NVIDIA: systems, not just GPUs
NVIDIA is publicly pivoting from a pure GPU vendor to a full‑stack systems company—selling data‑center pods, CPUs, LPUs and networking—and analysts on the GTC roundtables pegged NVIDIA’s AI‑chip opportunity near $1 trillion by end of 2027 ( ). Panelists also flagged an inference shift—roughly 80–85% of AI workloads will be inference in the next 1–2 years—which fuels demand for efficient inference hardware and NVIDIA’s new agent frameworks (OpenCLAW/NemoCLAW) plus NemoLLM and domain world models ( ).
Nvidia showed discrete rack SKUs at GTC including the Vera Rubin NVL72 CPU rack, a Blackwell GPU rack, a Groq‑LPU inference rack and a BlueField‑4 DPU storage/network rack as part of its new data‑center systems push. (crn.com)) CEO Jensen Huang told the GTC audience Nvidia sees at least $1 trillion of AI‑chip demand/orders through 2027, a forecast he framed around inference‑heavy deployments. (msn.com)) The company’s networking segment generated roughly $11 billion in revenue in the most recent fiscal quarter, a surge analysts and company filings say has turned networking into a second pillar alongside compute. (quartr.com)) Independent panelists cited at the GTC roundtables and in analyst writeups estimated an inference pivot where roughly 80–85% of AI workloads will be inference over the next one to two years. (fierce-network.com)) NemoClaw was unveiled as an open‑source agent platform that layers NVIDIA’s OpenShell runtime and policy guardrails to run always‑on, long‑running agents securely; the stack is already public on NVIDIA’s developer site and GitHub. (nvidia.com)) Nvidia also introduced Nemotron models and a Nemotron Coalition of eight AI labs; Nemotron‑3 variants include MoE architectures described in vendor benchmarks as having ~120 billion total parameters with roughly 12 billion active at inference to boost cost‑per‑token efficiency. (opentools.ai))
