Nvidia: platform and backlog
Nvidia’s new Blackwell platform is running at a PUE of 1.25 with ~20% of capacity air‑cooled, and the company says orders for Blackwell/Rubin chips extend through 2027 — pointing to multi‑year demand for AI inference. Networking revenue is surging (reported +263%), and analysts see Nvidia shifting from chips to high‑value platform contracts. (developer.nvidia.com, finance.yahoo.com)
At GTC 2026 CEO Jensen Huang said Nvidia now sees roughly $1 trillion in orders or revenue opportunity tied to its Blackwell and Vera Rubin families through 2027. (cnbc.com) Nvidia describes Vera Rubin as a rack‑scale “AI factory” that combines Rubin GPUs, a Vera CPU, NVLink 6 fabric, ConnectX‑9 NICs, BlueField‑4 DPUs, Spectrum‑6 Ethernet and Quantum‑X800 InfiniBand with a Groq 3 inference rack in a single reference architecture. (nvidianews.nvidia.com) The company reported about $11.0 billion in networking revenue in the latest quarter — up ~263% year‑over‑year — and said full‑year networking exceeded roughly $31 billion, reflecting heavy attach rates for Spectrum‑X/InfiniBand hardware. (finance.yahoo.com) Nvidia disclosed that total supply‑related purchase commitments jumped to roughly $95.2 billion and that inventory rose about 8% quarter‑over‑quarter as the company pre‑booked wafer, package and capacity to meet multi‑year demand. (qz.com) Multiple analysts and industry write‑ups frame the announcements as a strategic shift toward selling integrated platforms and long‑term site or rack contracts — citing Vera Rubin partner tie‑ins with OpenAI, Anthropic and Meta and hyperscalers accounting for just over half of data‑center revenue. (venturebeat.com) Nvidia’s recent moves to lock in inference IP and talent — including a roughly $20 billion deal covering Groq assets and LPU technology — are being read as an effort to own inference economics as well as training silicon. (tomshardware.com)
