The Battle for Open Weights: Why Jensen Huang’s First Post Matters to Us All

The AI industry has officially drawn its battle lines. Then, within about twenty-four hours, it redrew them.

On Friday, 24 July 2026, NVIDIA CEO Jensen Huang made his first-ever post on X. He didn’t launch a GPU or announce an earnings milestone. He shared a policy letter titled “Open Weights and American AI Leadership,” signed at that point by 25 companies and institutions including NVIDIA, Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Mistral, Andreessen Horowitz, Mozilla, the Linux Foundation and Y Combinator. The message to Washington was blunt: do not hastily impose bans on downloadable AI models.

As Huang put it, The world needs both frontier closed models and frontier open models. The post cleared 11 million views within hours.

To understand why the fight over open weights is arguably the most critical tech policy debate of the decade, we first need to decode what the term actually means.

The restaurant analogy: closed APIs vs open weights

If you want to understand the next two years of AI news, you need to grasp the difference between closed models and open-weight models.

Think of closed models like ChatGPT or Claude as a high-end restaurant. You can order the food (prompts) and pay the bill (API costs or subscription fees), but you are not allowed inside the kitchen. You can’t see the recipe, and the door is locked.

Open-weight models, like Meta’s Llama, Alibaba’s Qwen, or Mistral, are the exact opposite. They hand you the recipe. By publishing open weights, meaning the trained neural network parameters themselves, companies allow you to download the model and run it locally on your own hardware. Nobody can suddenly revoke your access, change the model’s behaviour overnight, or hike the price. If you have wondered what class of machine that actually requires, I have written a comparison of Blackwell-generation GPUs for exactly this kind of local workload.

The catalyst: the Kimi K3 controversy

So why the sudden urgency? The trouble started with the rapid advancement of international models.

On 16 July, Beijing-based Moonshot AI released Kimi K3, a 2.8-trillion-parameter open-weight mixture-of-experts model with a one-million-token context window, which the company bills as the largest open-weight model ever released. It is worth being precise about the performance claims, because they have been widely overstated. Moonshot’s own materials concede that K3’s overall performance still trails the strongest proprietary systems, naming Claude Fable 5 and GPT 5.6 Sol. What K3 did do is post frontier-level results across an evaluation suite and beat Fable 5 on at least one public benchmark, Arena’s Frontend Code evaluation. Full weights were scheduled for release on 27 July, so until then independent scrutiny is limited.

Then it escalated. Michael Kratsios, director of the White House Office of Science and Technology Policy, posted on X that the US government has information that Moonshot distilled Anthropic’s Fable model to build K3. He alleged the company ran an internal platform for large-scale distillation against US models, switching between access methods to avoid detection, and that it obtained servers with NVIDIA GB300 chips and reached GB300 hardware in Thailand. He drew a careful line: ordinary distillation to build smaller, cheaper models is legitimate and valuable, but covert industrial-scale distillation aimed at stealing proprietary U.S. technology is not.

Two caveats belong next to that allegation. Kratsios published no supporting technical evidence, no prompts, model fingerprints, watermarks or account infrastructure. And the timing is awkward: Fable 5 only returned to public availability on 1 July, and K3 shipped on 16 July, a fifteen-day window that is hard to square with distillation explaining K3’s overall capability. The strongest public evidence remains Anthropic’s own February report, which attributed more than 3.4 million Claude exchanges to Moonshot as part of a wider campaign involving roughly 24,000 fraudulent accounts across several Chinese labs. Moonshot has not responded to the latest accusations.

The line moved in twenty-four hours

Here is where the story got genuinely interesting, and where most takes published on Friday are already out of date.

The initial framing was irresistible: hardware and open-source champions on one side, and on the other the companies whose valuations depend on selling access to closed APIs. OpenAI, Google and Anthropic were all absent from the original 25 signatures, and plenty of commentary treated that as the tell.

By Saturday that framing had collapsed. The letter’s signatories doubled to 50. OpenAI formally signed. Google publicly endorsed it, as did Elon Musk. AMD, Cisco, Cloudflare, GitHub, Block and Ollama were among the names added.

Two companies did not join: Amazon and Anthropic.

That is a much sharper story than the original one. When a coalition swells from 25 to 50 in a day and absorbs the very frontier labs it was implicitly aimed at, the interesting question stops being “who is pro-open?” and becomes “what does signing actually cost?” A letter opposing broad statutory bans on open weights is cheap to sign for almost everyone, including firms with entirely closed product lines. It is worth noting that every signature on the letter belongs to an organisation that profits if its arguments win. NVIDIA sells more silicon when more people train and run their own models. Meta benefits when Llama pressures closed competitors. Y Combinator needs cheap infrastructure for its portfolio. Self-interest is normal in technology policy; it just shouldn’t be mistaken for principle.

What this means for the rest of us

Why should a freelancer in Dhaka, a university researcher, or a small digital agency care about a lobbying fight in Washington?

Because it affects everything.

Independent developers and small startups are not building the future on enterprise API plans. They are surviving, experimenting and innovating on free tiers and open-weight models. If regulators restrict open weights under the banner of security, the large conglomerates will comfortably afford to pay each other for API access. Everyone else gets priced out.

Huang has argued that roughly one in four tokens generated today already comes from open models. Whatever the precise figure, it describes a decentralised ecosystem that a broad clampdown would not so much secure as relocate. China’s DeepSeek, Qwen and Kimi models do not disappear because American firms stop publishing weights. Developers route around blocked tools, foreign labs keep shipping, and domestic builders are left asking why their own government made the cheaper path harder to reach.

There is also a hardware dimension that rarely makes the policy summaries. Running capable models locally means owning or renting real compute, which is why this debate keeps colliding with the infrastructure story unfolding inside AI data centres. Access to weights without access to silicon is a thinner freedom than it sounds.

We are watching the biggest battle in AI history unfold. It is no longer just a race to build the smartest model. It is a fight over whose hands those models are allowed to stay in, and this week demonstrated how quickly the sides can rearrange themselves.