The Open-Source Trap

Sponsored by

In August 2025, OpenAI released its first open-weight model since GPT-2. The weights were free. The compute to run them costs money. Someone pays that bill.

Eight months later, most developers still haven't asked who.

In August 2025, OpenAI released gpt-oss-120b, a 120-billion-parameter open-weight reasoning model from its new "gpt-oss" family. Developers can download the weights and run it on their own infrastructure, including a single H100 GPU, because the model uses a Mixture-of-Experts architecture that keeps only 5.1 billion parameters active at any moment. It is efficient by design. OpenAI pays nothing to serve it once the weights are public.

This is not generosity. It is a distribution strategy.

The model OpenAI kept closed is GPT-5.4 Pro. That is the frontier. gpt-oss-120b is not the frontier. It is a capable, well-engineered model designed to embed OpenAI's architecture, tooling conventions, and benchmark framing into every developer environment that adopts it, at zero compute cost to OpenAI. When those developers build products, they build on OpenAI's assumptions. When those products scale, they eventually migrate to OpenAI's paid infrastructure. The open model is the funnel.

This is not a new idea. It is the oldest idea in enterprise software: give away what costs you nothing to distribute, sell what only you can run.

What 200K+ Engineers Read to Stay Ahead

Your GitHub stars won't save you if you're behind on tech trends.

That's why over 200K engineers read The Code to spot what's coming next.

• Get curated tech news, tools, and insights twice a week

• Learn about emerging trends you can leverage at work in just 5 mins a day

• Become the engineer who always knows what's next

Join 200k+ engineers

What makes this moment different is the infrastructure context surrounding it.

While gpt-oss-120b was announced, OpenAI was simultaneously finalizing a $300 billion cloud deal with Oracle and managing a $100 billion letter of intent from Nvidia, structured as a circular investment where OpenAI uses Nvidia funds to purchase Nvidia hardware. By January 2026, reports indicated Nvidia had paused the commitment, citing valuation concerns and the optics of the loop becoming too visible. The deal that was announced as a milestone in AI infrastructure is now being described in financial press as fragile.

Meanwhile, xAI raised $20 billion to build its Colossus 2 supercomputer. The split is $7.5 billion in equity and $12.5 billion in debt. The debt sits inside a Special Purpose Vehicle, which means it does not appear on xAI's main balance sheet. The SPV buys the GPUs. The SPV leases them back to xAI. The company shows the compute. The liability stays offstage.

Two of the largest AI infrastructure announcements of the past six months share a structural feature. The headline describes power accumulation. The balance sheet describes something more complicated.

The gpt-oss-120b release and the fragile Nvidia deal are not separate stories. They are the same story at different scales.

At the company level, the largest AI players are building circular funding structures, SPV-shielded debt, and open-weight distribution plays because they cannot actually afford to run the frontier they are building on their own. The compute required to serve GPT-5.4 Pro at scale demands infrastructure that no single company's balance sheet can cleanly absorb. So the community runs the open model. Oracle holds the compute contract. Nvidia sits inside a letter of intent that may or may not complete.

At the developer level, the invitation to run gpt-oss-120b is real. The weights work. The model is competitive on coding and reasoning benchmarks. But every developer who adopts it is absorbing infrastructure costs that OpenAI has externalized, while embedding OpenAI's architectural conventions into their production stack. The switching cost grows with every integration.

Sebastian Raschka, AI researcher and author of the AHEAD OF AI newsletter, called gpt-oss-120b's tool-use training a paradigm shift, noting that relying on external tool calls rather than internal memory is currently the most viable path to reducing hallucinations. He is right about the technical property. He is describing, without naming it, why OpenAI chose this architecture for the open release. A model that requires external tool calls to perform reliably is a model that requires a compatible tool ecosystem. The schemas gpt-oss-120b is trained to call are OpenAI-compatible. The developer who builds on those schemas is not running an independent stack. They are running a dependency with downloaded weights.

The weights are free. The ecosystem is not.

You don't need to be technical. Just informed.

Most AI newsletters are written for engineers. This one isn't.

The AI Report is read by 400,000+ executives, operators, and business leaders who want to know what's happening in AI — without wading through code, jargon, or hype.

Every weekday, we break down the AI stories that matter to your business: what's being deployed, what's actually working, and what it means for your team.

Free. 5 minutes. Straight to the point.

Join 400,000+ business leaders staying ahead of AI — without the technical overwhelm.

Read it free today

The model is the commodity now. What nobody can externalize is the judgment about which workflow to build, which client to serve, and which problem is worth solving. That is in today's High Stakes Human Skills.

404 Found covers AI developments from a European Insider, three times a week.