Open-Source AI Was Unleashed. Now It Has Become National Policy.

Open-Source AI is National Policy.

11:09

Open-Source AI Was Unleashed. Now It Has Become National Policy.

In October 2023, I published a short article called "Open Source AI UNLEASHED."

At the time, open-source AI felt like a developer story. New models were appearing every week. Small teams were releasing weights, researchers were shrinking architectures, developers were building retrieval systems, and robots were beginning to connect computer vision with language. The names came so quickly that keeping them straight was nearly impossible.

Beneath the noise, something important was happening. Artificial intelligence was escaping the laboratory. It was becoming something ordinary people, small companies, universities, and independent developers could download, alter, and build upon. I wrote then that these tools had "immense potential to benefit humanity." That sounded optimistic in 2023. Today, it sounds almost understated.

On July 24, 2026, NVIDIA joined Microsoft, Meta, IBM, Palantir, CrowdStrike, Dell, Hugging Face, Andreessen Horowitz, Y Combinator, and more than two dozen other organizations in signing a letter titled "Open Weights and American AI Leadership."

Their argument was no longer merely that open AI helps developers innovate. Their argument was that open AI is necessary for America to remain America.

The progression is almost perfect.

In 2023, the conversation was: look what developers can build. New models arriving rapidly, smaller architectures becoming more capable, new retrieval and vector tools, multimodality and robotics, broader access, concern about transparency and alignment. The argument was fundamentally optimistic — put powerful tools into more hands, steward them responsibly, and innovation accelerates.

In 2026, the conversation is: a country that cannot download, modify, and operate advanced models on infrastructure it controls is not technologically sovereign. The rebellion became infrastructure. Open source began as a challenge to the idea that software could only advance behind corporate walls. Open-weight AI is now challenging the idea that intelligence itself should be metered through a handful of corporate endpoints.

The letter's central claim is that U.S. AI leadership will not be determined by whether one American company owns the best frontier model. It will be determined by whether advanced AI capability spreads through the entire economy — factories, hospitals, farms, universities, small businesses, government systems, and critical infrastructure.

That sounds obvious until you think about what it implies.

The signatories are collectively arguing that the AI economy cannot be allowed to collapse into a handful of hosted APIs controlled by a few companies. They are drawing a distinction between two layers: frontier closed models for the hardest, most expensive, highest-capability tasks, and open-weight and specialized models for the billions of routine tasks that will actually embed AI throughout society. This is a practical argument, not an ideological one. Paying frontier-model prices for every inference would be economically irrational. The math does not work at scale. Organizations need the ability to run smaller models locally, modify them, own the institutional knowledge those models accumulate, and deploy them wherever operational requirements demand — including places where you cannot send data to a commercial API.

The word that keeps appearing beneath the whole letter, quietly, is sovereignty.

Open weights mean that a company, university, hospital, military organization, or country can control its own data, run models on its own infrastructure, avoid permanent dependence on one provider, retain the improvements it makes, and continue operating even when commercial relationships, regulations, or geopolitical conditions change.

That maps directly onto the infrastructure argument I have been making all week. Sovereign AI is not simply owning a model file. It requires compute, power, cooling, data, networks, secure facilities, deployment expertise, and an application ecosystem. The model weights are one piece. Everything else is the physical and operational layer underneath.

Jensen Huang's statement — "AI will be built by every country" — is not merely a defense of open development philosophy. It is legitimizing the idea that nations and major institutions will operate their own AI systems, on their own infrastructure, using combinations of closed frontier models and locally controlled open models.

For NVIDIA, that framing is also commercially elegant. A world dominated exclusively by a few closed model providers concentrates enormous compute purchasing power in a small number of hyperscalers. A world where every nation, company, university, and agency builds sovereign AI creates far more customers for GPUs, networking, and AI factories. This is both principled policy and market expansion doctrine, which is probably why the argument is as clear as it is.

The security section is the most aggressive part of the document. The letter directly rejects the assumption that closed models are automatically safer. Its position: closed systems can still be compromised, misused, or fail invisibly, while concentration creates a small number of catastrophic single points of failure. Open models allow wider testing, benchmarking, red teaming, vulnerability discovery, and defensive development.

That is the old cryptographic argument, proven correct multiple times over the last thirty years: security should come from resilient architecture and rigorous scrutiny, not from hiding how the system works. Security through obscurity has a long track record of failing exactly when it matters most.

The letter acknowledges the obvious counterargument — once weights are released, they cannot realistically be recalled — but argues that defenders need access to capabilities comparable to those available to attackers. You cannot defend against a weapon you are not allowed to understand.

The distillation paragraph is one of the most politically consequential parts of the document, and it is easy to miss. Distillation — using one model's outputs to train or improve another — is how a significant portion of AI progress happens. The letter argues this is a legitimate and longstanding technique that should not automatically be treated as theft. It leaves room for targeted action against clearly unlawful extraction from closed systems, but it draws a line between that and broad restrictions that would effectively prevent competitors from learning from existing systems.

Do not let intellectual property enforcement become a mechanism for freezing the current AI hierarchy in place. This is politically significant because some of the most capable closed-model companies have an obvious incentive to define distillation broadly. If they succeed, they could use IP law to prevent open-weight models from ever catching up — not because those models are using stolen data, but because learning from the outputs of another system is defined as infringement. The letter is an attempt to draw that boundary before governments or dominant companies define it for everyone.

I have been writing this week about the technology stack reorganizing — Bezos's Prometheus targeting the engineering organization, Musk's Macrohard targeting the business organization, the hyperscalers trying to become vertically integrated intelligence systems rather than cloud providers. This letter is the open-weight layer of that same map.

The AI landscape is separating into several layers. Frontier intelligence labs building the most capable models. Compute and infrastructure empires supplying chips, power, data centers, and training systems. Open model ecosystems distributing capability beyond the frontier labs. Sovereign and enterprise AI factories adapting models to specific national, industrial, and institutional purposes. And application layers where AI finally becomes operationally useful.

The winners may not be the company with the single smartest chatbot. They may be the companies and countries that control the infrastructure, distribution, and adaptation layers around intelligence. This letter is a public declaration that the open-weight layer is now part of the official architecture of American AI leadership, not a fringe alternative to it.

The Bell Labs framing connects here in a specific way. A civilization-scale Bell Labs would not consist of one proprietary model hidden behind an API. It would require shared research infrastructure, open technical foundations, national compute capacity, public-private laboratories, specialized models, abundant energy, and thousands of organizations capable of experimenting at the edges.

This letter is basically making the same argument in policy language. America should not build one artificial mind. It should build the ecosystem from which millions of artificial capabilities can emerge. The open-source software movement did not win because it was ideologically appealing. It won because the economics were right, the architecture was resilient, and no single company could out-compete a global community of contributors building on a shared base.

In 2023, open-source AI looked like a torrent of experimental models and developer tools. I was documenting an ecosystem moving so quickly that model names, companies, and architectures were blurring together weekly. Some of the details in that article were garbled — model names slightly wrong, company attributions mixed up. But the underlying signal was exactly right.

By 2026, the largest technology companies in America were describing open weights as essential to national security, economic competition, and technological sovereignty.

In 2023, open-source AI was unleashed. In 2026, America finally understood what had been unleashed. Not a cheaper chatbot. A new industrial foundation.

---

Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital.