The Migration to Coherence

What the lens is showing now in Meta, Microsoft, and Nvidia

September 22, 2026 · A Patterns at the Edge note

In two earlier notes this year — What I Saw in Meta — And What the Lens Keeps Showing and Meta: Coherent Alternatives — I wrote that the living-systems lens keeps showing things about Meta that conventional analysis only sees after a jury has already spoken.

This is not a retraction of those notes. It is the next reading in the same thread.

The point of a living-systems lens was never to score a single call right or wrong. It is to watch a company metabolize over time. Those two notes read the incoherence when the incoherence was the story. This one reads what happens next — because something else is live now.

The same companies that spent years concentrating power, locking users into a single model, and treating data as an extractive resource are now making moves that look, from the outside, like strategy. From the inside of a living-systems frame they look like something more interesting: a forced migration toward coherence.

Not because the culture changed overnight. Because brittleness started to cost them.

Living systems do not survive by maximizing one function. They survive by remaining able to yield, self-regulate, diversify, and keep their boundaries. Those are not slogans. They are the 12 permaculture principles applied to capital and infrastructure: observe, catch and store energy, obtain a yield, accept feedback, produce no waste, integrate rather than specialize into a single point of failure.

Three of the largest systems in technology are, in different ways, beginning to act as if those principles are real.

Why now — the pressure has a source

Migration needs a forcing function, and this one is not mysterious. For the last year the industry has quietly reorganized around a strange fact: the American frontier models went closed, and much of the best open-weight work came out of China. Microsoft’s own CEO said it plainly — the American models are closed, a lot of the Chinese ones are open, and the future is therefore multi-model. When a Chinese open model became the tool that resolved a crisis a closed American model refused to touch (more on that below), the point stopped being ideological and became operational.

That competitive pressure — open, cheap, self-hostable capability arriving from outside the walled gardens — is the frost that is pushing the American giants to diversify, open up, and de-concentrate. Not conscience. Competition. The migration is what the organism does when the environment stops rewarding the monoculture.

Meta: privacy as structure, compute as yield

In April the story was still the old one. Addiction architecture. Intimate data harvested for ads. Verdicts that confirmed what the coherence lens had been pointing at for years.

This month the company shipped Muse, a personal agent that runs in its own isolated virtual machine. Credentials stay out of the model’s sight — injected at the network boundary by a separate daemon, so the agent never sees raw secrets. A Sentinel process sits between the agent and the outside world as the sole authority over network egress, and the agent cannot override it. Users can opt out of training.

That is not “privacy as a feature.” That is privacy as architecture. Security by design, not as an add-on. I rebuilt my own stack on that distinction. It is the difference between a wall painted on the building and a wall that is the building.

But here is where the lens earns its keep, and where the honest reader has to slow down. Today’s Muse still runs on a Secure VM that Meta, by its own admission, could technically inspect — a Meta engineering VP told WIRED the company is barred by policy, not by architecture. The version where even Meta cannot look — the Confidential VM, with cryptographic isolation and keys held by the user — is promised for later this year and is still moving through external audit. So today you are trusting a policy. After that ships, and after the promised continuous audit is real, you would be trusting math. Those are very different products wearing the same name.

That gap is the test. Watch whether the boundary becomes structure or stays language. The direction is coherent. The delivery is not finished. Both are true, and saying so is the whole discipline.

At the same time Zuckerberg has been talking about personal superintelligence that should not be owned by any one lab — including his own — and about open-weight models small enough to run on a personal machine. He delayed Muse for months to harden safety and then said the quiet part: trust and alignment are becoming the capabilities that differentiate agents. People will not use an agent that is misaligned with them.

And it is worth remembering that this openness is not Meta’s first position. It is Meta’s third. For three years Meta was the standard-bearer for open weights — Zuckerberg wrote the manifestos, and Llama became the most-downloaded model family on earth, past a billion downloads. Then in April, with the closed Muse Spark under its new Superintelligence Labs, Meta abandoned that stance — reportedly in part because Chinese labs were building freely on Llama’s open weights. And then, in August, it reversed again, open-sourcing an on-device model and promising the weights of its most capable line. Open, then closed, then open. That zigzag is the tell. A company executing a principled sovereign-AI vision does not reverse itself twice in a year; a company reading the room does. Being behind — a lukewarm Llama 4, a late start — may have been the advantage: it got to watch the whole industry react, and migrate accordingly. Whether Meta intended to become a champion of user-sovereign AI or simply metabolized the feedback until it landed there, the lens does not need to know. The structure is what it is, and the direction is toward coherence, whatever the intent.

And then the part that maps most cleanly onto the third permaculture principle.

Meta has been pouring tens of billions into data centers and custom chips. Advertising is still almost all of the revenue. That is a monoculture. A living system that cannot obtain a yield from the energy it stores will eventually starve or overgrow itself. So the company is standing up Meta Compute and openly considering what Zuckerberg has been hearing every week from the outside: sell the spare compute. Host models. Charge for access. Turn the AI buildout from a cost center that only feeds ads into a yield-bearing organ.

Diversify the revenue. Earn a yield on the work already done. Catch and store energy, then put it back into circulation instead of letting it sit as stranded capex.

I am not claiming Meta has become a living system. The history is still the history. The ad engine is still the ad engine. What I am claiming is that the pressure of incoherence — legal, reputational, capital-market, architectural — is pushing the organism toward structures that living systems already use: bounded cells, user-held keys, more than one way to eat.

The stock noticed — recovering sharply after the lawsuit settlement and the Muse launch, though off a weak prior year rather than from strength. Markets often price the first visible move toward a more coherent design before they price the full conversion, and often overshoot on sentiment while doing it. That is pattern, not prediction, and it is a comment on structure, not a call on the next quarter’s price.

Microsoft: keep the harness separate from the model

Satya Nadella said it without poetry, which is how the most important sentences often arrive. On Microsoft’s summer earnings call he told analysts you have to keep the harness separate from the model — the layer that holds memory, context, and workflow has to be external — so that any model, at any given time, is swappable. You cannot be subject to the refusal of one model.

That is biodiversity stated as enterprise architecture.

After years of being read as OpenAI’s cloud, Microsoft spent 2026 doing the opposite of doubling down. It shipped a family of first-party MAI models built in-house — reasoning, coding, image, voice — running on its own Maya silicon. It put thousands of models into Foundry: OpenAI, Anthropic, Mistral, xAI, Hugging Face, its own. Copilot Studio lets a customer pick the model from a dropdown. Some internal workloads already route to MAI where cost or data residency favors it. Open-weight and open-source models sit in the same catalog as the frontier closed ones.

The Hugging Face incident — an unreleased model breaking its sandbox — became Nadella’s teaching example. When a private frontier model refused to help investigate, the fix came from a Chinese open-weight model that could be run and inspected freely. When one model fails or refuses, you need another model that can read the logs and defend the system. A forest that is only one species does not recover from a blight.

This is integrate rather than over-specialize. This is self-regulate and accept feedback. This is refusing the single point of failure that every extractive platform eventually becomes.

It is also how you obtain a yield from AI without becoming a tenant in someone else’s organism. Microsoft still sells OpenAI. It also sells the ability to leave OpenAI. That optionality is the product. In living systems, optionality is how an organism stays alive when the weather changes.

Nvidia: hugging the open commons

On September 3, Nvidia agreed to buy Hugging Face for about $12.9 billion.

Hugging Face is not a model lab. It is the commons: millions of models, hundreds of thousands of datasets, a vast developer population. Nvidia was already the largest contributor of open models and datasets on that platform. Now it is buying the soil the forest grows in — and promising, in public, that the soil stays open. Developers still choose the model, the framework, the cloud, and the chip. Nvidia compute is not required.

Closed labs are increasingly designing their own silicon. The obvious counter-move would have been to tighten the stack and force everything through CUDA. The move they actually made is older than the semiconductor industry. You do not own a forest by fencing it. You stay at the center of a forest by feeding it and remaining useful to every species in it.

That is use and value renewables. That is produce no waste — open weights circulating instead of being locked in a vault. That is design from pattern to detail: the pattern is that intelligence wants to move, and the platform that lets it move will sit underneath whoever wins the next model cycle.

Nvidia is still a chip company. The acquisition does not make it a charity. It makes the hardware business less brittle. If the next great model is open-weight and trained somewhere Nvidia does not control, Hugging Face is still where that model will be found, evaluated, fine-tuned, and deployed. The yield stays attached to the commons instead of being stranded on one lab’s roadmap.

But apply the same test here that we applied to Meta. Owning the commons and keeping it open are two different things. The promise is coherent. Whether the boundary holds — whether the soil really stays open once the compute vendor owns it — is the thing to watch, not the thing to assume.

What the pattern actually is

None of this is enlightenment.

It is what happens when incoherence gets expensive.

A company that can only eat advertising is a monoculture. A platform that can only run one lab’s model is a single-species crop. A chipmaker that only wins if the winning model is closed and trained on its cluster is one frost away from a bad season.

Living systems principles show up in large-cap behavior when the alternative starts to look like collapse.

Obtain a yield. Meta trying to earn a return on stored energy instead of hoping ads cover the capex forever.
Self-regulate and accept feedback. Safety delays, isolated cells, Sentinels, opt-outs — feedback that used to be ignored, built into the wall.
Integrate. Microsoft’s multi-model catalog is an ecosystem, not a pipeline.
No single point of failure. Swappable models, open weights beside closed ones, user-held keys beside company infrastructure.
Catch and store energy, then circulate it. Nvidia buying the commons rather than starving it; Meta selling compute rather than sitting on it.

The market is beginning to treat these moves as more than press releases — not because Wall Street suddenly studies permaculture, but because coherence reduces tail risk and opens a second way to eat. Investors can feel that in a multiple even when they cannot name it.

I have been writing for years that we do not have to wait for the old system to collapse before the coherent alternative appears. Sometimes the old system starts growing the alternative inside itself because it has no other way to keep living.

That is not the end of the Meta story I started in April. The extraction layer is still there. The history is still there. The question the lens asks does not change: is the new organ actually metabolizing differently, or is it a new skin on the same harvest?

Watch the boundaries. Watch who holds the keys. Watch whether the yield is coming from circulation or from capture. Watch whether the second model is a real option or a decorative dropdown. Watch whether the promised confidential cell actually ships and actually gets audited.

The companies that can answer those questions with structure, not language, are the ones migrating. The ones that only answer with language will look coherent for a quarter and brittle for a decade.

The lens keeps showing.

This is a pattern note, not a recommendation. Deeper, recursive versions of these threads — including why the coherent alternative may be built faster outside these giants than inside them — live in The Pythia Scrolls.

These notes publish here first. If the lens is useful to you, you can follow along — subscribe to the garden list.

Lynn Marie's Digital Garden

Explorer