- calling for slowing the pace of AI growth
- Focus on security means more, NOT less spending
- Frontier Wars: Anthropic vs.
One essay took the AI trade down this week.
We think the fear is misplaced, which opened an attractive entry point.
Hardware is still the largest and, in our view, most attractive part of the opportunity. But the opportunity set is broadening fast. We’ll walk the value chain and show how risk and reward change as you move up the stack.
1. Anthropic Calling for Slowing Pace of Growth
On Saturday, Anthropic’s Dario Amodei published We Must Pace the Frontier, arguing capability is advancing toward the point where it would be hard for humans to control. Both Sam Altman and Elon Musk agreed.
For background AI models have been improving at a rapid pace. The first models were mostly prompt-and-response with limited reasoning ability (GPT -3 & 3.5 and Claude 1&2). This was followed by Strawberry-o1, the first model with “thinking” capabilities before responding.
Now, the latest development, which is both the biggest upside surprise and also what is spooking the markets, is “self-improving,” a.k.a. “recursive,” models. There were several developments over the past week that alluded to us being on the cusp of a new breakthrough:
- OpenAI showed early signs of self-improvement with Astra.
- Google is reportedly making progress towards a self-improving model.
Anthropic sounded the alarm bell that their next self-improving model is so good that it’s dangerous.
2. What Does This Mean for AI Hardware
AI Hardware got hit the harde=st by these comments.
Chips fell. Semicap fell with them. The is now roughly 17% off its highs, after a 67% year-to-date gain.
But more cautious pacing of frontier model development affects positioning between AI Labs and open-source models much more than broader AI spending.
Three things are worth clarifying:
Frontier progress and agentic adoption are different curves. Even if the largest models slow, the agentic ecosystem keeps expanding. You don’t need GPT-7 for the infrastructure buildout to continue.
According to a McKinsey survey on the use of AI, nearly nine in ten respondents reported regular use in at least one business function, but the majority of the organizations are using AI chatbots vs. more advanced Agentic AI.
- Only ~15% of the small enterprises and 25-30% of the large enterprises surveyed use AI Agents, and ~50% of them are either piloting or experimenting.
- Slower frontier development doesn’t mean slower spending. It changes where capital goes: cybersecurity, AI governance and control, agent infrastructure.
- Platforms and cybersecurity companies are likely to benefit disproportionately as focus shifts toward responsible development.
- Caution is also a positioning strategy. The warnings may be sincere. They also raise barriers to entry and reinforce the labs already at the frontier, ahead of potential IPOs.
Ultimately, this debate is more about market share and positioning of closed vs. open models rather than overall spending.
Binding regulation is an unlikely outcome, especially during this administration. Washington has treated AI as strategic in the competition with China and publicly rejected a slowdown. Industry-led self-regulation is the realistic outcome.
3. Frontier Wars
The biggest implication of the Anthropic alarm is dominance of frontier models.
Closed-source models have been winning, especially on wallet share (value) rather than token usage (volume), and this news sheds further light on how capable they are becoming.
Model introductions have been the biggest driver of market-share gains between the two leading AI Labs. As an example, on OpenRouter last week, OpenAI models overtook Anthropic models for the first time.
- Astra was the top model by spend last week.
- Luna led in tokens.
Stalling growth means losing share, which is why we don’t believe this is a likely outcome.
Bottom Line
AI adoption is still in its early innings regardless of the pace of frontier model development. What’s changing is the shape of the opportunity, not its size. In this new era of Agentic AI, growth will come from inference and from agents proliferating across the enterprise.
That broadening changes where the investment opportunities are. Agents need more distributed infrastructure — more networking, more CPUs, more cybersecurity — opening avenues beyond the original beneficiaries of model training.
The market keeps looking for the moment the spending stops. The contracts keep pointing the other way.
