Nvidia expects $1T chips
NVIDIA's CEO told investors at GTC that the company expects to make roughly $1 trillion in AI chip revenue through 2027, and the company doubled down on inference and Rubin/Vera platforms as the next growth engines reported. That roadmap pressure is shaping supply, partner deals, and buying decisions across seed‑to‑Series B AI teams.
Vera Rubin is a rack‑scale platform now stated to be in full production, with seven chips that include the NVL72 GPU rack, a Vera CPU rack, Groq 3 LPX inference racks, BlueField‑4 STX storage and Spectrum‑6 SPX Ethernet components, NVIDIA [announced]. (nvidianews.nvidia.com) OpenAI, Anthropic and Meta are listed as early Vera Rubin partners or integration customers, according to coverage of the GTC rollout. (venturebeat.com) Starburst and other software vendors announced day‑one optimizations for Vera CPU and NVL stacks, and Pegatron said it will ship NVL72 and HGX Rubin systems for customers later this year. (marketwatch.com) NVIDIA disclosed an order backlog that exceeded roughly $500 billion for 2025–2026 in late‑2025, a figure analysts and reporters have linked to its ability to promise large future volume. (cnbc.com) Wall Street has fluctuated around those disclosures, with analysts and market coverage flagging renewed investor scrutiny after recent earnings and guidance cycles. (bloomberg.com) Multiple supply‑chain moves surfaced alongside the GTC announcements: independent reports say NVIDIA cut GPU allocations to add‑in‑card partners by roughly 15–20% amid constrained components, and other outlets reported changes to VRAM bundling practices for board partners. (videocardz.com) Meanwhile, suppliers including Samsung have been linked to HBM4 memory modules expected to support the new racks’ bandwidth needs. (techradar.com) NVIDIA added a co‑designed Groq 3 language processing unit (LPU) and Groq LPX inference racks to the Rubin family to address low‑latency, high‑concurrency token decoding for agentic systems. (developer.nvidia.com) At the same time, cloud‑provider ASICs — Google’s TPUs and AWS’s Trainium family — and emerging inference specialists continue to position custom silicon as a cost‑efficient alternative for large deployers. (cnbc.com)
