Hugging Face CEO Says Enterprises Are Done Renting AI

Clem Delangue has a pitch he repeats often enough that it now sounds like a mission statement: companies start on frontier APIs because it's easy, then the bill arrives, and they migrate to open models they can actually own. He made the case again this week on TechCrunch's Equity podcast, and the numbers he can point to are real. Half the Fortune 500 use Hugging Face. Chinese labs now supply the majority of what gets downloaded on his platform. Nvidia offered $500 million to buy in and got turned down.

None of that is fabricated. It also isn't the whole picture. A separate dataset published two days earlier, drawn from a16z's own survey of Global 2000 technology executives, shows enterprise AI budgets moving toward closed models, not away from them.

The Claim, and What Backs It

Delangue's argument rests on a pattern he says he sees across Hugging Face's customer base: teams prototype on GPT or Claude, scale up, watch the token bill outpace the value delivered, and start looking for a model they can run themselves.

"Companies start out on frontier APIs because it's easy," he told TechCrunch's Rebecca Bellan on the Equity podcast. "But when they start deploying at scale, the costs become unsustainable. They realize they're renting their AI, and they want to own it."

That last figure is worth sitting with. Alibaba's Qwen family now has more derivative models on the hub than Google and Meta combined, according to Hugging Face's own State of Open Source writeup published in March. The platform crossed 13 million users and 2 million public models in 2025, and the share of models coming from independent developers rather than corporate labs rose from 17% to 39% over four years. By the measures Hugging Face controls, the open ecosystem looks bigger and more distributed than it did a year ago.

Clem Delangue has run Hugging Face since 2016 and has publicly opposed the idea that AI infrastructure should consolidate around a handful of companies. Source: commons.wikimedia.org

Delangue also raised a concern that cuts against his own company's growth story. Nvidia's investment offer came with no board seat attached, he said, but he still worried about what a single dominant backer could mean for a platform that's supposed to stay neutral. "If we let a few big players own the frontier models, we risk creating a world where innovation is bottlenecked by a few corporate gatekeepers," he said on the podcast.

The Numbers Say Something Different

Two days before Delangue's interview ran, 24/7 Wall St. reported on a different dataset entirely: a16z's own CIO survey of 100 verified Global 2000 technology executives. It measures something Hugging Face's report doesn't. Money.

Open-source models accounted for 19% of enterprise AI spending a year ago. That figure has since fallen to 11%. Closed models climbed from 81% to 89% of the average $7 million annual LLM budget over the same period, up from $4.5 million two years prior. David Sacks made the same argument on the All-In podcast around the same time, telling listeners to ignore GitHub stars and download counts and look at what companies are actually buying. "Enterprises continue relying on closed models for production systems where reliability matters most," he said.

What Hugging Face Measures

Downloads. Accounts. Derivative models. Every metric in Hugging Face's own report tracks activity on its platform, and by that measure, open weights are winning. A developer downloading a Qwen checkpoint to fine-tune over a weekend counts the same as a bank quietly running it in production. The report can't distinguish between the two.

What A16z Measures

Dollars actually spent, self-reported by the executives who sign the checks. It captures production budgets, the money going toward systems companies are willing to bet revenue on, and it says nothing about experimentation or the workloads running quietly on hardware the company already owns.

Open-weight models are cheap enough to self-host on rented or owned hardware, but the infrastructure and staffing costs of running them in production still favor teams with existing platform budgets. Source: pexels.com

Neither figure is wrong. They're measuring different things, and the gap between them is the actual story: enterprises are downloading open models in huge numbers while still writing their biggest checks to closed ones. Both trends are true at once, which is a less satisfying headline than either company's press appearance suggests.

Where the Money Might Still Move

There's a version of Delangue's argument that survives the a16z numbers: price per token. Open models often cost five to twenty times less to run than proprietary equivalents, and vendors selling multi-model routing have reported enterprise token costs falling sharply this year as customers mix cheaper open models into high-volume, lower-stakes pipelines. That figure comes from a router vendor with an obvious stake in the trend, so it deserves the same skepticism as any vendor claim, but the underlying economics aren't in dispute. Cheaper open inference paired with rising overall AI spend can make dollar-share statistics misleading in both directions at once.

That's likely closer to reality than either "open source is winning" or "open source is a mirage." Companies are routing cheap, high-volume work to models like DeepSeek V4 or Qwen while keeping expensive reasoning and coding tasks on GPT and Claude, where reliability still outweighs per-token price. Our guide to self-hosting open LLMs covers what that split costs in practice, and our rundown of the strongest open-weight models tracks which ones are worth the overhead.

Should You Care?

If you're deciding where to run your next workload, Delangue's framing and the a16z numbers point to the same practical answer: use open models for the volume, and keep closed ones for the parts of the product you can't afford to get wrong. The fight over whether open source is "winning" is mostly a fight over which metric gets to define winning. Hugging Face counts activity. A16z counts invoices. Both counts are accurate, and only one of them determines who gets paid.

Sources:

• Hugging Face's CEO on why companies are done renting their AI - TechCrunch
• Open source AI matters more than ever, according to Hugging Face's Clem Delangue - TechCrunch
• State of Open Source on Hugging Face: Spring 2026 - Hugging Face
• Why AI start-up Hugging Face turned down a $500mn Nvidia deal - OODA Loop
• The Open-Source AI Boom May Be a Mirage - 24/7 Wall St. via Yahoo Finance
• Hugging Face raises $235M from investors, including Salesforce and Nvidia - TechCrunch
• Enterprise AI Costs Crash 67% As Open Source Models And Multi-Model Routing Go Mainstream - Open Source For You