A clear geopolitical split has emerged in approaches to artificial intelligence. In Washington, policymakers increasingly treat AI as a national security asset that should be sequestered to prevent adversaries from gaining advantage. In Beijing, the approach is opposite: China is lowering barriers, encouraging adoption, and using open-source AI to help spread Chinese-developed technologies around the world.
Public endorsements and the July 24 turning point
On July 24, a high-profile public push for open source moved a debate that often happens behind closed doors into the open. Jensen Huang, CEO of Nvidia, posted for the first time on X and linked to an open letter signed by 35 companies—including Palantir, Andreessen Horowitz and Microsoft—urging Washington not to over-restrict open-source software. Huang argued that “open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.” Prominent figures such as Elon Musk and Mark Zuckerberg voiced support as well.
Mozilla signed the letter, stating that although it differs from many signatories in structure and perspective, it agrees with the letter’s core claim that “openness may be one of the most important paths to AI safety and security.” The letter was followed by the announcement of the Open Secure AI Alliance for AI Safety and Security, intended to “build and share open tools that promote responsible use of and trust in AI.”
The gap between usage and revenue
Mozilla’s new "State of Open Source AI" report highlights a striking economic gap: open models now perform roughly one third of the world’s AI work while capturing only about four percent of the money. That discrepancy frames much of the current struggle.
The report shows that while gaps between open models (those whose weights anyone can download, run, and adapt) and top proprietary systems remain in certain areas, the average performance difference has narrowed sharply over the past year. Costs are falling: 79 percent of developers now build with open models. Yet revenue has not followed usage—creating a clear prize in the middle of the debate.
The limits of a "kill switch"
The debate goes back nearly three years, when Anthropic’s Dario Amodei warned Congress that advanced open-source AI was on a “very dangerous path.” On its face, the argument is that once a model’s weights are public, access cannot be revoked and abuse cannot be fully prevented. But revocation is not a feature of a model itself; it’s a property of API contracts and the service relationship. That applies only to a lab’s customers—not to actors who obtain or run the model independently.
Many labs releasing frontier-class open weights—such as DeepSeek, Alibaba, Mistral, and Moonshot (which recently released Kimi, a 2.8 trillion-parameter model)—largely operate outside Washington’s direct reach. Two million open models already sit on Hugging Face, and many can run on a laptop. In practice, there is no single global kill switch to throw.
Where power is actually concentrating: the layer around the model
Increasingly, the decisive factor is not the model weights themselves but the software wrapped around them. The same open model can power very different products depending on the orchestration layer: a scam-call operation or a nurse-advice hotline. The model does not know—or care—about that difference; the agentic harness between user and model controls access, memory, and behavior.
As models grow cheaper and more interchangeable, the harness becomes the locus of power. Large labs are quietly locking that layer down. The analogy is familiar to farmers who bought tractors but the manufacturer retained control of the software—years of litigation and, as noted recently, action by the Federal Trade Commission were needed to pry that lock open. A similar dynamic is forming for the software that reads your email, books your travel, and remembers every detail of your life.
This is an economic choice, not an engineering accident: a closed wrapper makes leaving expensive; an open one survives by remaining worth using. The same underlying technology delivers opposing incentives, and that helps explain why open models account for about a third of usage but only four percent of captured value.
What developers can do to keep the future open
The relevant debate is not merely which models are released or regulated but who controls the orchestration layer. That decision is being made now—largely by developers, often without full awareness of their responsibility. The early web was shaped by developers choosing HTML and HTTP over proprietary walled gardens; AI is at a similar turning point, and two million open models indicate many builders have arrived early.
Mozilla and the author recommend four practices that matter most for preserving an open future:
- Build on open harnesses, not just open models. Capability is concentrating above the weights, and closed labs are already sealing that layer.
- Own the memory layer. Store accumulated context in portable, controllable formats so it remains retrievable if a vendor changes terms.
- Keep a second model warm. Integrate an open model and keep it production-ready even while primarily running on a closed API, so switching is inexpensive if needed.
- Don’t assume all open stacks are equal. Open models cluster by region and provider; a genuinely open orchestration layer requires actively supporting a geographically distributed set of options rather than defaulting to the cheapest model this quarter.
None of these recommendations assumes bad faith on anyone’s part. It is notable, however, that the loudest safety arguments intensified as models became cheap enough for real competition to move up a layer. That’s not proof of conspiracy—just a reflection of where incentives point, and why much of the debate targets the wrong object.
Closing: who holds the keys?
Most of the report’s evidence is positive: performance gaps are closing, costs are collapsing, and millions of developers are building with open models. The central question for developers is whether they will hold the keys to the systems they build. For governments, the question is whether the keys they reach for will actually change anything. For now, the door is still open—and keeping it that way will require collective effort.
A final note: AI Codecon: Building with Open Source AI, a free half-day virtual conference, will take place on August 31, 2024, featuring developers and technical experts working with open-weight models, self-hosted infrastructure, and real-world AI workflows. Registration is encouraged to reserve a spot.
