Open-Weights AI: The Big Directional Fight, With a Pinch of Hypocrisy

Set Trending Topics as a preferred source on Google.

Nvidia, Microsoft and Meta are warning Washington against restrictions on open AI models. Anthropic and Amazon didn’t sign. What looks like a fight over ideology is a fight over business models – and over a Chinese ecosystem that is currently destroying American pricing power.

When Jensen Huang posted on X for the very first time on July 24, 2026, it wasn’t about chips. The Nvidia CEO shared a three-page letter titled “Open Weights and American AI Leadership“ initially signed by 25 companies and organisations: Nvidia, Microsoft, Meta, IBM, Dell, Palantir, ServiceNow, CrowdStrike, Perplexity, Replit, Mistral, Hugging Face, Mozilla, Andreessen Horowitz, Y Combinator and the Linux Foundation. The message to US policymakers: no “premature restrictions” on models whose weights can be freely downloaded – that would stifle competition or push innovation overseas.

The letter’s core claim is that America’s AI leadership won’t be decided by a single frontier model, but by whether an open ecosystem diffuses into every sector of the economy. And: betting exclusively on closed models is not automatically safer – those can be breached or misused too, only without outsiders being able to detect it.

Within a day the list doubled to roughly 50 signatories. OpenAI and Google, absent at launch, joined later, along with AMD, Cisco, Cohere, GitHub, Cloudflare and Block. Satya Nadella, Mark Zuckerberg, Sundar Pichai, Sam Altman and Elon Musk publicly backed the initiative.

Who didn’t sign: Anthropic and Amazon.

Who stands where – and why

The battle lines track the balance sheets almost exactly.

Nvidia sells compute. The more models are freely available and the more companies self-host them, the more GPUs are needed. Huang put a number on the shift at CES 2026: one in every four tokens generated today comes from an open model. In his X post he wrote that the world needs “both frontier closed models and frontier open models.”

Microsoft and Meta likewise have no core business at risk. Microsoft makes its money on Azure and on software, not on selling access to a single frontier model – and has been working to reduce its dependence on any one model provider anyway. Meta has bet on open weights with Llama for years, because its business model is advertising, not API revenue.

Startups and VCs – a16z, Y Combinator, and the nearly 200 firms behind the newly formed Little Tech Association – make the economic argument: open models are cheaper, can be run on your own hardware, and can be fine-tuned. On their reading, banning Chinese models would achieve one thing above all – cementing the market power of OpenAI and Anthropic by government decree.

Anthropic holds the clearest opposing position. CEO Dario Amodei has long argued that released weights cannot be recalled: you can’t revoke access, you can’t update guardrails after the fact, you can’t prevent misuse. Precisely what supporters see as the strength of open weights, Anthropic treats as the risk. The company has also lobbied actively in Washington for tighter chip export controls on China. OpenAI sits in between: it has shipped open models of its own with gpt-oss, while echoing the administration’s warnings about distillation by Chinese labs. According to Axios, the two rivals have converged in Washington on exactly this issue.

The hypocrisy charge – pointing both ways

The sharpest critique of the letter came from an Austrian. Julian Schrittwieser, member of technical staff at Anthropic and previously at Google DeepMind, where he worked on AlphaGo, AlphaZero and MuZero, responded to Huang’s post with a dig: he was delighted that Jensen Huang now believed in open source, and was looking forward to the “CUDA and GPU driver open source release.” Addressing Nadella, he added that Windows and MS Office could presumably follow.

Schrittwieser subsequently clarified that he doesn’t want open models banned – he finds them useful. What he found notable was that companies with a long history of hostility to open source were suddenly all in favour of openness.

The point is hard to argue away. Nvidia organised the letter and simultaneously controls CUDA, the industry’s most proprietary software layer. Microsoft is not opening any operating systems. Meta’s Llama licence contains usage restrictions and, according to legal analysis from the EU AI Office’s advisory group, does not meet the AI Act’s open-source standard. Openness costs these companies nothing where they are demanding it.

The counter came quickly, among others from investor Bill Gurley: Anthropic’s concerns happen to arise exactly where open models threaten its own business model. Both readings can be true at once – a sincere safety position and a hard commercial interest are not mutually exclusive.

One point of terminology belongs here: “open weight” is not “open source.” What gets published are the trained parameters, not the training data or the training pipeline. Under the Open Source Initiative’s definition, most of these models don’t qualify.

China’s rise in open models

The trigger for the current escalation came from Beijing. On July 16, Moonshot AI released Kimi K3, a model with 2.8 trillion parameters; the weights are due on July 27. In independent evaluations it landed third on the Artificial Analysis Intelligence Index, second on Vals AI, and first in the Frontend Code Arena – ahead of Claude Opus 4.8 and GPT-5.5, behind Claude Fable 5 and GPT-5.6 Sol. Moonshot itself is more measured, saying it still trails the strongest proprietary models. Its API price is reportedly less than a third of Fable 5’s.

Shortly before that, Z.ai released GLM-5.2 under an MIT licence, claiming 62.1% on SWE-bench Pro against 58.6% for GPT-5.5 – a vendor-reported figure that varies by benchmark and agent harness.

This is a pattern rather than a one-off, and it started with DeepSeek in 2025. Chinese models accounted for 41% of all model downloads on Hugging Face over the past year. On the OpenRouter marketplace, Chinese open-weight models – from Tencent, Xiaomi, DeepSeek, MiniMax and Z.ai – hold the top five spots by weekly token usage. In a speech in July, Xi Jinping explicitly positioned China as a champion of open access to AI as a global public good.

Why this hurts US business models

The threat is commercial rather than technical. If companies can handle 90 to 95% of their queries with a self-hosted model that costs a fraction as much, the addressable market for premium APIs shrinks to a remainder. Mozilla CTO Raffi Krikorian told Axios that using frontier models for everyday work is like driving a Ferrari to the grocery store.

For OpenAI and Anthropic, both heading toward IPOs, a great deal rests on the assumption that frontier AI stays scarce, indispensable and high-margin. A price collapse in “good enough” intelligence hits that assumption directly.

There’s a further argument, and it comes from a security incident. After the Hugging Face hack in July, forensic analysis initially failed because of the frontier models’ own safety mechanisms – Anthropic’s Fable 5 couldn’t tell whether an attacker or a defender was asking it to analyse attack code. Hugging Face switched to China’s GLM-5.2, ran it on its own hardware, and says it reconstructed the incident in hours rather than days. Cheaper, self-hostable, no data leaving the building: three reasons US firms reach for Chinese models regardless of geopolitics.

Which governments want restrictions

United States. According to Axios, the Trump administration has been weighing restrictions up to an outright ban on Chinese open-weight models for some time; one source close to the administration says leading AI labs or their allies lobby for such bans every three to five months. It got concrete in July: OSTP director Michael Kratsios accused Moonshot AI of running an internal platform for large-scale distillation against US models while deliberately evading detection; Treasury Secretary Scott Bessent said sanctions and Entity List designations were on the table. Trade Representative Jamieson Greer treats distillation as IP theft. Notably, on June 12 the Commerce Department had already imposed a worldwide licence requirement on the public distribution of Anthropic’s Claude Mythos 5 and Fable 5 – meaning the controls first hit an American company.

Pushback is coming from Congress itself: Representatives Liccardo, Obernolte, Franklin and Lieu wrote to Commerce Secretary Lutnick in June opposing blanket restrictions.

China. According to Reuters and the Financial Times, Beijing is considering limiting overseas access to its best models – export controls on its own open weights. The logic mirrors Washington’s, with the signs reversed: the more capable the models, the less appealing it becomes to put them on the internet unfiltered.

United Kingdom. The AI Security Institute published research in July finding that leading open-weight models trail frontier models by only four to seven months on cyber capabilities, depending on the measure. The report covered models from mid-April – Kimi K3 has likely narrowed that further.

EU. The AI Act takes a different route: Article 53(2) provides exemptions for models under free and open licences, but these don’t apply to models with systemic risk above the compute threshold. The Commission’s enforcement powers – information requests, model access, recall – take effect on August 2, 2026. For European providers such as Mistral, open weights double as a sales pitch: sovereignty, on-premise deployment, auditability.

The open question: how much does distillation explain?

Even inside the US industry it’s contested whether Chinese progress rests primarily on scraped model outputs. Dean Ball, head of strategic futures at OpenAI, wrote that Kimi K3’s performance can’t be explained away by distillation. AI researcher Nathan Lambert also considers the effect overstated and argues that distillation is becoming less important in modern training pipelines. Moonshot AI has not publicly responded to the allegations.

The enforcement problem remains unsolved as well: you can stop a chip shipment, but not a weights file that has already been downloaded, mirrored and deployed locally.

The Kubernetes analogy

One framing widely cited in the debate comes from Tobi Knaup, co-founder of Mesosphere. He draws the parallel to Kubernetes: it didn’t win because its repository was public, but because it became a neutral substrate that engineers, cloud providers and enterprise vendors could all build on. Once an open, customisable platform becomes the industry’s centre of gravity, he argues, no single vendor can match the ecosystem’s combined rate of innovation.

Knaup flags the limits of the comparison himself: with models, improvements don’t flow back into a shared upstream project, frontier weights still require expensive hardware, and there is no neutral governance body for AI along CNCF lines. His policy conclusion: instead of a ban, independent testing and standards – plus American models released openly at comparable capability.