TL;DR. Jake Saper, General Partner at Emergence Capital (early backers of Salesforce, Veeva and Zoom), just published a thesis piece called The Open-Weight Unbundling. His argument for enterprises: the model layer is commoditising, and the durable moat is migrating to the proprietary data plus workflows built on top. He is describing, for the enterprise, the same structural shift Ostler is built on for the individual. Same inversion, different vertical. This post is what the personal version looks like.

Saper’s frame is worth reading in full, and I would encourage you to. But the short version is this. The gap between what an enterprise can run on its own hardware using open-weight models and what is available from the frontier labs behind an API has compressed to about four to six months, and it is still shrinking. Open weights now carry a quarter to half the volume across major aggregator platforms like OpenRouter and Vercel, and are used by roughly 80% of startups. Enterprises are having what Saper calls a tokenmaxxing hangover: teams paid a fortune for cloud-AI subscriptions, discovered the money mostly went on wasteful spend, and are now routing the work to cheaper open alternatives. The labs, having lost the moat at the model layer, are being forced to move upmarket into applications and services, which means competing directly with their own API customers. That in turn accelerates the shift to open, because customers can see what is happening.

The consequence Saper draws is that value in enterprise AI is migrating from model access to the proprietary data and workflows built on top of it. His winners are the vertical software companies that own the full outcome cycle (Veeva in pharma, Epic in healthcare, Harvey in legal, Together AI as the open-weight infrastructure). His losers are the frontier labs whose margins are compressed and whose customer relationships are eroding.

That is his enterprise thesis. It is a good one, and I think it is correct.

What I want to add is that the same argument, applied to individuals rather than enterprises, is even stronger. And it is what Ostler has been quietly built for.

The personal version of the same shift

An enterprise has a data moat because it holds proprietary information about pharma trials, or legal precedent, or customer support tickets, that a competitor would have to spend years and money to accumulate. That moat is real. But it is replicable in principle. A better-funded, better-positioned competitor can, over enough time, build the same corpus. Saper’s vertical-AI winners depend on being first, being embedded, and being difficult to displace. Not on being unique.

Your personal data is not like that. Nobody else can ever have your iMessages. Nobody else knows what you promised your brother in February, or the specific phrasing you used when you finally wrote the resignation email you did not send. The data that would make an AI actually useful to you is inherently non-replicable. It is the accumulated log of one life, and there is exactly one copy of it, and it belongs to you.

That is a data moat with a different structural quality. An enterprise’s moat compounds with time and market position. Your personal moat compounds with the fact that only you have lived your life. There is no funding round that can replicate it, no acquisition that can absorb it, no better product that can leapfrog it. The moat is you.

And that is exactly the moat that a personal AI has to be built on, if it is going to be worth anything at all. Which is why the same unbundling Saper describes for the enterprise has such a specific, and much sharper, shape for the individual.

Where the enterprise version and the personal version diverge

Enterprises do the unbundling on a spreadsheet. Someone in finance notices the line item labelled Anthropic has quintupled in a quarter, the CFO calls a meeting, work gets routed to a local model, done. Rational. Boring. Someone else’s money.

The personal version is not boring. It is the moment you realise you cannot use the assistant the way you actually want to use it, because every good question is metered and your credit card statement is watching. The paradox of cloud personal AI is that the better it is, the more you want to talk to it, the more it costs, and so the less you use it. You develop a little internal accountant that pipes up every time you open the chat window: is this question worth two cents? Is this one worth three? And so you ration. You save the big questions for later. You do not paste the whole document in. You get half the value because you cannot afford the whole value.

Ostler does not have that accountant. Your Mac’s compute is a sunk cost. You already paid for the chip. The model runs on your chip. The inference happens on your chip. There is no per-token bill, no monthly reset, no context window you are paying to keep full, no premium tier that unlocks the good model. You can throw everything at it – every meeting, every project, every question you have been saving up – and the cost does not move.

That is not a small ergonomic advantage. It is the difference between an assistant you use freely and one you ration. The people who get the most out of AI in the next five years will be the ones who can afford to have long, meandering, unhurried conversations with it. Ostler makes that affordable for anyone who owns a modern Mac. Because you already paid.

The enterprise unbundling is about margin on someone else’s books. The personal unbundling is about your own credit card, your own custody, and the freedom to stop being your own AI-usage accountant.

What Ostler is, in these terms

Ostler is a personal AI that runs entirely on your Mac. It ingests your iMessages, WhatsApp, Gmail, calendar, contacts, browser history, meeting recordings, reminders, and other on-device data, and builds a private graph of the people, places, moments, decisions, and commitments in your life. That graph is what makes the assistant actually useful to you, rather than useful in the generic sense that any large model already is.

The model itself runs on your machine. The graph sits on your machine. The wiki, the vectors, the assistant, the answers, and the memory all sit on your machine. Nothing about you ever leaves it, because there is nowhere for it to go. That is a design decision that predates Saper’s piece by a year, and it makes Ostler a specific kind of product in a way the enterprise-facing framing does not quite reach.

The value is not the model. The model is a downloadable, upgradeable, interchangeable component. The value is the memory of you. That is what Ostler is here to hold, and to hold on your own machine, so that the memory belongs to the person the memory is about.

The market is voting, twice

On 24 July, thirty-plus of the largest technology companies signed an open letter in support of open-weight AI models. The signatories represent roughly eight trillion dollars of market capitalisation. That is the supply-chain vote: the industry has decided the model layer is going to be open enough to be usable.

Three days later, one of the enterprise-SaaS VC firms with the sharpest track record in the last twenty years (Emergence Capital’s early cheques went into Salesforce, Veeva, Zoom, Box, Doximity) published the demand-side companion piece: the value has moved off the model layer entirely. That is Saper’s vote. The industry and the money have arrived at the same conclusion from opposite ends.

Ostler was built for the personal instance of exactly that conclusion. Local model, private memory, no server between you and the answer, and a design that cannot be undone by a policy change or an acquisition because there is no server to change the policy for.

The enterprise unbundling is already happening. The personal unbundling is next. Ostler is what it looks like.