The Open-Weight 'Letter' Is a Lobbying War

On July 24, 2026, Jensen Huang made his first-ever post on X. It was not a selfie in a leather jacket or a quarterly earnings flex. It was a three-page policy letter — co-signed by NVIDIA, Meta, Microsoft, IBM, and 21 other companies — arguing that open-weight AI models are essential to American competitiveness. Within 48 hours, the signatory list doubled to 50, adding OpenAI, Google, AMD, Cisco, and Cloudflare. Sam Altman endorsed it publicly. Yann LeCun amplified it. The amplification was staggering — Soumith Chintala's repost alone pulled 170,000 likes, the single biggest engagement item on X that day.

📖 Read the full version with charts and embedded sources on The Arc of Power →

View original post on X →

It looked like a manifesto. It read like a movement. But read the lobbying disclosures filed the same week, and a different picture emerges: one where the companies signing the letter and the companies spending millions to shape the rules are playing the same game from opposite sides of the same table.

This is not a debate about principles. It is a coalition war over who writes the regulatory framework for the most consequential technology since the internet.

View original post on X →

Three Lessons from the Open-Weight Coalition

1. The Letter Protects the Supply Chain, Not the Ideal

The open-weight letter's argument is straightforward: restricting downloadable AI models would harm American competitiveness, weaken cybersecurity, and concentrate power among a few proprietary vendors. Every word of that is defensible. But the composition of the coalition tells you what is actually being defended.

NVIDIA makes GPUs. Every open-weight model downloaded, fine-tuned, and deployed on-premise requires NVIDIA hardware. Meta distributes Llama because it eliminates API dependency on competitors. Microsoft backs open weights because Azure hosts them. The letter advocates openness in model weights — as Forbes noted, "the one layer where almost no signatory keeps its moat" — while every signatory protects its own proprietary advantage elsewhere. NVIDIA controls CUDA. Microsoft controls Azure. Meta controls the social graph.

This is not hypocrisy. It is rational coalition-building. Each signatory has identified open weights as the market structure that maximizes its own revenue, and the letter gives that position the vocabulary of public interest.

The question is not whether the letter is correct — it mostly is. The question is why it appeared now, in this form, signed by this particular set of companies.

2. Follow the Money: The Lobbying Spend Is the Real Mechanism

The letter is theater. The lobbying spend is the mechanism.

Federal disclosures released the same week reveal that AI firms have broken lobbying records in H1 2026. Anthropic nearly tripled its spending to $3.53 million — already surpassing its entire 2025 total of $3.1 million. OpenAI nearly doubled to $2.22 million. Together, these two closed-model labs have spent $5.75 million in six months to shape AI policy in Washington.

View on Hacker News →

The Hacker News thread on the Financial Times report captured the reaction with characteristic precision: "Never ceases to amaze me how cheap lobbying is. That's pocket change for these companies." And that is precisely the point. At the scale these companies operate — Anthropic's latest funding round was $2 billion, OpenAI's $6.6 billion — lobbying spend is a rounding error that buys regulatory architecture.

What are they lobbying for? Their filings list cybersecurity, copyright, cloud computing, and defense procurement. But the subtext is structural: reports indicate that representatives from both OpenAI and Anthropic have urged key federal figures — including Treasury Secretary Scott Bessent and White House technology adviser Michael Kratsios — to restrict accessible AI models.

⚠️ The uncomfortable math: the two most prominent absent names on the open-weight letter — Anthropic and Amazon (Anthropic's largest investor) — are spending more on DC lobbying than most of the letter's signatories combined. The letter has 50 signatures. The lobbying spend has $5.75 million. In Washington, dollars vote louder than names on a PDF.

Jeremy Howard, the fast.ai founder and one of the most credible voices in the open-source AI community, put the contradiction in terms that were impossible to misread: "Hey sir. We are not asking you to open source Anthropic. Just don't lobby the government to shut down open source."

View original post on X →

3. China Is the External Threat That Makes the Coalition Viable

No political coalition forms without a shared enemy, and the open-weight letter has a convenient one: Beijing.

The letter arrived three days after Moonshot AI released the weights for Kimi K3 — a 2.8-trillion-parameter model that is now the largest open-weight system publicly available. On the same day the weights dropped, it hit 1,031 points on Hacker News with 417 comments, and blind arena evaluations showed it outperforming most U.S. models on front-end coding tasks. Polymarket prices Moonshot at 90% on the WebDev Arena benchmark — even as it sits at just 11% for "best Chinese AI company" overall (Alibaba's Qwen holds 88%).

As we analyzed when K3 first launched, this is not merely a product release. It is standards warfare. Xi Jinping used the World Artificial Intelligence Conference to announce WAICO (the World AI Cooperation Organization), headquartered in Shanghai, with founding members spanning BRICS, ASEAN, the African Union, and the Arab League. The 130+ nations outside Washington's orbit need AI capacity they cannot build themselves — and China is offering to give it to them for free.

Nathan Lambert's analysis in Interconnects frames the escalation clearly: the frontier performance gap between open and closed models has narrowed from 6-9 months to roughly 3-5 months. Chinese labs appear to achieve better capital efficiency, with Moonshot reporting 2.5x improvement in overall scaling efficiency versus its predecessor.

This creates a paradox for Washington: restricting open weights does not prevent China from distributing them. It only prevents American companies and American cybersecurity teams from accessing the best available open models. As LeCun argued in his amplification of the letter: "Attackers have frontier AI. Defenders need a frontier AI ecosystem — the best open and closed models, force-multiplied by a global community."

View original post on X →

The Real Map of the War

Strip away the rhetoric and the coalitions become legible:

Team Open-Weight (the letter signatories): NVIDIA, Meta, Microsoft, AMD, Dell, IBM, Palantir, Hugging Face, a16z, Y Combinator — plus late additions OpenAI and Google. Their business models benefit from distributed inference on commodity hardware. They want regulation that permits open distribution.

Team Closed-Model (the lobbying spenders): Anthropic ($3.53M H1), OpenAI ($2.22M H1). Their business models depend on API access as the primary distribution channel. They want regulation that creates compliance barriers for open distribution — framed, naturally, as safety requirements.

Team China (the external catalyst): Moonshot/Kimi K3, Alibaba/Qwen, DeepSeek. Their releases make the open-weight argument politically viable for Team Open-Weight while simultaneously creating the security concern that Team Closed-Model uses to justify restrictions.

ℹ️ The irony: OpenAI signed the letter AND is spending $2.22M lobbying for restrictions on open models. This is not a contradiction — it is hedging. Sign the letter (cost: nothing) to avoid being seen as anti-openness. Spend $2.22M (cost: pocket change) to shape the actual rules. In Washington, you can hold both positions simultaneously because the letter has no legal force and the lobbying does.

The Commerce Department has reportedly considered multiple restriction approaches: adding Chinese labs to entity lists, implementing liability frameworks for U.S. companies hosting Chinese models, and circulating draft supply-chain security rules. Each approach would differently advantage the closed-model or open-weight faction. The lobbying spend is calibrated to influence which approach prevails.

The Deeper Power Dynamic: NVIDIA's Leverage Play

One detail in The Register's analysis deserves special attention. NVIDIA has reportedly committed $250 billion in future orders and infrastructure guarantees to U.S. AI companies — including substantial commitments to OpenAI itself. The subtext is not subtle: NVIDIA controls the compute supply chain that every AI company depends on, and it has chosen to use that leverage on behalf of open weights.

This is the kind of structural power that makes the letter more than a PDF. When the company that manufactures the hardware everyone needs takes a public position on model distribution policy, and has the capital commitments to back it up, the letter becomes a pressure campaign with actual teeth.

Jensen Huang's choice to make this his first X post — rather than a product announcement or earnings highlight — signals that NVIDIA views the regulatory fight over open weights as a strategic priority equal to or greater than any product cycle. NVIDIA's future revenue depends on the maximum possible number of organizations training, fine-tuning, and deploying models. Restrictions on open weights would concentrate inference on a handful of API providers, reducing the total addressable market for NVIDIA hardware.

Contrarian Corner: What If the Letter Is Right Despite Itself?

⚠️ Here is the uncomfortable truth for critics of the open-weight coalition: the letter's arguments are mostly correct, even if its signatories' motives are entirely self-interested.

Restricting open weights in the United States would not prevent Chinese labs from distributing their models globally. Kimi K3's weights are already on Hugging Face. They cannot be recalled. A ban would accomplish one thing: preventing American organizations from legally deploying the best available open models for cybersecurity, research, and enterprise applications — while adversaries face no such restriction.

The distillation paradox compounds this: closed-model APIs have already been extensively used to train and distill open-weight alternatives. The knowledge has diffused. Banning downloads now is like banning photocopiers after the documents have already been copied.

The safety argument against open weights — that released weights cannot be recalled and could be fine-tuned for misuse — is technically accurate but strategically incomplete. As we analyzed in our weights embargo coverage, the asymmetry runs the wrong way: every restriction on defensive use of frontier models increases the relative advantage of attackers who face no such restrictions.

What Comes Next

Three predictions:

First, the letter will win the PR battle and lose the lobbying war. Fifty signatures make good headlines. $5.75 million in targeted spend shapes actual regulation. Expect a "compromise" framework that nominally supports open weights while imposing compliance requirements — liability frameworks, safety evaluations before release, export controls on specific architectures — that create structural advantages for well-capitalized closed-model labs.

Second, Chinese open-weight releases will accelerate, not slow down. Kimi K3 is not the endpoint; it is the proof of concept. Beijing has identified open-weight distribution as a geopolitical tool for building AI dependency across the Global South — the same strategy the U.S. used with the internet in the 1990s, now being replicated with AI model access.

Third, the real regulatory action will move to export controls on compute, not restrictions on weights. This is where NVIDIA is most exposed and most motivated. The company's support for open weights may, in part, be a strategic trade: defend open distribution of models (which drives GPU demand) while accepting tighter controls on chip exports (which it has already absorbed via existing entity-list restrictions on China).

The open-weight letter is a significant document — not because of what it says, but because of what it reveals about the structure of power in the AI industry. When 50 companies agree on policy language in 48 hours, what you are witnessing is not consensus. It is coordination. And the real question is not whether they are right about open weights. The real question is who gets to write the rules — and how much they are willing to pay for the privilege.

---

The Arc of Power analyzes the intersection of technology, geopolitics, and power dynamics. For more on the AI regulatory landscape, see our coverage of China's K3 gambit and the distillation paradox.

Originally published at The Arc of Power