Frontier 'pacing' calls may not spell an end to AI capex cycle

Podcast: Europe’s political pendulum swings wider, on Apple and Spotify (26 mins)
Video:Market Playbook | Why higher rates alone don't mean lower equities (6 mins)
Video: UBS Explains | What is happening in the US bond market? (5 mins)

AI-linked chip stocks came under pressure on Monday, after several leading US frontier artificial intelligence labs publicly backed efforts to pace development of the most advanced models. Leading North Asia memory stocks saw mid-single-digit declines, and the US Philadelphia semi index fell 5.9%, though pressure has eased into the Tuesday session.

Leaders from frontier labs over the weekend made the public case for stronger safeguards and coordination. This follows closed-door talks between these labs to form an “industry-led AI safety standards body,” according to The Information. The safety debate has intensified in recent weeks as departing “alignment” researchers at leading labs have warned that advanced AI development could threaten human survival.

A bipartisan US Senate effort to establish national AI rules is also progressing, according to Reuters, and California enacted standards for independent AI auditors last week. US President Donald Trump on Monday characterized the AI misuse fears as a "hoax" and suggested sufficient US regulations were already in place.

The high-profile, coordinated calls to pace AI development are certainly notable. But we would caution against equating stronger safeguards with an end to the AI capex cycle:

Pacing does not necessarily imply lower capex. First, AI compute demand stems from both training models and running them, a process known as inference, with the latter estimated to account for roughly two-thirds of demand. Importantly, inference demand is driven primarily by real-world adoption and monetization, in our view, with the practical challenges of integrating AI into existing enterprise operations a greater constraint than current model capabilities. Training demand, by contrast, is more directly tied to the costs and expected returns of developing more capable models. The safety proposal also explicitly distinguishes pacing from halting model training, and xAI’s Elon Musk confirmed on Sunday that further training was under way for its leading-edge Grok 4.8 model. OpenAI also recently reiterated its expectations for continued growth in memory demand. We retain our 2027 AI industry capex forecast of USD 1.2tr, a rise of 33% from our estimate of USD 900bn this year.

Regulation could reshape competition, not just development. This may represent the industry acknowledging a more challenging political backdrop into the US midterm elections, with voter concerns spanning employment and data center electricity and water consumption. The push to address safety concerns also appears to reflect growing internal pressure from employees within frontier AI labs, rather than political positioning alone. The labs’ efforts to establish common standards may reflect a desire to shape future AI regulation, such as limits on the liability of platforms for what their users do with AI. More stringent evaluation requirements could also favor larger, better-funded incumbents, while raising the barriers to entry for startups, and closing the door to overseas competitors who are unable to accept external oversight.

Pacing calls arrive at a less frothy moment for AI stocks. The SOXX traded at around 21 times forward earnings prior to Monday's decline, versus 33 times in June and its post-ChatGPT average of 24 times. The "Magnificent Six" group trades at roughly 23 times forward earnings, or about 25% below its October 2025 valuation. Earnings continue to grow fast, with the latest consensus estimates implying semiconductor earnings per share growth of 103% this year and 46% next year. Meanwhile, OpenRouter’s token volumes have risen about 176% since the end of June, pointing to growing AI usage.

So, we think the key question for investors is not whether frontier development slows, but whether AI demand and monetization will continue to expand. We believe the answer is still yes. We continue to favor a diversified approach across the AI value chain, combining infrastructure beneficiaries (including semiconductors, networking, power, and cloud) with larger platforms and software companies positioned to monetize adoption. Investors with concentrated holdings within AI may consider rebalancing toward this broader mix. Stronger AI safeguards may reshape competition, but the proposals so far do not establish that the AI capex cycle is ending. Stepping back, we remain constructive on equities overall, despite a more restrictive policy outlook, supported by earnings growth, AI investment, and a resilient economy.