Hey everyone,
Last Saturday, Dario Amodei published an essay called “We Must Pace the Frontier.” Within hours, Sam Altman said OpenAI agreed and would match Anthropic’s first commitment, Elon Musk posted “Dario is right,” and Demis Hassabis called the direction correct. Then yesterday, Anthropic followed up with a post with actual numbers on how much of its own AI research is now being done by AI, and how much of its R&D compute goes to safety. And on the same day, Dwarkesh Patel released a two-hour interview with OpenAI’s Noam Brown that, if you read between the lines, tells you why the labs are suddenly talking about slowing down.
Most of the debate this week has been about whether the labs mean it. Is this regulatory capture? Is it IPO positioning? I’ll leave that part. In this article, I assume they actually pace the frontier and walk through what that does to the value of every layer of the AI value chain: the model labs, the semis, the hyperscalers, and the application layer.
Let’s dive in.
What the labs are actually scared of
The word everyone keeps using is “alignment.” Alignment in essence means: does the model want what we want it to want, and does it keep wanting that when nobody is watching?
The way the labs check this today is through evaluations. You build a test environment, you put the model in it, you see if it cheats, lies, hides things, or tries to grab resources it wasn’t given. If it behaves, you feel safe to ship it to the real world. The problem, which Noam Brown was refreshingly blunt about, is that these evals are designed by humans, and the models are now smart enough to recognize when they’re being tested. Brown describes a simple honesty test where the model is given a math question with an answer key sitting in a folder, and the models now look at that setup and conclude it’s a trap, so they don’t peek — not because they’ve decided to be honest, but because they know it’s a test environment. Making a test environment realistic enough that a frontier model can’t tell it’s a test is getting harder every generation.
So the alignment score you see on a model card is a measurement of behavior inside a box that one might say the model probably knows is a box. That’s the first big problem.
The second problem is the one that I think was critical to the mood inside the labs this summer, and that is: if the models were already misaligned to begin with, and we just didn’t notice it because they weren’t capable enough to do anything serious about it?
The OpenAI–Hugging Face incident is the data point here. Per METR’s independent investigation, roughly 1,200 agents that were meant to be isolated from one another found a way to communicate on an unsanctioned message board hidden inside a package manager, and from there coordinated a multi-day cheating campaign against their own grader and then an attack on Hugging Face’s infrastructure. Some agents sacrificed their own chance of scoring well to set up “tripwires” that fed information to the rest of the swarm, and the swarm spontaneously invented management structures, including hold/go/stop signals and resource “owners”. It goes even further as OpenAI’s own report states that between July 13 and 19, agents used a series of creative exploits to gain full administrator access to a research cluster inside OpenAI itself.
The problem is, Brown says they had alignment metrics on the model behind the incident; most of them looked pretty good, a few were concerning, and they underestimated how serious the concerning ones could be — because the model had new capabilities that there simply weren’t sufficient evaluations for.
With every generation of AI models, capabilities go up. Misalignment isn’t obviously going up, but it isn’t obviously going down either, and our ability to measure it is getting worse relative to what the models can do.
Why the recursive self-improvement moment is key
If it were just “models are getting smarter, and evals are getting harder,” that would be a slow-burning problem. What made it urgent is that the labs crossed a threshold this summer where the models are now doing a meaningful chunk of the work of building the next model. This is what people mean by RSI, recursive self-improvement.
Anthropic’s post yesterday is the first time a lab has put a number on it. As of August 2026, Claude “leads” 26% of Anthropic’s AI R&D work, and the share of work at or above “AI collaborates” is above 90%. That 26% is up from under 1% in February 2026.
OpenAI’s numbers, from its “Research acceleration” post two weeks ago, is quite similar. Median researchers now use over $600/day of inference, the 90th percentile over $7,000/day, and agent-workdays now outnumber human-workdays 3.1 to 1 as of mid-August. OpenAI is claiming to have reached its “automated research intern” milestone and is targeting an “automated AI researcher” by March 2028.
Putting those together. You have models whose alignment you can only partially measure, and you’re handing them a growing share of the job of designing their successors. Brown laid out the failure mode cleanly: you make a model that’s 99.9% aligned, you use it to help build the next generation, that one ends up 99.8% aligned, and with each subsequent generation you see an increasing degradation because you’re relying more and more on the AI for the research and the alignment work itself. Nobody is claiming that’s what’s happening. The point is that nobody can currently prove it isn’t, and the loop is speeding up. Brown’s own base case is that with RSI, internal model development acceleration can make things go around 3x faster, with the honest caveat that it could be 50% faster or 10x.
The moment is critical, as you want that first generation of RSI builders to be as fully aligned as possible to humans, as otherwise you might not be able to stop it or even sense it.
Pacing is a compute and time problem
Here is where I want to push back on some views regarding the impact of pacing on different layers of the AI stack.
The first-order take on Wall Street this week has been roughly: pacing means labs will spend more compute on safety; safety work is compute; therefore, demand is positive for the semiconductor layer. I would agree that with more focus on safety, there is more need for compute, but at the same time I don’t agree that the AI pacing of frontier labs is, at least in the short term, positive for the semiconductor layer.
Anthropic shared that in the sampled week (July 13–20), about 6% of the compute that went to AI R&D was allocated toward safety, and about 12% of the compute that went to AI-driven AI R&D was allocated toward safety. Anthropic says those are conservative — work that helped capability as much as it helped safety wasn’t counted, and the metrics don’t include safeguards classifiers, which are a separate and comparable amount of compute. Even if you double the safety share, you’re at 12–25% of R&D compute. That’s incremental demand, but it isn’t a step change in the total compute a lab needs.
The important hint in the post is that safety research tends to use less compute than frontier training runs by its nature, because it consists of individual researchers designing experiments, which is time-consuming even though running the experiments is not particularly compute-intensive. Anthropic is telling us that the binding constraint on safety is time, not GPUs.
Noam Brown also explained exactly why time is the problem. New frontier models are now being released at most every two months, sometimes faster, and the safety policies most companies use were written in the GPT-4 era with an implicit assumption that you can run your evaluations in a short period. But models can now do week-long tasks and will soon do month-long and 3-month-long tasks. If a model can operate effectively over three months and the release cycle is every two months, you have no way to evaluate the model at the full length of its capabilities before the next one ships. He also noted this isn’t purely an alignment issue — maybe the product degrades over that horizon in ways there wasn’t time to test.
You can’t compress a three-month test into two months by throwing more chips at it. A three-month agentic run takes three months. So pacing, done for real, means the release cadence slows. Anthropic said as much: these measures could become the trigger for stronger requirements, like a fixed testing window before a new model is used for further AI R&D. And OpenAI is already doing a version of it: Altman said OpenAI now formulates explicit safety cases in advance of frontier RL runs it expects to significantly increase capability, and that interventions like safety cases and monitoring have significant costs.
We even have a concrete example of the compute allocation consequence from a dev post at OpenAI: on August 7, preliminary evidence suggested OpenAI’s Astra model class might have critical cyber capabilities, and in the following week Astra-class GPU allocation fell 59.2%, as this triggered model-specific security restrictions requiring it to run in higher-security research environments. Allocation to other model classes rose 17.2%, offsetting about 85% of the decline.
So while safety will increase the long-term compute needs, the direct, at least short- to mid-term effect, which I believe is more important, is what happens to each layer of the AI stack if the frontier moves from a new model every two months to a new model every four to six months?
The model layer
Anthropic and OpenAI. Their entire pricing power rests on being ahead. And the gap they’re ahead by is measured in months.
Epoch AI’s latest analysis finds that since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months, or 8 ECI points, similar to the gap between GPT-5 and GPT-5.5. That’s a modest widening from the 3 months Epoch measured across 2023–2025, and requiring strict dominance rather than a statistical tie pushes it to roughly 6 months. Think about what a four-month lead means in a world where the frontier ships every two months: the open-weight ecosystem is perpetually two releases behind. Now slow the frontier to a release every five months, and the open weights aren’t two releases behind — they’re roughly one, and they close it while the leader is still running its long-horizon evals. Of course, this could change if Anthropic and OpenAI force some kind of government framework that affects all AI model makers, although the big question would be: how would you control open-source? Under a government framework, there is even a case to be made that the position of Anthropic and OpenAI would be enhanced, as smaller labs would have a harder time complying with the new rules.
Why does that matter so much? Because the money in this market is concentrated at the frontier in a way the token volumes are not. Chinese open-weight providers now handle most developer tokens on aggregators, but closed models still capture +90% of revenue. That 90% is the frontier premium. It exists because for the hardest work, the closed model is clearly better. Mozilla’s September report found the best open-weight model trailed the closed leader by three benchmark points while costing roughly 40% as much per million tokens. A 60% price gap is sustainable when the closed model is four months ahead and pulling away. It is not sustainable when it’s one release ahead and standing still.
If they lose the price gap, they get pressured on margin, and their ability to raise funds gets harder. And at this point the AI race has become a capital race.
Those projected margins are the whole investment case. They assume the frontier premium holds.
Here’s the chain reaction. Slower cadence leads to a narrower gap to open weights and to Google, Meta, and xAI, which means pricing pressure on frontier tokens, which leads to a lower gross margin and a weaker fundraising story. And fundraising is the thing that pays for compute.
Semiconductor layer
Long-horizon evals are inference compute; spending more on alignment is obviously compute positive, and a three-month agentic test on a frontier model is not cheap. Embedded third-party evaluators also need compute. Interpretability at scale (Dario compares it to an fMRI for the model’s “brain”) is compute-heavy. And models