David Sacks: Anthropic, OpenAI Should Slow AI If Deemed Necessary

David Sacks said Anthropic and OpenAI should voluntarily slow development of frontier AI models if they determine the systems pose serious risks, rather than relying on new regulation. He opposed antitrust exemptions and a broader regulatory framework for AI labs, arguing the companies already control the pace of frontier-model development. Sacks said U.S. policy must also account for competition with China, which is unlikely to participate in a global agreement to curb AI development.

Analysis

The investable read-through is a lower near-term probability of a U.S. regulatory pause or mandatory frontier-model licensing regime. That modestly reduces policy-discount risk for hyperscaler AI capex—MSFT, GOOGL, AMZN and META—and for the compute supply chain led by NVDA, AVGO, ANET, VRT and CEG. The effect is primarily multiple support rather than an immediate earnings revision: enterprise deployment and inference monetization, not model-training freedom, remain the binding variables for 2026 estimates.

A permissive domestic posture also raises the likelihood that competition shifts from a small number of proprietary frontier models toward distribution, data access and power availability. MSFT and GOOGL retain the strongest enterprise channels, while META benefits if open-weight models face fewer restrictions; the second-order loser is smaller standalone model vendors whose safety/regulatory moat thesis becomes less defensible. For infrastructure, sustained model iteration extends demand duration, but it may intensify the existing bottlenecks in grid interconnection and data-center power, favoring VRT, ETN, PWR and selected independent power producers over software names priced for immediate AI revenue conversion.

The consensus risk is treating less regulation as unambiguously bullish for all AI equities. Faster competitive cycles can compress model-layer pricing and force higher capex without proportionate returns; this is most relevant to MSFT and GOOGL if inference costs remain high. Over the next 1-3 months, watch for White House executive actions, Commerce export-control changes, and hyperscaler capex commentary; a new binding safety regime or China-linked security incident would reverse the policy relief quickly. Over 6-18 months, the key falsifier is a deceleration in cloud AI workload growth despite continued capex, which would expose excess infrastructure ordering.

Key Decisions for Investors

• Maintain a 1-3 month overweight in AI infrastructure versus model/application exposure: long VRT and ETN, paired against an equal-dollar short in a broad software proxy (IGV). The trade captures power-and-cooling bottlenecks while reducing risk from slower enterprise AI monetization; reassess if hyperscaler capex guidance falls below current consensus or data-center lead times normalize.
• Use any policy-driven weakness in NVDA and AVGO to build, not chase, 6-12 month exposure. The relevant confirmation is continued aggregate capex acceleration from MSFT, GOOGL, AMZN and META; invalidate the thesis on two consecutive quarters of materially reduced AI infrastructure budgets or a meaningful restriction on advanced-chip domestic deployment.
• Avoid adding to pure-play frontier-model valuation narratives absent disclosed recurring revenue and inference-margin data. A less restrictive policy environment lowers barriers to iteration but does not create pricing power; favor distribution incumbents MSFT and GOOGL over unlisted/early-stage model-layer proxies.
• Set an event-risk alert around export-control or AI-security announcements. A China-related incident could widen semiconductor restrictions and harm NVDA/AMD near term even while supporting U.S. domestic infrastructure spend; reduce gross AI beta if regulatory rhetoric shifts from voluntary commitments to mandatory licensing or compute thresholds.