When AI Builds Better AI, the Moat Moves to What the Model Can't Own
An independent group called METR published an assessment this week that most people will read as a curiosity and a few will read as a threat. It looked at how fast Anthropic's models are actually improving and found that the rate has roughly doubled compared with the prior stretch. The details matter less than the reason. METR estimates that most of that acceleration, on the order of 80 percent of it, comes from AI being used inside Anthropic's own research. The models are helping build the next models. That sounds like a technical footnote about one lab, but it is a business event, because it changes what anyone should expect from the frontier and therefore where the durable value actually sits.
I want to be careful not to overstate one nonprofit's estimates. METR is measuring its own definition of capability on well-specified tasks, not revenue or product usefulness, and those two things have drifted apart over the past year. A model can get dramatically better at a benchmark while the way it shows up in a customer's workflow changes less. Still, when a lab's internal research loop starts feeding on itself, the direction of travel is unmistakable. The compounding is real even if the exact number is approximate. You do not need to trust the 80 percent to notice that the people building the next generation increasingly use the current generation as part of the tool chain.
Here is the part that founders should actually weigh. If the frontier is accelerating, the obvious reaction is to worry that the big labs will pull away forever and that everyone else will be left behind. I think that instinct gets the market wrong, and the history of every previous layer of technology suggests why.
Whenever the thing that builds the next thing becomes cheap and fast, the leverage in the market does not go to the thing. It goes to whoever sits between the thing and the customer, the distribution, the workflow, the relationship, the data, the switching cost. Software itself went through this. For two decades, the cost of writing and shipping an application kept falling, and the winners were not the people who wrote code fastest. They were the people who owned the distribution, the customer relationship, and the workflow that companies did not want to leave. The same logic applies to models now that models have crossed the threshold where they can meaningfully help build themselves.
The clearest consequence is that the frontier premium shrinks faster than people assume. If the best model improves twice as quickly as it did a year ago, then whatever lead yesterday's best model had is eaten away in half the time. A capability that looked like a moat in January is a commodity by June. For a founder building on these models, that is both a warning and a release. It is a warning that betting everything on being the first to the newest model is a treadmill, because the newest model is never going to stay the newest for long. It is a release because it means the thing you need to own was never the model. It was the layer above it.
There is an honest counterargument, and I should not wave it away. The acceleration is not shared equally. The labs that run the research loop own the compute and the scale to feed it. A small team cannot reproduce 80 percent catalytic research improvement on its own hardware, and pretending otherwise would be naive. So the near-term picture does show the leaders compounding genuinely hard to catch on capability. What I push back on is the conclusion that follows from that, which is that mere capability is where the spoils end up. Being hardest to catch at raw benchmark improvement is not the same as being hardest to catch at owning the customer.
The two get confused because, for a moment, they looked identical. When a frontier model was the only thing that could do a task, the lab that made it also owned the only way to access it, and capability and distribution were fused. That fusion is dissolving. The moment a task is good enough on an open or cheaper model, the moat reverts to whoever controls the interface, the data that accumulates around it, the trust, and the workflow. We have seen exactly this in the enterprise. Companies do not switch off a workflow that is integrated into how they run because a newer model is fractionally better. The switching cost lives in the surrounding system, not in the weights.
This is the point of the METR finding that I keep circling back to, and it deserves emphasis. When AI builds better AI, the frontier improves, but it also commoditizes the thing above it. Every release makes the previous release's advantage more ordinary. That is good news for anyone whose business is the layer that survives model churn: the interface, the process, the relationship, the accumulated data. It is uncomfortable news for anyone whose only claim to defensibility was being slightly ahead on the latest benchmark, because that lead is now a wasting asset with a shorter half-life than it used to have.
I also suspect this deepens, rather than solves, the dependency problem I keep writing about. A frontier that compounds faster gives the labs more leverage over the small companies that rent their intelligence by the token, because those customers have fewer good alternatives at the very top and because the rate of change makes it easy to fall behind. The answer is not to reflexively distrust every lab. It is to notice where the power actually sits and to build in the layer where it does not. Whether you run open weights locally or subscribe to a closed API, the moat you can actually own is the one that does not get rebuilt by someone else's research cycle: your workflow, your customer relationship, your data, your trust.
The irony worth sitting with is that the more capable AI becomes at building itself, the less the raw capability of any single model is what you should be maximizing. The ratio that matters has shifted. If the frontier is improving twice as fast, then being locked into the wrong provider, the wrong interface, or the wrong accumulated data is twice as penalizing. Flexibility and ownership of the surrounding layer become the durable assets, not because they are exotic, but because they are the part that does not get rebuilt underneath you.
Let me end with a balanced note, because there is no clean either/or here. The acceleration is real, and the labs deserve credit for it. The people pouring compute into the frontier are not wrong; capability will keep moving and that will open plenty of genuinely new workflows. But the lesson of the METR finding is not that the labs are unstoppable. It is that the things you can own, the workflow, the relationship, the data, the distribution, are more durable than the thing that keeps moving. When the machine that builds the future is itself getting faster, you want your position on the side of the table that the machine cannot leapfrog: the side that owns the customer. The models will keep getting better. That is exactly why what you build around them matters more, not less.