Thinking Machines Bet You Want to Own Your AI, Not Rent It

Here is a thing almost nobody says out loud about the current AI market. When your company calls a frontier model through an API, you are renting the exact same intelligence your competitor is renting, from the same handful of labs, priced the same way, updated on the vendor’s schedule, and shaped by the vendor’s choices. You do not own it. You cannot open it. You get a very good stranger who works for everyone in your industry at once.

This week a new lab decided that arrangement is the opportunity, not the default. On around July 15 and 16, 2026, Mira Murati’s Thinking Machines Lab released its first model, Inkling, and the pitch underneath it is blunt: the companies that win the next phase of AI will be the ones that own and shape their own model instead of renting a general one from a lab that also serves their rivals.

Own it, don’t rent it. That is the whole bet, and Inkling is the opening move.

The shift, named

The frontier-lab model, the one OpenAI and Anthropic run, is a rental economy. You send text to a closed endpoint, you get an answer, you pay per token, and the weights that produced the answer live on someone else’s servers where you will never see them. It is a genuinely good deal for getting started. It is a strange deal to build a company on.

Thinking Machines is arguing that the rental economy has a ceiling, and that the ceiling is sameness. As Axios put it, the company thinks enterprises want AI they can customize and own rather than rent from a small set of frontier labs. When every business in a sector calls the same closed model, the model cannot become anyone’s advantage. It is a utility, like electricity. Useful, necessary, and identical for everyone plugged into the grid.

“

A model everyone rents cannot be anyone’s edge. The differentiated model is the one your competitor cannot download, because you shaped it on data they do not have.

”

The counter-move is an open-weight AI model: a foundation model whose actual weights you can download, inspect, fine-tune on your own proprietary data, and run on infrastructure you control. That last part is the quiet radical bit. You are not asking permission to specialize the model. You are not waiting for a vendor’s fine-tuning feature. You have the weights, and what you do with them is your business.

What Inkling actually is

Inkling is a large mixture-of-experts model, and the numbers are worth sitting with before we talk about why they matter.

Per Thinking Machines, Inkling carries 975 billion total parameters but activates only 41 billion of them on any given query. That is what mixture-of-experts buys you. Instead of running the whole network for every token, a router picks a small subset of “expert” sub-networks for each input, so the model has the knowledge capacity of something enormous while the per-query compute stays closer to a much smaller model. You store a giant. You run a specialist. For a company that plans to serve this model on its own hardware, that ratio is not a spec-sheet flex, it is the difference between an affordable deployment and a cluster you cannot justify.

The context window runs up to 1 million tokens, enough to hold a large codebase, a quarter of legal filings, or a long multi-document research task in a single pass without hacking around the limit. And it was pretrained on 45 trillion tokens spanning text, images, audio, and video, which makes it a multimodal base rather than a text-only one.

I do not have published benchmark scores for Inkling, so I am not going to invent a leaderboard position for it. What matters for the argument is not whether it edges out the latest closed model on some eval this month. It is what kind of thing it is, and who it is built for.

This is a base to build on, not a chatbot to talk to

The most important design decision in Inkling is who it is aimed at. It is not trying to be your ChatGPT or your Claude, the friendly assistant you open in a browser tab. As the reporting frames it, Inkling aligns with the lineage of Meta’s LLaMA, a base model that enterprises fine-tune into their own systems, rather than the consumer-chatbot lineage of OpenAI and Anthropic.

That distinction changes everything about how you would use it. A consumer chatbot is a finished product you consume. A foundation base is raw material you shape. Thinking Machines is explicitly selling the raw material, and it built the tooling for shaping it first.

That first product, Tinker, is the fine-tuning API the company launched in October 2025. It is the mechanism that makes “own your model” a real workflow rather than a slogan. You take Inkling’s open weights, you point Tinker at your proprietary data, the support tickets, the internal docs, the transaction histories, the domain nobody else has, and you get back a model that is specifically good at your problem. Then you deploy it where you choose. The ownership is not symbolic. You hold the weights, you did the training, and the resulting model encodes knowledge that never left your control and that a competitor calling a shared API simply cannot reproduce.

Why now

Timing is half of any bet, so it is worth being precise about the inflection this rides.

Open-weight demand is surging, and until recently the most capable open-weight models were coming disproportionately out of China. That created an odd situation for a lot of Western enterprises: the closed frontier models were American but un-ownable, and the ownable open models were, increasingly, not American. Inkling is being positioned squarely into that gap. Computerworld framed it plainly: a Thinking Machines lab offering enterprises a US alternative in open-weight AI. For a company whose procurement or compliance team cares where its model’s weights originate, a frontier-grade, US-developed, downloadable base is a new option that did not exist at this quality a year ago.

The other half of “why now” is compute, because a lab cannot pretrain on 45 trillion tokens on good intentions. In March 2026, Thinking Machines and Nvidia announced a partnership: an undisclosed equity investment from Nvidia plus a multi-year agreement to deploy 1 gigawatt of Vera Rubin compute. A gigawatt is data-center-scale power, the kind of commitment that signals this lab intends to keep training frontier-scale models, not ship one and stop. The rental incumbents are not the only ones with a hyperscaler behind them anymore.

The argument, taken seriously

Strip the framing and here is the actual claim: a centralized lab selling one identical product to everyone will, over time, underperform a model that an organization can shape to itself. It is worth pressure-testing that, because it is not automatically true.

Where the argument is strong. For any problem where your advantage is your data, ownership wins on structure, not just on price. A bank’s fraud patterns, a hospital network’s clinical notes, a manufacturer’s defect images, none of these live in a frontier model’s training set, and none of them should be shipped to a third-party API if you can avoid it. Fine-tuning an open base on that data produces a model that is better at your task and keeps the data inside your walls. You also get control over the two things renters cannot touch: hosting (run it in your own region, your own cloud, air-gapped if you must) and cost curve (a model you host has a fixed-ish compute bill instead of a per-token meter that scales with your success). Murati’s lab is betting that as AI moves from experiments to load-bearing production systems, those become the properties companies actually care about.

Where it is harder. Owning a model is not free the way calling an API feels free. You need people who can fine-tune, evaluate, and serve a 975-billion-parameter mixture-of-experts model without setting the cluster on fire. You inherit the operational burden that the rental model exists to hide. And the frontier closed models will, for now, keep pushing raw capability at a pace a fine-tuned open base has to chase. So the honest version of the thesis is not “ownership beats renting for everyone.” It is “for a growing set of serious companies, the calculus that used to obviously favor renting no longer does, and Inkling is what tips it.”

What this changes for you

Here is the reframe, and it is bigger than one model launch. For most of the last few years, the build-versus-rent decision for AI had an obvious default. Renting won, because the only models worth using were closed and rentable, and rolling your own frontier model was a fantasy for all but a few labs. So every serious company quietly accepted the rental economy and its ceiling.

When a frontier-grade base is downloadable, that default breaks.

“

The question is no longer “which API do we call.” It is “which parts of our intelligence do we rent, and which do we own because they are the actual business.”

”

You do not have to go all-in to feel the shift. The realistic near-term shape is a split: rent the general frontier model for the broad, commodity reasoning where sameness is fine, and own a fine-tuned open base for the narrow, high-value, data-rich problem that is your real moat. A month ago the “own” column was mostly theoretical for a normal enterprise. With Inkling plus Tinker, and the Nvidia compute behind the lab that made them, it is a line item you can actually plan.

The labs selling you access have spent years training you to think of a model as a service you consume. Thinking Machines just made a serious case that, for the part of your business that matters most, a model is an asset you own. Whether or not Inkling specifically is the one you build on, that question is now on your table, and it will not leave it.

Sources

• Thinking Machines Lab, “Introducing Inkling”
• Axios, “Mira Murati’s Thinking Machines releases open-weight model Inkling”
• Fortune, “What is Mira Murati’s Thinking Machines first AI model, Inkling?”
• Computerworld, “Thinking Machines Lab offers enterprises a US alternative in open-weight AI”
• Simon Willison, “Inkling”