NVIDIA lined up compute, networking and DPUs on the same day, shifting the pitch from chip speed to the power budget of a whole AI factory.
Four posts landed on NVIDIA’s developer blog at once: performance per watt for Vera Rubin and Blackwell, long-context interactivity via Groq 3 LPX on Vera Rubin, Spectrum-X Ethernet, and BlueField-4 scale-in networking.
Four of the five driving items are NVIDIA’s own announcements, with little third-party benchmarking or named adoption yet. What is readable is the shared framing: the unit of competition is the energy cost of keeping agents running, not a single accelerator. Whether the claimed efficiency holds in practice is not confirmed.
GA timing for the Vera Rubin generation, and independent benchmarks or production deployments. Whether Spectrum-X and BlueField-4 get chosen alongside rival silicon will show if this is lock-in or an emerging standard.