Building on open source AI? Hugging Face’s policy chief on the costs, risks, and trade-offs

The era of the closed-door AI model is hitting a wall. For the last two years, we have been told that building on anything other than a proprietary API was a fool's errand. The narrative was simple: big labs have the most compute, therefore they have the best models, therefore you should just rent their brainpower and be happy.

But the wind is shifting. We are seeing a massive migration toward open-source or, more accurately, open-weights AI. Hugging Face has become the town square for this movement, and their policy chief, Irene Solaiman, recently laid out the stakes for what this means for founders. If you are building today, you are likely caught between the convenience of a closed API and the sovereignty of an open model. Choosing wrong could kill your margins or your autonomy.

The Control vs. Convenience Trade-off

When you build on a closed model, you are essentially a tenant. You don't own the infrastructure, you can't see the updates before they happen, and you are subject to the pricing whims of a provider that might eventually become your competitor. Open models change that dynamic by putting the weights directly in your hands.

The biggest misconception I see among founders is that open source is "free." It is not. While you don't pay a per-token tax to a provider, you pay in engineering hours and infrastructure. You have to handle your own orchestration, your own fine-tuning, and your own security. For a builder, the trade-off is clear: do you want to spend your time optimizing the model, or do you want to spend your money outsourcing the headache?

The Global Arms Race

We cannot talk about open AI without talking about the geopolitical reality. There is a lot of noise about Chinese models like Qwen and DeepSeek catching up to or surpassing Western counterparts. Some policy hawks use this as an excuse to argue for closing down open-source releases in the U.S., fearing that we are giving away the secret sauce.

Solaiman points out a nuance that often gets lost: open source is a defensive strategy as much as an offensive one. When the U.S. ecosystem releases open weights, it sets the standard. It creates a global infrastructure that relies on Western-aligned architectures. If we stop releasing open models, we don't stop the progress of others; we just lose our seat at the head of the table. For builders, this global competition is a win. It creates a buyer's market for high-quality weights, regardless of where the training run happened.

Capturing Value in a Post-Scarcity World

If models are becoming a commodity, where does the value go? This is the question that keeps founders up at night. If anyone can download a Llama-3 or a Mistral model that performs at 90% of GPT-4 levels, you can no longer win just by having "the best AI."

The value is moving toward the data and the workflow. The model is just the engine; the car is your proprietary dataset and the specific problem you solve for the user. Builders who rely solely on the model's intelligence as their moat are going to get crushed. Those who use open models to build deep, vertically integrated solutions will own their segment.

The Risks Nobody Mentions

Open source isn't a silver bullet. There are real risks involved in deploying these models at scale. When you use a closed API, the provider handles safety filters and alignment. When you host your own model, you are the one responsible for what it spits out. This requires a level of maturity in your stack that many early-stage startups simply don't have yet.

• Infrastructure debt: Self-hosting requires a team that understands CUDA, quantization, and high-availability GPU clusters.
• Model drift: Without the constant updates of a managed service, your implementation can become stale.
• Compliance: Governments are starting to look at who is responsible when an open model causes harm. If you are the one hosting it, that person is you.

The Founder's Perspective

I have spoken to dozens of founders who started on OpenAI and are now frantically trying to migrate to open weights. Usually, it's not because of a political stance on open source—it's because of the bill. When you hit a certain scale, the per-token cost of a closed model eats your entire margin. Open source is the only path to profitability for high-volume applications.

My advice to anyone starting a project today: start with the best API available to prove the product-market fit, but design your architecture as if you will be switching to an open-source model in six months. Don't get locked into proprietary features that you can't replicate on-prem. The builders who win this cycle won't be the ones who built the best model, but the ones who were flexible enough to use whatever model was cheapest and fastest at that specific moment.

The democratization of AI isn't just about making models available; it's about making sure that builders don't have to ask for permission to innovate.

We are moving into a world where the "intelligence" part of the stack is effectively free. The winners will be the ones who focus on the unglamorous parts: data pipeline, user experience, and solving actual human problems rather than just chasing benchmarks.

Read the original at TechCrunch Startups →