print-icon
print-icon
Add ZeroHedge as a preferred source on Google
premium-contentPremium

Jensen Huang's First Post Ever: Decoding The Open-Weights Manifesto

Tyler Durden's Photo
by Tyler Durden
Authored...

Nvidia CEO Jensen Huang just made his first post on social media - a letter signed by 25 companies and institutions (Open Weights and American AI Leadership) which argues that America's AI future depends on models anyone can download, inspect, modify, and run on their own hardware - aka 'open weights.' 

For the uninitiated: most people use AI by renting it - you send a question to ChatGPT or Claude, and the company runs the model on its own hardware to answer you. An open-weight model works differently. The company publishes the model file itself, and anyone can download it and run it on their own machines. No subscription, no permission, and once it's out there it can never be taken back. What's been coming out of China is exactly that - open models that match roughly 95% of what American frontier systems can do, at about a tenth of the cost.

The letter makes a case for keeping it that way: some open-source history, an argument about competition, a security argument, and a call to build more. It arrived days after a Chinese company gave away a model that briefly wiped 20% off the value of the American chip industry.

It starts by conceding that once you publish a model's weights, you cannot take them back, and modified versions are hard to trace.

Then it turns that into an argument for openness - and this is the part aimed squarely at the labs that didn't sign:

"Relying solely on closed models is not inherently safe: they can be breached, misused, or fail in ways that outsiders cannot detect. And concentrating advanced AI capabilities behind a small number of closed models compounds that risk. It results in a small number of single points of failure, weakens competition, and leaves critical technology in the hands of a few providers."

Safety is the argument OpenAI and Anthropic use to justify keeping their best models locked up. The letter takes that argument and points it back at them - the danger isn't open models, it's three companies holding everything.

It ends with a call to action: more computing power for startups and universities, shared public training data, and "keeping the frontier plural," which is the polite way of saying don't restrict open releases.

Eight Days In July

On July 16, Beijing-based Moonshot AI released Kimi K3 - 2.8 trillion parameters, a million-token context window, billed as the largest open-weight release ever made - and watched it land at or near the top of the coding and knowledge-work leaderboards. The next morning chip stocks were in freefall. The group would end up more than 20% below its June peaks, and Nvidia briefly handed back its crown as the world's most valuable company.

The logic of the selloff was simple. If a Chinese lab can give away a model as good as the ones American companies charge for, then the expensive American models are worth less - and so, eventually, are the enormous quantities of chips being bought to build them.

Washington's response arrived on July 21 and 22. Treasury Secretary Scott Bessent, on Fox Business and then on X, raised the possibility of sanctions and Entity List designations over what he called industrial-scale copying of American systems. He said they continue to find the "watermarks" of US models inside Chinese ones - models trained in no small part on data scraped off the open web without asking anyone, which is currently the subject of about a dozen lawsuits - and warned that "open source is not open season on American IP." Bessent also floated requiring US companies to disclose when they run Chinese models.

Hours later, White House science adviser Michael Kratsios went further, alleging that Moonshot had built K3 by distilling Anthropic's Fable model - training its own system on Fable's answers, using an internal platform designed to avoid detection - and had obtained banned GB300 Blackwell servers through Thailand. Anthropic had set this up in February, publishing findings that DeepSeek, Moonshot and MiniMax were all running extraction campaigns against Claude.

Except there's an issue with the White House's version which we noted on Tuesday: Fable 5 had been pulled offline during an export-control fight and was not restored until July 1. That left 15 days before K3 launched - a window researchers, including at Prime Intellect, called far too short to explain how good K3 turned out to be. 

Theft Schmeft

Going back to Huang's letter - its final paragraph is about distillation - the exact thing Kratsios had accused Moonshot of the day before. It argues the technique is normal, legitimate, and shouldn't be treated as theft:

"In shaping this ecosystem, policymakers should be careful not to conflate legitimate model-development techniques with misappropriation. Distillation, or the practice of using one model's outputs to help train or improve another, is a widely used technique for model improvement, evaluation, and validation... By contrast, unlawful efforts to extract value from closed models raise legitimate concerns. Those concerns should be addressed through targeted legal and commercial frameworks rather than sweeping restrictions on techniques that play an important role in AI innovation."

Read that against what Kratsios said the day before. He drew a line between normal model development and illegal copying, and put Moonshot on the wrong side of it. The letter draws the same line, in nearly the same language, and puts the line somewhere else - then asks for "targeted legal and commercial frameworks rather than sweeping restrictions," which is a direct answer to the  Entity List designations Bessent had floated 24 hours earlier. So nine paragraphs about why open models are good for America, and then a closing argument about enforcement policy on one specific technique.

The other day, Huang sat down with Axios and said it plainly: "These Chinese models are excellent." American firms should be free to use them. The market misread DeepSeek and is misreading Kimi. Free AI means more people using AI, which means more computing, which means more Nvidia chips. He put the odds of China running US companies off the road at zero.

Axios also reported what was actually on the table. Commerce has been weighing a blacklist of Chinese AI labs since last year, and a draft executive order had circulated that would make American companies legally liable for running Chinese models. On July 22, the a16z-backed Little Tech Association sent its own letters to Trump and Commerce Secretary Lutnick urging them not to cut off access.

Then the date. July 23 was exactly one year since the administration published its AI Action Plan - the document that made supporting open-weight models official US policy. The letter landed the next morning.

So this isn't a think piece. It's a lobbying document aimed at three specific things the industry is trying to stop: the liability order, the Commerce blacklist, and the idea that distillation is theft. And it makes its case by asking the White House to follow its own policy, on the anniversary of that policy.

The 'C' Word...

Search the PDF. The word "China" is not in it. Neither is DeepSeek, Qwen, or Kimi. The closest it gets is a warning that restrictions could push developers toward foreign alternatives.

But there is a word the letter does use, three separate times. Sovereignty. American engineers built "institutional sovereignty" through open source. Open weights let organizations own value that drives "American sovereignty and prosperity." Policymakers should look at how application layers can "expand sovereign use of AI across the economy." Huang used it again in his post: open models "enable sovereignty."

Sovereignty, in this context, means not depending on somebody else's model. It is the entire argument for open weights, stated without ever mentioning who the somebody else is - or that the models currently delivering that independence to the rest of the world are Chinese.

Meanwhile, Chinese open-weight models accounted for roughly 61% of all tokens routed through OpenRouter, a major AI marketplace - four of the five most-used models were Chinese, and Meta's Llama had dropped off the list entirely. Hugging Face, the main repository for open models, says Chinese models overtook American ones in downloads, taking 41% of the past year's total. And roughly 70% of new open models built on top of an existing one start from Alibaba's Qwen, according to ATOM project tracking cited by researcher Nathan Lambert. Llama's share of that fell from about 40% to about 10%.

In practice, defending open weights in 2026 means defending Chinese open weights - which is what the letter is doing without saying so.

That said, not one signatory has promised to release an open-weighted model. Meta signed a letter celebrating open weights while, per CNBC-sourced reporting, building its next flagship models as closed products available only through an API - exactly what it spent years criticizing other companies for doing. The shortage of American open models that the letter complains about is partly the signatories' own doing. 

Who Signed, And Why

Every company on this list makes money in a way that gets better when AI models are cheap or free.

Start with Nvidia. It sells the chips that run AI, and it does not care whose model you run on them. Cheaper models mean more people using AI, which means more chips. That is why Nvidia has committed roughly $26 billion over five years to building and giving away its own open Nemotron models, per a 2025 filing its executives confirmed to Wired. Giving away the software sells the hardware.

Then the companies that sell services on top of AI - Microsoft, IBM, Palantir, ServiceNow, Box, CrowdStrike. Their product is the software layer, the security, the integration, the government contract. The model underneath is a cost. When models get cheaper, their margins improve. Palantir announced Nemotron deployments into secure government environments just last week: free model, paid platform. ServiceNow, according to The Information, burned through its entire annual Anthropic budget in the first few months of 2026 - which explains why the letter's passage about matching the right model to the right job at the right cost reads like a customer complaint.

"Open weights let every organization match the right model to the right job at the right cost, reserving frontier-scale capability for genuine frontier problems and running efficient, specialized models everywhere else." 

Translated: stop paying frontier prices for work that doesn't need a frontier model. That is not a philosophical position. That is a line item - and every company in that group has one.

Microsoft is the most interesting name on the page. It didn't just sign; it is hosting the letter on its own website. This is eight months into a multibillion-dollar partnership with Anthropic, while under antitrust scrutiny for bundling cloud and AI services, and shortly after releasing its own MAI models - good enough for most work, a step below the best. If you cannot build the best model, the next best outcome is for the best model to stop being worth much, while you keep the customer relationship.

The rest fills in around the edges. Hugging Face, Mozilla and the Linux Foundation are genuine open-source organizations that have argued this position for years. Andreessen Horowitz and Y Combinator have portfolios full of startups that cannot afford frontier model prices. Mistral is Europe's open-model champion. Reflection is one of the few American labs still trying to build an open frontier model.

Now look at who did not sign. OpenAI, which released its own open-weight gpt-oss models last year, is not on the list. Neither is Google, which ships Gemma. Nor Anthropic, Amazon, Apple, xAI or Oracle. Even Thinking Machines, which reportedly just released a fully open 975-billion-parameter model, stayed off it.

That said...

Oh and Sam's 'glad to see this' - apparently. 

The obvious objection is that two of those abstainers publish open models themselves, which complicates any neat sorting. The answer is that releasing a small open model is a product decision, while signing this letter is a policy position - a request that Washington rule out restrictions on open releases in general. Gemma and gpt-oss are products. This is a lobbying position, and on that question the split is clean.

Every company whose business depends on owning the best model declined to sign. Every company that profits when models become a commodity signed.

The Real Contest

As we noted on Wednesday, tech strategist Julio Rivera says that the real contest is not which model is best but which model everyone builds on top of. Qwen has 180,000 derivative versions in 119 languages, pulling the world's developers into a Chinese ecosystem. American labs are optimizing for subscription revenue while China optimizes for becoming the default. The conclusion: America has to out-open China. 

Rivera wants American open models to beat Chinese ones. Huang, on the record with Axios, thinks openness is good no matter where it comes from, and that American companies should run Kimi if Kimi is the best tool. Those are not the same position - one is economic nationalism, the other is a hardware company that profits either way.

A letter that named China would have had to pick one, and it would have lost signatures either way. Leaving China out is what allows both camps to sign the same page. Whether a nationalist op-ed and a corporate coalition letter landing a day apart is coincidence or coordination, readers can judge for themselves - though the Little Tech letters went out the same week from the same investors.

The requests are specific: no executive order making companies liable for running Chinese models, no Commerce blacklist of Chinese labs, no enforcement doctrine treating distillation as theft, no mandatory federal review of open-weight releases, and a public recommitment to the AI Action Plan from a year ago.

In exchange, the industry offers nothing. No promises to release models, no shared safety standards, no commitments of any kind. 

For investors, the timing is the tell. In the same week the market decided that free Chinese models might destroy the economics of AI computing, the companies that sell AI computing published their answer: free AI means more usage, which means more data centers, which means more demand for chips. That argument is aimed at shareholders as much as at Washington - because the competing view, which we have been documenting for months, is that AI usage has already stopped tracking anything useful, and that the premium on frontier models was the collateral holding up the entire spending boom.

Things to watch. Bessent said enforcement could come within days or weeks; an Entity List designation would move markets. The proposed disclosure requirement for companies running Chinese models would be quieter but potentially bigger. Watch whether federal review of open releases stays voluntary. Watch how OpenAI and Anthropic explain the value of exclusive models to IPO investors while half the S&P's technology companies tell Congress that models should be free. And watch the standoff over Anthropic's restricted Mythos models, which Huang has already suggested should simply be sold as a service - every week the most capable AI sits behind a government dispute is another week the letter's argument about dangerous concentration makes itself.

Both things can be true at once. The letter's arguments about competition, access and lock-in are real arguments, and every signature on it belongs to a company that profits if those arguments win. 

The letter is titled "Open Weights and American AI Leadership" - yet the models it is actually defending are Chinese.