Modal plans to use Nvidia Vera processors for AI agents for the first time — Channel NewsAsia - UA.NEWS

Modal's decision to deploy Nvidia's Vera CPU rather than a GPU for AI agent workloads signals a structural break in inference architecture. Agent workloads are orchestration problems—multi-step reasoning, tool invocation, context scheduling—where memory bandwidth and latency matter far more than raw FLOPS. Vera sits precisely in that sweet spot. The industry is moving past the GPU-everything paradigm toward genuine heterogeneous CPU-GPU division of labor. Upstream, Nvidia's CPU business shifts from supporting component to primary revenue driver. Downstream, inference frameworks must re-architect CPU-GPU data paths from scratch. Risk exposure is acute: Modal's single-vendor lock-in, compounded by TSMC (Taiwan, China) foundry dependency, creates concentrated geopolitical fragility. Any export-control expansion to the CPU side would force a costly rebuild with minimal switching runway. Competitively, AWS Graviton and AMD EPYC 9005 are theoretical alternatives, but the CUDA-plus-NVLink ecosystem makes migration prohibitively expensive for independent clouds. Expect CoreWeave and Lambda Labs to follow within two quarters; Azure will accelerate Maia adaptation for agent-specific paths. Within 18 months, agent-native infrastructure becomes a distinct product category. CPU share of AI-stack revenue should climb from roughly 15% to over 30%, with Nvidia's CPU revenue crossing the 20% threshold for the first time.