Overview
- Nvidia announced on Monday, Aug. 24, that Groq 3 LPX racks are in full production and said Nebius will deploy the racks and bring them online before the end of 2026.
- SpaceXAI has officially committed to standalone Vera CPUs for Grok and to expand Vera Rubin acceleration, and Nvidia and SpaceX plan a space‑optimized Vera Rubin NVL72 for the Starmind satellite with a Q4 2027 launch window.
- Commercial demand has surged: AM Intelligence placed a roughly 9,000‑system order for Vera Rubin NVL72 racks for a Hyderabad AI factory with delivery targeted by Q1 2027.
- Nvidia claims Vera CPUs (88 Olympus cores, LPDDR5X up to 1.2 TB/s) and the Groq LPX + Vera Rubin pairing deliver large latency and efficiency gains but those performance and token‑cost figures come from vendor benchmarks and need independent verification.
- The announcements mark a shift to extreme codesign for agentic AI where CPUs handle orchestration and tool use, Rubin GPUs and Groq LPX handle large‑context and low‑latency decode work, and investor attention is focused on hyperscaler concentration, Rubin shipment ramps, and supply constraints.