Last week, around 25 technology companies put their names to a short open letter titled Open Weights and American AI Leadership. Nvidia and Microsoft led it. Meta, IBM, Dell, Palantir, Hugging Face, Mistral and Mozilla signed alongside them. Jensen Huang, who had never posted on X before, used his first message on the platform to share it, and Satya Nadella amplified it the same day.
The more telling detail is who stayed away. OpenAI, Anthropic and Google, the three companies with the strongest closed frontier models, did not sign. Read the list of names and the argument writes itself. This is an industry position from the companies that benefit when models circulate freely, and it is aimed squarely at a decision brewing in Washington.
For months there has been talk of the US government restricting open-weight AI, partly on national security grounds and partly as a response to Chinese models catching up. The immediate trigger was Moonshot AI's Kimi K3, a Chinese open-weight model released in July that its makers claim performs on par with leading American systems. Moonshot has been accused of distilling Anthropic's Fable 5 to build it, and DeepSeek made similar waves earlier with a capable model at a fraction of the usual cost. The US Treasury has since floated the idea of sanctions over model "theft".
The letter never mentions China by name, but the timing makes the intent clear. Its authors want policymakers to draw a line between legitimate model development and intellectual property misappropriation, and to go after proven misuse with targeted legal action rather than blanket restrictions on a whole class of technology. Alongside that, they ask for expanded access to compute for smaller developers and public investment in shared datasets and evaluation tools.
There is a fair amount of self-interest here, and it is worth naming. When models become interchangeable, people buy more chips, rent more cloud capacity and build more applications. Nvidia and the infrastructure providers do well in a world of commoditised models. That does not make their case wrong, but it does mean you should read the letter as a commercial argument dressed in the language of national interest, not as neutral advice.
For a CTO or head of engineering, the Washington fight matters less than the shift it points to. Open-weight models have quietly become good enough for real production work. Huang put a number on it earlier this year: roughly one in every four tokens generated today already comes from an open model. Weights that anyone can download now run on enterprise clusters, regional clouds and on-premises racks, not just a handful of API endpoints.
What this changes for you is the nature of a choice that used to be invisible. For the past two years, "which model" was effectively a default. You called an API, you paid per token, and the provider handled the rest. That default is dissolving. The model layer is becoming plural, and plural means you now have a genuine architectural decision to make rather than a box that was ticked for you.
We think the useful way to frame this is as a portfolio decision, not an allegiance. Open and closed are not teams to join. They are tools with different properties, and mature engineering organisations will run both.
Closed frontier models tend to earn their place where the work is genuinely hard and the volume is comparatively low. If you need the strongest available reasoning, the fastest path to a working feature, and you have no appetite for running infrastructure, a frontier API is often the right call. You are paying a premium for capability and convenience, and for many use cases that trade is sound.
Open-weight models make more sense as the work becomes repetitive, sensitive or large in scale. The conditions that push in their favour are fairly consistent:
The letter itself makes a version of this point when it argues that open weights let organisations reserve frontier-scale capability for genuine frontier problems and run efficient, specialised models everywhere else. On that narrow engineering claim, we agree. Most of what a business actually asks of AI is not a frontier problem. It is a large number of ordinary tasks that need to be reliable, affordable and under your control.
There is a quieter implication in all of this.
If governments start restricting which models can be used, and if disputes over how models were trained lead to sanctions, then your model supply chain now carries policy risk. A model you depend on today could become harder to access, or awkward to justify, six months from now.
This is not a reason to avoid any particular provider. It is a reason to know where your models come from, to avoid designing yourself into a corner around a single one, and to keep a working alternative in reach. The organisations that will handle the next few years well are the ones that treated the model as a component they could swap, rather than a foundation they poured concrete around.
The open letter is a lobbying document, and it should be read as one. Strip away the politics, though, and it confirms something we have been telling clients for a while. The model is no longer a settled default. It is a live design choice, with real consequences for cost, control, compliance and resilience.
The practical response is unglamorous and effective. Treat models as a portfolio. Design your systems so you can move between them without rewriting everything. Put the frontier where the frontier is needed, and put efficient open models everywhere else. Keep an eye on where your providers sit relative to policy, because that ground is moving.
If you are working out how open and closed models should sit together in your own stack, that is exactly the kind of decision we help teams make. Let's talk about how this applies to your product.
.jpeg)
