually pools

The post identifies six ways to charge for removing the pain of production: hosted inference, managed cloud deployment, enterprise governance, fine-tuning services, commercial licensing and consulting. A model that is free to download but requires a GPU fleet, operations staff and sign-off from a risk committee is not free in any meaningful sense. This aligns with the question TechCrunch puts at the centre of its conversation: when capable models circulate freely, value has to be captured somewhere else in the stack.

Licences do not settle the politics

The dev.to analysis also flags regulatory exposure. It states that the US Department of Homeland Security barred DeepSeek from federal agencies in October 2026, despite the model's permissive MIT licence, and that members of Congress are already scrutinising organisations running Qwen and Moonshot models. The lesson it draws: a licence governs what you may legally do with a model; it says nothing about whether regulators will let you do it.

Benchmarks will not answer your question

On capability, the post is agnostic, noting that research has not settled whether open models currently match frontier proprietary ones. Its recommendation is to validate candidates against your own workload before committing, because any gap that matters will surface in production, where it is expensive to discover.

Why it matters

Both pieces converge on the same reframe. If model weights become a shared, commoditised resource — increasingly supplied by Chinese labs with high release cadence — then competitive advantage and profit migrate to the surrounding layers: hosting, deployment tooling, governance, distribution and trust. For startups picking a foundation, licence fine print, release rhythm and regulatory exposure matter as much as benchmark league tables. For the industry, the question TechCrunch raises about whether American players can keep up becomes commercial rather than ideological. One caveat worth holding onto: the dev.to piece is an opinion analysis, and its download figures and the homeland security claim are presented as its own reporting rather than independently verified news.