We are not there yet, and recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for.
That is the Anthropic Institute, writing about AI that helps build the next, better AI. And they put a hard number on it: as of May 2026, "more than 80% of the code we merge into Anthropic's codebase was authored by Claude." Two years ago that figure was a rounding error. Whatever you think about where this ends up, the first leg of it is not hypothetical. It is already running inside one of the best AI labs in the world.
Most people file this under safety, and Anthropic encourages that reading: AI improving AI, faster than anyone can keep watch, is the oldest worry in the field. The worry is legitimate. It is also only half of what is going on, and the other half barely comes up.
The half that gets skipped has almost nothing to do with alignment. It is about what becomes scarce once raw intelligence stops being the constraint. The shortage moves to physical things, chips and memory and electricity. Advantage starts collecting wherever those things sit. The timeline slips out of the hands of the people trying to govern it. And the contest between countries turns into a fight over power and manufacturing.
(I write about the physical side of AI full-time at BEP Research. The pieces I draw on are linked as we go.)
What "Recursive Self-Improvement" Actually Means
Strip away the science fiction and the mechanics are mundane. An AI gets good enough to do the grind of building the next AI: writing the code, running the experiments, reading the results, and folding all of it back into a better version of itself. That version does the same work a little faster, and so on. Each pass shortens the next one, which is how something slow becomes something you can't catch.
If you already work with these tools, you have felt a small version of this. The model has stopped being a faster way to type and started behaving like a faster colleague. Anthropic's internal data backs up the feeling. On its hardest, most open-ended coding problems, Claude's success rate "reached 76% in May 2026, up 50 percentage points in six months" — six months earlier it was failing most of them. On one repeatable lab task it went from modestly useful a year ago to far past the human baseline today. The phrase Anthropic uses is that Claude went "from super helpful to superhuman in under a year."
So take the premise seriously and assume the snowball is real. Where does it roll to, then, and what finally gets in its way?
The Bottleneck Just Relocates
Recursive self-improvement does not lift the ceiling on AI progress. It just moves the ceiling somewhere else, and most people keep watching the wrong room.
For decades, the scarce ingredient in better AI was human brainpower: the researchers who could have the idea, the engineers who could build it. That is exactly the ingredient this loop automates. But a loop that runs in software still has to actually run, and running it costs real-world resources. Every cycle burns computing power, fills up specialized memory chips, and draws electricity from a grid that takes years to expand. Speed the loop up, and it demands more of all three, faster.
I made a version of this argument in The Compute Confession, about the strange fact that Anthropic built the most capable model in the world and then couldn't afford to run it at the price it wanted: "the constraint is not intelligence, or talent, or capital. It is atoms." Recursive self-improvement is that same problem one level up. It takes the one limit we actually knew how to loosen, human effort, and hands the rest to physical infrastructure, which doesn't care how clever you are.
Then Anthropic goes and makes the point for me. Buried in the section where they list the ways the trend might stall, they name the exact limit I have spent the past year writing about: "the binding constraint to AI progress could be in the supply chain, not the model... The pace of chip fabrication, grid expansion, or interconnect bandwidth may be the constraint, rather than intelligence itself." That is their sentence, not my reading of it.
Where we part ways is what to do with it. Anthropic treats the supply chain as a risk, a thing that might slow them down. I read it the other way around, because all that demand for chips and grid capacity has to land somewhere, and everything it lands on is somebody's revenue. They have already watched the mechanism play out inside their own walls. Once Claude started flooding the company with code, the jam did not disappear; it relocated to the humans who had to review all of it. Anthropic even reaches for the textbook name for this, Amdahl's law: speed up one stage of a process and the stage you left alone becomes the whole problem. Push the researchers faster and the holdup just slides downstream, toward whatever software cannot accelerate, and sooner or later that is something you have to build, power, or wait for.
The loop closes, but it closes onto atoms.
Advantage Piles Up at the Top
If the scarce ingredient is physical infrastructure, then the advantage flows to whoever owns the most of it. That is a very different world than the one AI optimists usually picture.
A race decided by ideas can be run from anywhere a clever person sits. A race decided by chips and electricity can only be run where there is a mountain of computing power, a reliable supply of specialized hardware, and enough electricity to light a small city. There are not many places like that, and not many companies that can pay for them. So recursive self-improvement, taken seriously, concentrates power. It pushes the frontier toward a handful of big labs, governments, and cloud giants, not out toward everyone else.
Builders should sit with this one, because it runs backwards from the last few years. From 2023 to 2025 everything was spreading out: open models, cheap customization, capability leaking everywhere. A loop gated by hardware pulls the other way, back toward whoever already owns the hardware, unless the next idea gets there first.
Does It Get Cheaper Faster Than It Gets Hungrier?
Here's where I could be wrong, and I'm not going to pretend otherwise.
The loop pulls in two directions at the same time. It drives demand for computing power up, but it can also drive the cost of computing down, because smarter models keep finding ways to run leaner, shrink themselves, and waste less. Whether the physical limits keep biting is really a race between those two curves. If each turn of the loop gets cheaper faster than it gets hungrier, the constraint loosens, capability spreads back out, and the whole "advantage piles up at the top" story starts to come apart.
Anthropic concedes this in the same sentence where it names the constraint. In its fullest scenario, progress ends up set by the supply of compute "(or the speed of discovering various efficiencies in algorithmic training or inference)." That parenthetical is the entire counter-argument, written by the people making the bullish case: either you run out of capacity, or you get clever enough to stop needing so much of it. The clever side has been moving fast, too. In one recent industry test, engineers pulled 2.7 times more work out of the very same chips with no new hardware at all, purely by improving the software (I went through it in 2.7x on the Same Iron). A self-improving loop that keeps turning that crank on itself is exactly the thing that could outrun the grid.
My own bet is that the physical limits keep biting for the next several years. Demand has been outrunning efficiency at the frontier for a decade, and the sheer amount of money going into buildout says the people closest to it are betting the same way. But anyone who tells you this is settled is selling you a position, not an honest read.
The Rules Can’t Keep Up With the Concrete
Anthropic frames this as a governance problem. Maybe so. Except the physical world may end up setting the clock more firmly than any rulebook can.
A regulator can move in a few months. A new power line, a chip fab, a backup generator with a years-long order book, those move on their own schedule and they do not care about anyone's deadline. If the loop really is gated by electricity and hardware, then whatever actually controls the timeline is not a committee or a piece of legislation. It is a substation. For a debate that has mostly been about ethics boards and policy frameworks, that should be a little unsettling, because it means the pace is being set by procurement departments and construction crews working through backlogs measured in years.
And Then It Turns Into an Energy Race
The last consequence is about geopolitics, and it reframes the contest between the US and China. If intelligence is no longer the scarce ingredient, then "who has the best model" stops being the right question. The right question becomes who can build and power the most intelligence, and that is a question about factories, energy, and the electrical grid, not about software.
I made the hardware version of this case in The Chip Is Dead: against a rival with more raw electricity, the only answer is "more intelligence produced per unit of power." Recursive self-improvement is what that race looks like up close. Whoever can run the loop more cheaply per unit of energy wins, and that is decided by power plants and transformers, not by clever code.
Where This Could Fall Apart
The strongest objection is that efficiency just wins. If the loop makes AI cheaper faster than it makes it hungrier, the physical limits stop mattering and the whole "advantage at the top" story comes apart. The odds are against that over the next several years, in my read, but they are not zero.
It's also possible the loop stalls before it really gets going. Anthropic flags this itself: today's steep trends "may actually turn out to be S-curves," the kind that climb hard and then flatten. The judgment that separates a competent researcher from a great one might be the one thing scaling can't buy. Eighty percent of the code being AI-written is a productivity jump, not a machine doing original research, and Anthropic is candid that "large performance gaps persist" in the harder, taste-and-judgment parts of the work. The jump from "writes the code" to "designs its own successor" has never happened, and it may be far harder than the trend lines make it look.
And the whole thing assumes the world keeps wanting this much AI. As I wrote in The Token Dollar, "oil is scarce because of war; tokens are scarce because of physics." If demand turns out smaller than the buildout assumes, AI gets cheap and a lot of expensive hardware sits idle, which is the bubble case, and not a crazy one. It's just hard to square with the revenue run-rates and the power contracts already signed.
None of that breaks the core claim, I think. It changes the timing, and how bumpy the ride gets on the way there.
The race that decides everything: does AI get cheaper faster than it gets hungrier?
What It Means, if You Build or Invest
A few things follow from this, depending on where you sit.
If you build: bet on the part that gets cheaper, not the part that gets scarcer. If power is concentrating at the frontier, the lasting opportunity is downstream, in the apps, tools, and workflows that get better and cheaper as AI gets cheaper to run. Owning a frontier model is a game most builders can't win. Building something that gets more valuable as AI costs fall is a game they can.
If you invest: follow where the loop spends its money. Every cycle of self-improvement cashes out as demand for the same handful of physical things: the chips, the specialized memory, the high-speed connections between them, and the electricity to run it all. An essay that reads like an AI-safety paper is, underneath, a demand signal for all four.
For everyone: watch the cost of running AI over time, not the benchmark headlines. That one number tells you which world you're in. If it keeps falling faster than demand climbs, AI keeps spreading out and the optimists win. If it doesn't, the physical limits bite and the frontier closes in around whoever owns the hardware. Either way, that ratio is the thing to track, not the next flashy demo.
Where To Go Deeper
If this was useful, the fuller version of the argument lives at BEP Research. A few starting points:
The Compute Confession: how the price of using top AI models quietly reveals what they actually cost to run
The Token Dollar: why AI computing is becoming a major force in the global economy
The Chip Is Dead: why the energy race, not the chip race, decides who leads in AI
Free posts are public. Follow on X at @benitoz for real-time notes from the field.
Disclosure: The author holds positions in NVDA, LITE, CRDO, ALAB, LSCC, TSEM, BE, and ORCL. This is research, not investment advice.
