AI Safety Efforts to Drive Higher Compute Demand, Industry Leaders Say

Efforts to improve the safety and alignment of advanced artificial intelligence systems will require a significant increase in computing power, a development that is expected to fuel further demand for AI hardware. This key insight from leading frontier model developers at the AI Infra Summit contradicts some investor concerns that stricter safety regulations could stifle infrastructure spending, according to a research note from Citi.

The Rising Cost of AI Safety

Discussions at the Santa Clara conference highlighted that the primary bottleneck for the AI industry is no longer just model architecture but the ability to secure massive compute, power, and facility capacity. According to the Citi report, developers revealed that safety-related processes are heavily compute-intensive.

This demand is already influencing infrastructure design, with next-generation data centers being planned to accommodate 30% to 40% more GPUs. The industry is also focused on boosting efficiency through software orchestration platforms and hardware upgrades like Nvidia’s upcoming Vera Rubin platform to maximize system-level throughput.

Nvidia: 'Agentic AI' Is 100x More Demanding

Ian Buck, Nvidia Corp.’s Vice President of Hyperscale and HPC, explained that the shift from conversational chatbots to more autonomous "agentic AI" is causing a massive surge in hardware requirements. Speaking at the summit, Buck noted that agentic workloads are roughly 100 times more demanding than the chat applications common in 2023.

Ad

This exponential increase is driven by several factors:

• Significantly longer input sequence lengths
• Larger Key-Value (KV) cache sizes
• More complex multi-turn interactions
• The spawning of multiple sub-agents to complete tasks

Infrastructure Giants Detail Scaling Roadmaps

Major technology firms outlined their strategies to meet this growing demand. Intel Corp. CEO Lip-Bu Tan detailed the company's pivot toward becoming a full-stack AI infrastructure provider, emphasizing that CPUs remain critical for reinforcement learning and agentic workflows. Tan also confirmed that Intel's 18A process node is in volume production, with the more advanced 14A node approaching its launch.

Meanwhile, Meta Platforms Inc. is planning a massive expansion of its AI infrastructure, moving from its current 1-gigawatt-plus "Prometheus" cluster toward a 5-gigawatt "Hyperion" footprint. Santosh Janardhan, Meta's Head of Infrastructure, identified multi-region, loss-less networking as the single largest operational hurdle in this expansion.