AI safety efforts will require more compute, not less: experts

Citi says AI safety, alignment and agentic-AI workloads are likely to increase infrastructure spending rather than curb hardware demand, with Nvidia estimating agentic workloads require roughly 100x the compute of 2023-era chatbot applications. Next-generation data centers are being designed for 30%-40% more GPUs, while Meta is targeting expansion from its 1GW-plus Prometheus cluster to a 5GW Hyperion footprint over coming years. Intel is repositioning as a full-stack AI infrastructure provider, with its 18A node in volume production and 14A nearing launch, underscoring sustained demand for compute, networking, power and data-center capacity.

Analysis

The investable implication is a broadening of AI capex away from accelerators toward the constraints that determine usable cluster uptime: high-radix switching, optical interconnect, power delivery, cooling and grid equipment. NVDA retains the highest near-term earnings sensitivity, but its incremental dollar of spend increasingly leaks to AVGO, ANET, CRDO, VRT, ETN and GEV; these suppliers face less risk of model-level commoditization and may enjoy longer project backlogs. The key second-order effect is that reliability requirements favor premium, validated components over lower-cost alternatives, supporting margins across the networking and electrical stack.

META's spending trajectory is constructive for the supply chain but less unambiguously positive for META equity over 1-3 months: investors will require evidence that inference, ranking and advertising productivity monetize ahead of depreciation and energy expense. A rising capex budget can compress its FCF yield and multiple if quarterly ad-revenue acceleration does not offset it. For NVDA, the next catalyst is not another demand claim but whether hyperscaler orders remain firm through the platform-transition cycle; any evidence of deployment delays caused by power, networking, or facility readiness would shift revenue timing rather than necessarily eliminate demand.

INTC's strategic narrative remains an execution trade, not a capacity trade. A process-node milestone has limited valuation significance without disclosed external design wins, foundry gross-margin progression, and credible AI-system revenue; meanwhile its capital intensity leaves less room for error than fabless peers. The contrarian view is that the market may still be underpricing physical infrastructure scarcity, but is overpricing the speed at which agentic workloads translate into enterprise revenue and therefore sustained hyperscaler ROI.

Key Decisions for Investors

• Initiate a 3-6 month basket long AVGO, ANET and VRT versus a short SOXX hedge: target 10-15% relative upside if AI deployment bottlenecks shift spend into networking and power; exit if hyperscaler capex guidance is cut or lead times normalize materially.
• Maintain NVDA as a core long into the next earnings cycle, but fund part of the exposure with a 1-3 month NVDA put spread following sharp rallies. The thesis is intact, but facility-readiness delays are the principal near-term risk to shipment timing and a premium multiple.
• Use META as a watch-item rather than add aggressively before the next results: add only if ad pricing/revenue growth and operating-margin guidance absorb higher depreciation. A capex increase without monetization evidence is a 1-3 month multiple-compression risk.
• Avoid treating INTC as a value long solely on manufacturing roadmap milestones. Reassess after externally verifiable 18A customer commitments, foundry margin disclosure, or AI-platform revenue guidance; absent these, downside remains tied to cash burn and execution slippage.