Premium: Modular Inference

Key Takeaways

• Agentic AI creates new scaling law boosting GPU and CPU demand
• NVIDIA added Groq's low‑latency chip as seventh component
• Five modular rack types enable disaggregated compute and storage
• Vera Rubin expected to add $20 B revenue by FY27
• New racks aim for 25% GPU uplift, 20% storage uplift

Pulse Analysis

Agentic AI is reshaping inference workloads by introducing a second scaling law that drives not only GPU consumption but also a surge in CPU cycles for orchestration, tool use, and code execution. This dual‑demand dynamic forces data‑center operators to rethink architecture, blending high‑throughput GPUs with specialized CPUs that can manage the complex decision‑making loops of autonomous agents. NVIDIA’s Vera Rubin platform directly addresses this shift, offering a cohesive stack that couples its traditional GPU power with purpose‑built CPUs and DPUs, ensuring end‑to‑end performance for next‑generation AI services.

The centerpiece of Vera Rubin’s strategy is the integration of Groq’s SRAM‑based inference chip, announced just weeks before CES and positioned as the seventh component in the lineup. Groq’s architecture excels at ultra‑low‑latency inference, enabling premium‑priced tiers that deliver near‑real‑time interactivity for per‑user applications. By packaging this chip alongside five modular rack configurations, NVIDIA gives customers the flexibility to mix and match compute, storage, and networking resources, optimizing each rack for specific workloads such as high‑throughput training, latency‑critical serving, or agentic orchestration.

Financially, NVIDIA frames Vera Rubin as an incremental driver of its $1 trillion GPU sales target through 2027, forecasting a 25% uplift from Groq, a 20% boost from AI‑optimized storage, and a 5% lift from its new CPUs. Management also estimates a $200 billion total addressable market for the standalone Vera line, with $20 billion in revenue expected by fiscal year 2027—roughly 5% of the overall mix. The rollout begins in October, positioning the platform to contribute heavily in Q4 2026 and beyond, and signaling a competitive moat that could pressure hyperscalers and emerging AI chip makers alike.