TL;DR. On 24 July, Satya Nadella, Jensen Huang, and thirty-plus other technology leaders signed an open letter in support of open-weight AI models. Roughly $8 trillion of market capitalisation, aligned around a specific future for AI. Packy McCormick’s take at Not Boring frames it correctly as commoditising the complement: everyone who signed benefits when the model layer is competitive. What that argument misses is quieter and, for a small class of products, structural. Open-weight is not just good for consumer choice at the model layer. It is the necessary condition for personal AI to run on your own machine at all. Yesterday’s letter is the supply-chain agreement Ostler quietly needed.

On 24 July 2026, Satya Nadella and Jensen Huang each signed and shared an open letter in support of open-weight AI models. Jensen even set up an X account for it. Alongside them: leaders at Meta, Hugging Face, Dell, CrowdStrike, Box, ServiceNow, Replit, Perplexity, Google (a late but notable addition), Y Combinator, a16z, Palantir, Mariana Minerals, and more. Add up the market caps and you land somewhere around eight trillion dollars.

Packy McCormick wrote it up beautifully, and if you like watching technology strategy played at Super Bowl level, read his piece. His frame is Joel Spolsky’s Strategy Letter V: smart companies try to commoditise their products’ complements. Models run on Nvidia chips. Models power Microsoft products. Models let a16z-backed and YC-backed companies exist. Meta is already the most open of the American labs. Hugging Face is where open models live. The more competition there is at the model layer, the less any single lab captures the value, and the more the complementary businesses win. Signing the letter is a selfish act that happens to align with the interests of consumers and millions of businesses. That is capitalism working the way it is supposed to.

Packy is right about all of that. He is writing about the winners and losers in the industry.

This post is about what the letter means for a much narrower class of software: the one Ostler is in.

Open-weight is not a nice-to-have for us. It is the prerequisite.

Ostler is a personal AI that runs entirely on your Mac. Not “runs on your Mac with a cloud fallback”. Not “runs on your Mac and syncs to our servers”. On your Mac. The model that answers your questions is a file on your disk, and the inference happens on your machine’s neural cores. Pull the internet cable out and everything keeps working.

You can only build a product like that if the models you need can actually be run on customer hardware. That, in turn, requires exactly one thing: the model weights have to be available to be downloaded and executed locally. If the weights live only on someone else’s GPUs, everything downstream of the model layer is a chatbot with someone else’s hand on the tap.

Today, Ostler ships with Qwen 3 (Apache 2.0) as the reasoning model and nomic-embed-text (Apache 2.0) as the embedding model. Ollama, also Apache 2.0, is the runtime that loads them. Whisper (MIT) does speech transcription. Every one of those pieces is on your disk after installation, and none of them phones home. That entire stack is only possible because the maintainers of those models chose to publish the weights. And the maintainers of those models chose to publish the weights partly because a handful of platform companies decided it was in their strategic interest to keep the ecosystem open.

The letter is $8 trillion of tech saying, out loud, that they intend to keep the ecosystem open. Not out of virtue. Out of aligned self-interest. And it happens to be exactly the ecosystem Ostler needs to exist.

Your assistant’s brain gets better and cheaper. For free. Every quarter.

The second thing the letter locks in is a trajectory the customer benefits from silently.

Because the model layer is competitive, and because Nvidia and Microsoft and Google are pouring resources into making the open frontier keep pace with the closed frontier, the free open-weight models keep getting better. Qwen 2 was decent. Qwen 3 is good enough that most customers on a modern Mac cannot tell the difference between it and a hosted product for most personal-AI tasks. Qwen 4, or whatever succeeds it, will be better again. This happens without Ostler doing anything except pointing at the new weights.

The economic shape of this matters. Cloud AI subscriptions are priced today as if compute were a loss leader. Industry commentary increasingly expects those prices to rise once enough users are habituated to a given product. Every one of those subscription-price rises is a bill the cloud AI customer pays. The local-AI customer’s bill for “a better model shipped this quarter” is zero. It is structurally zero, because the model is a file, and downloading a new file to your Mac costs the customer nothing per query, today or ever.

Yesterday’s letter is $8 trillion voting for the compute-tap to stay open. That is a very useful bill to have paid on your behalf.

Nobody can pull the ladder up.

Every “we don’t train on your data” promise in the AI industry is a policy. Policies can be changed. Startups get acquired. Governments ask nicely, then loudly. Terms of service get rewritten in a quarterly memo you were sent an email about but did not read.

Open-weight models change the failure mode. If a single lab tightens its terms, the customer of a local product can swap the model out for another one. If a whole ecosystem of $8 trillion of platform incumbents is aligned on keeping open-weight alive, the customer of a local product can trust that the swap will always be possible. There will always be another Qwen. There will always be another Llama, another Mistral, another Gemma-shaped alternative from a lab that wants to keep the ecosystem competitive.

Ostler puts this in the strongest form we can. A policy letter can kill a closed model. It cannot kill the assistant on your Mac. The one on your Mac keeps running exactly as it did the day you installed it, because there is nothing between it and you that can be quietly changed under your feet. What could have happened is that the open-weight ecosystem starved and Ostler would eventually have run out of good models to point at. Yesterday made that scenario materially less likely.

What about the safety concern?

There is a genuinely held view, which Packy’s Matt Kaufman letter articulates well, that open weights are dangerous because they can be fine-tuned to remove safety measures. It is a real concern and worth engaging with, not just dismissing.

The honest answer, from where Ostler sits, is that the same argument would justify tightly regulating operating systems, compilers, and web browsers. Every one of those is a general-purpose capability that can be misused. We have not concluded, as a society, that we should therefore make them accessible only through a licensed cloud provider. We have decided that the value of general-purpose capability in the hands of individuals is enormous, and that the correct place to intervene against misuse is at the deployment layer, not the toolchain.

The bet Nadella and Huang and the other signatories have placed is that AI is closer to a general-purpose capability, and that consumers and enterprises should have direct access. It is the same bet we are placing when we ship an on-device model with Ostler. It is not a certain bet. It is a considered one, and it is one the customer benefits from every time a better open-weight model lands.

The close

The world does revolve around you.™ That claim requires that the intelligence which knows your world lives on the machine you own. Which requires that the intelligence is available to live on the machine you own. Which requires that someone, somewhere, keeps publishing the weights.

Yesterday, $8 trillion of aligned corporate self-interest signed a letter promising to keep doing exactly that. Ostler runs on the models that letter defends. We wrote nothing yesterday. Our supply chain got safer.

Questions, corrections, disagreements – [email protected].