Nvidia pivots: perf-per-watt is king
Nvidia’s GTC signals a shift from raw GPU count to full-stack efficiency—networking revenue surged 263% and Nvidia is pushing performance-per-watt and token-factory economics as the new battleground. That changes how AI infrastructure is designed and pressures device teams to optimize energy for inference, not just peak throughput. (developer.nvidia.com) (hpcwire.com)
NVIDIA’s Q4 fiscal 2026 results showed Data Center Networking sales of $10.98 billion, a 263% year‑over‑year increase that helped drive record quarterly revenue of $68.1 billion. (nvidianews.nvidia.com) The Vera Rubin NVL72 rack pairs 72 Rubin GPUs with 36 Vera CPUs and integrates NVLink‑6 switching, ConnectX‑9 SuperNICs and BlueField‑4 DPUs in a liquid‑cooled rack reference design. (nvidia.com) NVIDIA and third‑party benchmarks presented at GTC and CES claim Vera Rubin delivers up to 10× inference throughput per watt and roughly one‑tenth the cost per token versus the prior Blackwell generation at rack scale. (storagereview.com) Microsoft posted that it has powered on a Vera Rubin NVL72 system for validation, while cloud and hyperscaler commitments — which NVIDIA says account for just over 50% of data‑center revenue — are cited as immediate demand drivers for the new racks. (datacenterdynamics.com) NVIDIA framed the new economics at GTC with a simple equation — Revenue = Tokens per Watt × Available Gigawatts — positioning “tokens per watt” as the metric enterprises must maximize across compute, networking (Spectrum‑X/InfiniBand) and DPUs to increase revenue within fixed power envelopes. (developer.nvidia.com) Jensen Huang told GTC attendees the company is already seeing order momentum consistent with as much as $1 trillion of Blackwell‑ and Vera‑era demand through 2027, a claim that accelerates procurement and facility‑level power planning for data‑center operators. (cnbc.com)
