The 25-Company Open-Weight Alliance — and Why OpenAI and Anthropic Refused to Sign
Yesterday, twenty-five of the most powerful companies in Silicon Valley put their names to a single piece of paper, and the real story isn’t what’s on it — it’s who’s missing.
The letter, titled “Open Weights and American AI Leadership”, was published on July 24 and signed by NVIDIA, Microsoft, Meta, IBM, Palantir, Andreessen Horowitz, Hugging Face, Mistral AI, Mozilla, the Linux Foundation, Y Combinator, Telnyx, Perplexity, Reflection, and Replit, among others. Their collective message to Washington: don’t ban open-weight AI models. Not yet. Not without understanding what you’re breaking.
Jensen Huang, NVIDIA’s CEO, made his first-ever post on X (the platform he’s apparently never used in the ten+ years it’s existed) to champion the letter. The message: “Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.”
When Jensen Huang finally joins X and his debut post is about open-weight AI policy, the industry has reached an inflection point.
The signatories have something in common — and it’s not accidental
Every single company that signed this letter benefits directly from open-weight AI. Meta publishes Llama openly. NVIDIA sells the chips that run open models. Hugging Face’s entire business is hosting open weights. Mistral AI is an open-model company through and through. Microsoft distributes open models through Azure.
Not one of the signatories sells access to a closed frontier model.
And that’s precisely why OpenAI, Anthropic, and Google are conspicuously absent.
These are the three companies whose business model is built on keeping their best models locked behind an API. OpenAI’s whole pitch has been safety through control — you can’t misuse what you can’t download. Anthropic follows the same logic with its constitutional AI framework. Google, sitting on Gemini, has similar incentives.
The split is now formal. The closed-model camp and the open-weight camp are no longer just having different strategies — they’re on opposite sides of a policy battle.
Why now? The Kimi K3 factor
The letter didn’t emerge in a vacuum. On July 16, Beijing-based Moonshot AI unveiled Kimi K3, a 2.8 trillion-parameter Mixture-of-Experts model that the company says rivals OpenAI and Anthropic’s best. Open weights are promised for July 27 — just three days from now.
Kimi K3 isn’t just big. At 2.8 trillion parameters, it’s the largest model anyone has ever published a parameter count for. It uses a technique called Kimi Delta Attention with Attention Residuals, has native vision capabilities, and supports a one-million-token context window.
Axios reported on July 20 that the Trump administration is considering an executive order to ban Chinese open-source AI models within the US, specifically prompted by Kimi K3’s emergence. The concern is national security — what happens when any organisation, anywhere, can download a frontier model and run it locally, invisible to any regulatory oversight?
The Silicon Valley response has been two-pronged. On July 22, nearly 200 startup companies — including Proton and Y Combinator — separately urged the Trump administration not to cut off access to Chinese open-weight models [Politico]. Yesterday’s 25-company letter is the big-tech version of the same argument.
Amazon’s AGI layoffs: the deployment pivot
Running parallel to the policy debate, Amazon quietly cut jobs in its Artificial General Intelligence unit on July 22, according to Reuters. The layoffs hit the teams behind Amazon’s Nova models, and the company is reportedly closing its San Francisco AGI site.
The pivot is telling: Amazon is shifting resources away from building foundational models and toward embedding engineers directly inside enterprise customer organisations, helping them deploy AI that already exists. A $1 billion AWS effort to put engineers on-site at customer businesses.
This isn’t about abandoning AI. It’s about acknowledging that the race to build bigger models is increasingly a two-horse competition between OpenAI and Anthropic, while everyone else is figuring out how to profit from models that already exist.
The irony that nobody is talking about
From an analytical standpoint, there’s a rather neat irony here. The companies arguing most passionately for open-weight AI are the ones that make money when those weights are open — chipmakers, cloud providers, hosting platforms, and companies that publish their own open models to build ecosystems around them.
OpenAI and Anthropic, the companies that have been the most vocal about AI safety for years, are the ones whose business model depends on keeping the weights closed. You can’t fine-tune, audit, or run a closed model yourself. You have to trust the provider, subscribe to their API, and accept their terms.
Neither position is inherently wrong. But the fact that the “safety through openness” and “safety through control” camps map so neatly onto their respective business models deserves to be pointed out.
The 25-company letter makes a genuine argument — open diffusion of AI capability strengthens American leadership by building a broad ecosystem of developers, researchers, and enterprises who can build on top of shared foundations. That’s not just a business case; it’s a geopolitical one.
But the Kimi K3 question is real. When the largest open model in the world comes from a company in Beijing, the policy dilemma stops being abstract. The letter says “don’t rush into restrictions.” The White House reportedly says “we need to think about this carefully.” Both sides are right, which is exactly why the conversation is going to get much more uncomfortable over the next few months.
Sources: Reuters, Business Insider, CNBC, SiliconANGLE, Washington Post, Axios, Politico
