Over Labor Day weekend, I had one of those experiences with AI that left me shocked by the speed of improvement.
I was using Astra to build a trading model. I had read many X posts of people who had used it to build trading systems and was inspired. My pattern recognition knew something was different based on who was posting and the description of what they did. I took an idea I had been thinking about for weeks, gave Astra the logic, and let it start building. It turned the idea into code, tested it, found problems, fixed them, and kept refining it largely on its own. Since 1997, when I moved to Brazil, I have always had someone working with me who could turn an idea into something we could test and eventually put into production. In the past, I would have been deeply involved throughout that process. This time, my contribution was mostly the initial framing of the idea including the change I see in market structure. Astra did the work, and it did it at a speed that was difficult to comprehend. After nearly 30 years of doing this, it is hard to describe just how much time the model saved.
What stayed with me into the next morning was how quickly and seamlessly the whole process moved. I made a subscriber video about how anyone could do it the next day because the experience was different from the incremental improvements we have become accustomed to seeing every few months and I wanted to inspire others to have the same experience. I had just recently been blown away by Grok Bot and now this. Comparing Astra with an older model almost missed what I felt through experience was happening. The rate of improvement is accelerating at yes, a scary pace.
That is what I mean when I say the slope is becoming the story.
During that week, Jensen Huang congratulated OpenAI and said that “AGI has arrived,” pointing to Astra and the progression from ChatGPT to o1 to Astra in roughly four years. Astra was trained using more than 100,000 Nvidia Grace Blackwell systems, and Huang said another 400,000 GPUs were coming online. Whether Astra meets anyone’s preferred definition of AGI is almost beside the point for the investment question. There is no universally accepted definition anyway. What matters is how quickly economically useful intelligence is improving and how much compute is being used to produce it.
Then the story became much more interesting over the weekend. I was in the middle of writing a solving Navier Stokes paper and the impact to the economy when X started shaking over a new piece of AI news.
Dario Amodei published an essay arguing that frontier AI development should be paced more carefully. His concern centers on rapidly improving capabilities, misuse, loss of control and the possibility that AI systems themselves could accelerate future AI development. He proposed independent evaluation, common standards and eventually broader international coordination. Sam Altman expressed support for pacing the frontier, and Elon Musk also backed Amodei’s concern.
A headline saying that the leading AI labs want to slow development sounds bearish. It invites an obvious chain of reasoning: slower models, slower commercialization, lower returns on all that capital spending, eventually less demand for GPUs and data centers.
That interpretation makes sense if the slowdown reflects technological disappointment seen through a lack of demand. It means something very different if the people closest to the technology are becoming cautious because capability is advancing faster than their systems for controlling it.
The Sacks Objection Matters
David Sacks raised the strongest objection to reading the behavior of the labs too literally.
His response was essentially that if OpenAI and Anthropic genuinely believe their unreleased systems are becoming dangerous, they are free to slow down. They do not need a new regulatory structure to do it.
He also questioned their incentives, and that is an important part of the argument.
Frontier labs face product liability if powerful systems cause real damage. Reliability matters commercially. A model that is slightly less capable but much more predictable may be a better product. Incumbent companies can also benefit when regulation raises the cost of competing with them.
So there is an alternative explanation for all of this. A company can talk about danger because the danger is real, because caution is economically rational, because regulation helps its competitive position, or because some combination of those things is true.
That weakens any simplistic reading that the labs are worried, therefore superintelligence must be around the corner. But it does not make the capability signal meaningless.
Sacks himself separates the two issues. He challenges the regulatory solution while acknowledging that the labs may genuinely be seeing something internally that outsiders cannot see. Amodei wants a more formal framework for pacing frontier development; Sacks argues that companies should act voluntarily if they believe the risks justify it.
As an investor, that distinction is useful.
I do not need to decide whether Amodei’s preferred regulatory structure is good policy to notice that the argument has shifted. Some of the people with the best visibility into frontier systems are now debating how quickly capabilities should advance and how much restraint is appropriate.
That is a very different conversation from the one we were having a few years ago.
For investors, one way to think about it is to turn off the sound and watch the behavior.
The leaders of frontier AI companies are discussing whether they should deliberately slow capability development. Washington is arguing over whether they are acting out of safety concerns, commercial self interest, regulatory strategy, or some combination of all three. Behind those disagreements sits a more basic fact: the debate itself is happening because the technology has become consequential enough to force it.
Wall Street May Be Modeling the Wrong Distribution
Most AI forecasts are still built around variables that fit neatly into spreadsheets: GPU shipments, hyperscaler capex, data center construction, token consumption, cloud revenue, depreciation and return on invested capital.
Those are the numbers we should model. I use them too.
The problem is the range of outcomes surrounding them.
How many models seriously consider a world where AI agents become a large source of digital labor, where AI materially increases the productivity of software development, engineering and scientific research, or where AI researchers use increasingly capable AI to help build the next generation of models?
And how many semiconductor forecasts contain anything resembling a scenario in which wafer fabrication equipment spending eventually reaches several times its pre AI baseline?
I am not putting 10 or 20x WFE spending into a base case. I am saying it belongs somewhere in the scenario analysis.
Leopold Aschenbrenner’s paper, Situational Awareness, was valuable for this reason. We are about where he thought we would be with the models reaching this critical safety point. The important insight was to stop treating the current model as a static object and focus on the trajectory.
That is also how I think about the Visser Labs V. One side is linear; the other bends upward. Investing in an exponential technology requires spending more time thinking about the second side.
My experience with Astra was one small data point. The debate now taking place among the frontier labs is another.
Follow the Compute
This leads back to the part of the story that matters most for my thematic portfolio.
Better models will become more efficient. That seems inevitable. But I have never understood why efficiency is automatically treated as bearish for total compute consumption.
As intelligence becomes cheaper and more useful, the number of economically rational uses for it expands. I use my own experience as an example. I can test hundreds of models in a week now. That used to take years. I am using more compute every single week because the models can do more. The video I put out will allow people to replicate what I did and use the code stored in GitHub.
A better coding model produces more software. Better software tools expand into engineering. AI assisted research opens possibilities in biology, materials and drug discovery. Agents can take on business processes that nobody would have considered automating when the technology was expensive and unreliable. Robotics takes intelligence into the physical economy.
Efficiency can therefore increase consumption rather than reduce it. The Jevons paradox is the framework here: making a resource dramatically more productive expands the universe of uses for that resource.
AI may turn out to be an unusually powerful version of that phenomenon.
This is why I keep coming back to a simple loop: compute creates more capable intelligence; more capable intelligence creates new applications and new discoveries; those applications create additional demand for intelligence, which creates additional demand for compute.
If AI begins making a meaningful contribution to AI research itself, the loop tightens further.
A large increase in semiconductor demand would ripple well beyond GPUs. It would reach wafer fabrication equipment, foundries, advanced packaging, HBM, networking and optical infrastructure. Then come the physical bottlenecks: electricity, turbines, nuclear generation, transformers, switchgear, cooling, copper, land and data centers.
We are gradually building the industrial system for producing intelligence that Jensen Huang and Elon Musk have talked about the last couple years. The size of that system ultimately depends on the economic value of the intelligence coming out of it.
The Signal Inside the Slowdown
There will be periods when AI stocks fall sharply for many reasons including fears of a slowdown, hackings and regulatory or bottleneck delays. Positions will get crowded. Investors will degross. Capex estimates will move up and down. Some projects will disappoint.
That market volatility should be separated from what is happening to the underlying capability curve.
The question I keep asking is whether we are spending too much on AI infrastructure or whether the more consequential risk is that we are still underestimating the amount of useful intelligence the economy will eventually consume.
A week ago, I was sitting at my computer building a trading model with Astra and realizing how much faster the process had become. Days later, Jensen Huang was declaring that AGI had arrived while pointing to a model trained on more than 100,000 GPUs. Then the leaders of frontier AI companies began publicly discussing whether capability development itself should be paced more carefully. Now Washington is arguing about the motives, competitive implications and regulatory response.
There are reasons to remain skeptical. The labs have commercial incentives. Nvidia has an obvious interest in emphasizing the importance of compute. Regulation can protect incumbents. AGI remains an imprecise term.
I take those objections seriously.
I also think investors should pay close attention to what the argument itself is telling us.
The next time we hear that an AI lab is slowing development, becoming more cautious or discussing pacing the frontier, the first question should not be whether the AI cycle is ending.
It should be why they are slowing down.
Because in an exponential technology, the pause may not be the signal.
The reason for the pause may be the signal.
