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

Contrary to investor concerns that stricter safety requirements could dampen hardware demand, leading frontier model developers at the AI Infra Summit said AI safety and alignment are deeply compute-intensive and are expected to drive higher infrastructure spending, Citi said in a research report.

That conclusion came from the first day of the Santa Clara conference, where analysts observed that obtaining large amounts of compute, power and facility capacity—not model architecture innovation alone—has become the industry’s main bottleneck.

Next-generation data centers are being designed to accommodate 30% to 40% more GPUs, while orchestration software such as Astra and hardware upgrades including Nvidia’s Vera Rubin platform are intended to improve end-to-end system throughput.

Nvidia flags 100x demand jump from agentic AI

At the summit, Ian Buck, Nvidia Corp.’s vice president of Hyperscale and HPC, described how the move from conversational chatbots to autonomous “agentic AI” is multiplying hardware requirements.

Buck said agentic workloads are about 100 times more demanding than 2023-era chat applications, pointing to exponential growth in input sequence lengths, KV cache sizing, multi-turn interactions and sub-agent spawning.

Intel repositioning and Meta’s 5GW vision

Intel Corp. CEO Lip-Bu Tan spoke at the summit, outlining the chipmaker’s strategic shift from a traditional CPU vendor to a full-stack AI infrastructure provider.

Tan stressed that as competition moves from raw GPU horsepower to system-level optimization, CPUs will remain essential for reinforcement learning, orchestration layers and agentic workflows. To speed up decision-making, Tan flattened Intel’s management structure, reducing hierarchy from 10–12 layers to 4–5.

On manufacturing, Tan confirmed Intel’s 18A process node is in volume production, while the more advanced 14A node nears launch, backed by a 0.9 PDK release expected in October and a 1.0 PDK in the next quarter.

Meanwhile, Meta Platforms Inc. Head of Infrastructure Santosh Janardhan detailed the company’s experience scaling large-scale training infrastructure.

As Meta advances from its 1-gigawatt-plus “Prometheus” cluster toward a 5-gigawatt “Hyperion” footprint over the coming years, Janardhan identified multi-region, lossless networking as the single biggest operational hurdle.

To meet its goal of limiting GPU cluster interruptions to just two per 8,000 GPUs per day, Meta deployed a proprietary Back-End Aggregation (BAG) networking architecture.