The AI Specialist Brain Just Got Folded Into the General Model
For a while the story of AI has been that every serious problem needs a specialist brain. Robotics needed a model trained for robots, customer support needed a model tuned for support, coding needed a model built for code, and that belief underwrote a whole tower of startups and expert claims. Each domain, the reasoning went, was distinctive enough that a general model would always sit a rung below the one built for the job. Then, in a relatively quiet way, OpenAI did something that cuts against the whole arrangement. It disbanded its standalone robotics team and moved the researchers onto other teams, because its general purpose frontier models, including GPT-6 Astra, had started running robots directly. There was no longer a robotics team because the general model had turned out to be a robotics model too, which is a strange and slightly unnerving idea for anyone whose bet was that specialization would win.
Let me be careful about why this is more consequential than a single personnel move, because it would be easy to skim past it as internal housekeeping. OpenAI had circled robotics for years. Researchers began the year expecting progress to build on video models like Sora, and there were long intervals when the company seemed unsure the physical world was its job at all. The decision to fold the team in is not a retreat from robots. Altman has committed to a humanoid demo in 2027, which would put OpenAI directly against the very companies it once backed, Figure and 1X. It is instead a structural argument. The company concluded that the capacity to run a robot no longer lives in a separate model built for robots. It lives in the general model, and so a dedicated unit, far from accelerating the work, mostly slowed it down. That is the same logic an outside observer might stare at for a moment, because it quietly invalidates a founding assumption of an entire industry.
There is a familiar shape to this if you look back at software, and it is worth setting out because the market tends to forget it every cycle. For decades, vertical software was among the safest businesses there was. A company that understood one industry, dental practices, trucking, legal billing, could build something narrow enough to own that industry and hold it for years, because the general platform rarely bothered to chase the long tail of specialties and chased it slowly when it did. The moat was domain knowledge packaged as software. What happened over time, with remarkable consistency, is that the platforms absorbed the most valuable versions of those specialties into the core product, and the standalone vendor had to make its living on the edge, on distribution, on a head start, on a workflow the platform was slow to understand. The AI era is compressing that same cycle into a few years, except the absorption is happening at the model layer, where the thing that used to be a separate product is becoming a capability the general product already carries.
Robotics is the clearest case because it was the most visibly separate. For years the assumption was that controlling a physical body, planning a manipulation, understanding a world of objects and obstacles, demanded a model trained for exactly that. The robot startup raised money on that claim, and so did the foundation model companies selling access to the same idea. If a general model can be applied to a robot directly, then the performance spread between the specialist brain and the general one narrows fast, and the question for every robotics company becomes the question every other AI company is already facing. If the brain is increasingly available to anyone, where is the value that you alone hold? The moment a capability stops being scarce, the business built on selling that capability has to find something else, or it becomes a feature with a price tag attached to it.
And this is really not about robots. Robotics is just where the separation ran deepest, so it is where the collapse is easiest to see. The same dynamic is playing out in customer support, in code review, in legal research, in marketing, in any corner of the market where a startup has staked its identity on a model that knows its domain better than the general one does. The uncomfortable question running through all of it is whether that specialty is a durable moat or a feature the general model will grow into. If the general model absorbs it, the specialist's advantage does not necessarily disappear, but it moves. It moves to the workflow the product is embedded in, to the data a company holds that a model cannot simply obtain, to the trust and relationship with a customer who has to decide whether to hand the machine something real. The model stops being the moat. The surroundings become the moat.
Of course, this does not mean every specialist is doomed, and I want to hold the objection up honestly, because it is a good one. There is a real argument that specialty still matters for performance, that a model trained hard on one domain with real data and fine tuning will beat a general model on that domain's hardest tasks for a long time yet. There are also places the general model will not bother to go, niches too small, too messy, too far from a big platform's roadmap to justify the investment. Some verticals will survive precisely because they are uninteresting to the general players, and the history of vertical software is not a story of total extinction. It is a story of consolidation around the slices the platform chose to absorb, with the survivors holding the edges the platform left alone. So the question is not whether specialization dies. It is whether the specialization you hold is the core a platform will eventually want, or the edge it is content to ignore.
The investment implication is the sharpest part, because capital reads signals like this fast and reprices consequences into valuations before the rest of us catch up. OpenAI's move tells the market that even the preeminent general purpose lab no longer believes its most capable work requires a separate brain per domain. That is a statement about where the frontier is heading. If a general model can run a robot directly, then the robotics startups whose pitch rested on owning the brain have to revalue what they actually possess: the hardware, the deployment, the fleet, the operations, the relationship with a customer, rather than the model. Some of that repricing is already visible in how money has moved toward companies that bundle brains, bodies, and deployment together rather than pure model shops. Over and over, the durable asset turns out to be the part you own that a model cannot, the physical presence, the trusted workflow, the installed base.
There is a second reading that is less comfortable, and it connects directly to the dependency worry running through this whole buildout. The same movement that folds the specialty into the general model also concentrates more of each vertical's destiny in the laboratories that own the general model. When a single general model can run a robot, a support queue, and a coding team, then the companies that use those capabilities sit one more step away from owning the thing that does the work. The robotics company that borrowed OpenAI's brain is now downstream of every decision OpenAI makes about that brain, its pricing, its roadmap, its freedom to change the terms. Folding the specialist team in is good news for the breadth of the general model and quietly bad news for anyone who depended on a separate supplier for that specific piece. Consolidation and dependency are the same trade looked at from opposite ends, and the lonelier end belongs to whoever is doing the borrowing.
So I keep arriving back at the same place I land every time I watch this industry move. Capability keeps concentrating in the general model, and the value keeps drifting out to the parts of a business a model does not own. The specialist brain is being absorbed into the general one, and the honest response is for every AI company to ask what it would actually hold if its model suddenly became a feature that anyone could switch. For some companies, the answer is encouraging: a workflow no platform wants to bother with, data nobody else has collected, a customer relationship nobody can replicate. For others, the answer is mostly the model itself, and that has always been the weakest position to occupy, you just needed a well financed robotics team to make it obvious. The general model is getting better at everything, and the moat was never really the special brain. It was what you could do with that brain that nobody else could, which is a different and much harder thing to own.
The largest foundation model company just told its robotics team there was nothing special left to build, only consolidate into the general thing. Every vertical built on the same assumption should wonder whether that conversation is headed its way next, because the folding in never stops at the first specialty, and the only durable answer was never the model at all.