Why Big Companies Are Moving from Renting to Owning Their AI Models

Hugging Face CEO Clem Delangue says open-source AI is booming, driven by a clear pattern he has watched repeat across large enterprises: companies start by paying for access to a hosted AI service, then decide they want to own and control the model themselves. According to Delangue, roughly half the Fortune 500 now uses Hugging Face, which has grown into something like a GitHub for AI, hosting open models and datasets that teams can download and run on their own infrastructure.

What happened

Hugging Face CEO Clem Delangue spoke to TechCrunch about the state of open-source AI and why enterprise appetite for it keeps growing. His headline claim: roughly half the Fortune 500 now uses Hugging Face, a platform that lets AI builders share, download, and deploy open models and datasets. Think of it as GitHub, but for AI model weights instead of code repositories.

Delangue says he has watched the same sequence unfold at company after company. They start by paying a cloud provider or AI vendor for access to a hosted model. Then, at some point, they decide they would rather own their AI than keep renting it.

Why do companies stop renting AI and switch to open models?

The reasons tend to cluster around control, cost, and data privacy. When a company runs a third-party hosted model, the vendor controls the pricing, the update schedule, and often the terms around what data can be sent to the API. Switching to an open model that runs on your own infrastructure removes those dependencies.

Cost is a real factor too. At low usage volumes, paying per API call is convenient. But as a product or internal tool scales, the per-token bills grow fast. Running an open model on owned or rented compute can flatten that curve significantly, even after accounting for engineering time.

Data sensitivity is a third pressure. Industries like finance, legal, and healthcare often cannot send confidential documents to an external API without legal review. An on-premise or private-cloud open model sidesteps that problem.

Why it matters

The Fortune 500 statistic is worth pausing on. If Delangue’s figure is accurate, open-source AI is no longer a niche preference for startups and researchers. It is now a mainstream infrastructure choice for large organisations with serious compliance and procurement requirements.

That has knock-on effects for the whole market. It puts pressure on closed-model providers, since every enterprise that self-hosts an open model is one fewer paying API customer. It also raises the stakes for model quality in the open ecosystem: companies making this switch need models that are good enough to do real work, not just demos. Recent releases from Meta, Mistral, and others have made that bar easier to clear.

For smaller businesses, the trend is a signal worth watching. The tooling and support infrastructure that enterprises are building around open models will eventually trickle down, making self-hosted AI cheaper and easier to run at smaller scale.

If you are already exploring how AI fits into your own workflows, the question of build-vs-buy is directly relevant. Our work on AI integration for business clients regularly surfaces this exact decision point, and the right answer depends heavily on your data sensitivity, budget, and team capacity. For broader context on how the open-source and proprietary AI markets are moving, see our full AI news coverage.

Our take

Delangue has an obvious interest in talking up open-source AI adoption. Hugging Face is the platform that benefits most when enterprises go in that direction. So take the Fortune 500 figure as a directional signal, not an audited stat.

That said, the underlying dynamic he describes is real and we see it in client conversations. The first AI project often starts with a ChatGPT or Claude API key because it is fast to set up. Then someone looks at the monthly bill, or the legal team raises a data question, and the conversation shifts toward alternatives. Open models are the most common landing spot.

The practical caveat is that “owning your AI” is not free. You trade a predictable per-token cost for engineering overhead: model selection, hosting, fine-tuning, and ongoing maintenance. For most businesses without a dedicated ML team, a managed open-model service (still technically renting, but with more portability) is a more realistic middle ground than fully self-hosted infrastructure.

What to do about it

1. Audit your current AI spend: list every API you pay for, the monthly cost, and what data you are sending to it.
2. Flag any data that legal or compliance would not want leaving your infrastructure. That list defines your minimum case for exploring open models.
3. Run a small test: pick one internal use case, deploy a relevant open model via a managed service (Hugging Face Inference Endpoints, for example), and compare output quality and total cost over 30 days.
4. Decide on a hosting model before you commit: fully self-hosted, private cloud, or managed open-model API each carries different tradeoffs on cost, control, and maintenance burden.
5. Revisit the build-vs-buy question every six months. The open model ecosystem is moving fast enough that a model that was not good enough last year may now meet your needs.

The best starting point is an honest look at what you are actually paying today and what control you are giving up for that convenience.

Source: TechCrunch · AI