Chief Market Strategist for the Americas
Global Market Strategist
Over this past weekend, the rhetoric around the dangers of artificial intelligence (AI) models reached a crescendo. Anthropic’s chief executive, Dario Amodei, published an essay entitled “We Must Pace the Frontier” calling on the industry to slow the rate of model capability improvement to focus on safety. Within a day, the chief executives of OpenAI, Google DeepMind and xAI had spoken in agreement—a rare united front after years of intense competition over model capabilities. This follows a summer-long increase in public unease over AI and an uptick in model training incidents and warnings1. For investors, there are a few useful questions worth exploring amidst the ever-more-frequent volatility around the AI buildout: 1) what’s behind these sudden calls for a slowdown, 2) what’s the impact on projections of AI-related capex spending, 3) how can different areas of the market react and 4) how should portfolios be positioned.
1. Why are the AI labs focusing on safety – many things can be true at once
a) Genuine safety concerns given rapid model development: Labs have discussed the possibility that systems become powerful enough to be difficult to control and pose a threat to human safety (or cause significant financial harm) for years. Recently, alarm bells rang after a July incident in which a swarm of agents running inside an OpenAI safety test appeared to have coordinated on their own, broken out of their sandbox and hacked into another company (Hugging Face)’s systems (without being asked).2
b) Financial motivations given upcoming IPOs: Some have suggested that with public listings on the horizon, the labs have reason to slow things down and improve their financial metrics. This includes by shifting the focus onto model inference over training (where the economics are better) and quantifying any potential liabilities in the event of safety incidents.
c) Political motivations given souring public opinion: Data centers are drawing significant public (bipartisan) opposition, with a recent Gallup survey showing that 80% of U.S. adults support AI safety rules over faster development. Local proposals for AI regulation are growing with 1,561 AI-related bills introduced in U.S. state legislatures so far this year3. AI regulation looks like a question of when, not if, and there may be an advantage to the model labs in proposing the rules themselves, and potentially stifling competition (especially from open-weight Chinese models and smaller players).
2. What a slower frontier means for the pace of capex – training vs. inference
Current consensus estimates point to hyperscalers spending nearly $800 billion on capex this year and over $1 trillion next year. This spending has been hugely beneficial to the “picks and shovels” of the AI buildout, including semiconductors, hardware and power, which are seeing explosive earnings and returns so far this year. As a result, one of the most important questions for all investors is what will happen to these lofty capex assumptions?
The reality is that most AI compute is no longer spent on training models: Inference (the work of running models that already exist) is expected to account for roughly two-thirds of AI compute this year, up from about a third in 2023.4 A slower race to the frontier could shift the mix toward inference further, slowing related capex. Training is the marginal buyer of the newest and most expensive AI hardware systems, so certain “picks and shovels” names could see a deceleration in orders.
However, zooming out a bit: the pace of capex is going to be determined by broader AI adoption across the economy. In this case, a faster shift towards inference may actually be beneficial as it can improve the unit economics for the AI labs. The economics for inference have always been more clearly beneficial for the AI labs than training as it increases the commercial life of each model generation and raises the return on investment of those models. As a result, AI labs can focus more on the cost and practical applications of models.
When it comes to regulation itself, it’s too early to tell what it looks like in practice. Such policies could impact the contours of the AI race but are unlikely to derail AI adoption altogether.
3. How markets are reacting – builders vs. adopters
On Monday, semiconductor stocks fell 6%, but the AI theme as a whole was not for sale, with cybersecurity, software and hyperscaler names all faring better. While assumptions for certain “picks and shovels” may need to be tweaked, hyperscalers could see the opposite effect, with some relief around the magnitude of their capex spend and the related pressure on free cash flow. Additionally, software companies could benefit from a renewed focus on adoption, and a slower frontier lowers the risk that the next release could absorb what they have built on top of it.
4. What this means for portfolios – AI theme is not a monolith
Even the most bullish investors on AI still need to think really carefully about portfolio construction. This includes position sizing, concentration, leverage and quality. As shown by July’s momentum unwind – and the current volatility episode – market indigestion related to the AI development, buildout, adoption and regulation is a feature not a bug of the AI secular theme (and is becoming more frequent). The winners and losers are in a state of flux even amidst a still robust AI demand environment.
For investors, this is a reminder of the need to build resilient portfolios that have diversification within and away from the theme. True diversification from the “AI theme” is becoming harder to find – but includes government bonds, gold and real assets (core real estate and infrastructure).