NVIDIA’s Post‑GPU Push
At GTC 2026 NVIDIA unveiled a three‑system strategy—Groq LPX inference racks, Vera ETL256 CPU racks, and an STX storage reference architecture—signaling a move to own more of the AI stack beyond GPUs. That vertical push reshapes enterprise AI benchmarks and is a competitive signal for Apple to reassess accelerator and data‑infrastructure scenarios. (digitimes.com)
Groq 3 LPX racks are built around 256 LPUs (LP30 series) assembled into a single inference system. (theregister.com)) The Groq 3 LPU architecture exposes roughly 500 MB of on‑chip SRAM and ~150 TB/s internal bandwidth, contrasted with NVIDIA Rubin GPUs that ship with 288 GB of HBM and about 22 TB/s of memory bandwidth. (forbes.com)) NVIDIA’s Vera CPU rack reference packs up to 256 Vera processors in a liquid‑cooled MGX design, totaling more than 22,500 CPU cores and roughly 400 TB of system memory in a single rack configuration. (theregister.com)) Those Vera CPU racks are integrated into the Vera Rubin POD as one of five purpose‑built rack classes on NVIDIA’s third‑generation MGX rack architecture for agentic AI workloads. (developer.nvidia.com)) BlueField‑4 STX was announced on March 16, 2026 as a modular, accelerated storage reference architecture centered on a storage‑optimized BlueField‑4 DPU and ConnectX‑9 SuperNIC. (investor.nvidia.com)) NVIDIA and partner messaging credit STX with up to ~5× token throughput and up to ~4× energy efficiency for long‑context inference, and lists early adopters including CoreWeave, Lambda, Mistral AI, Oracle Cloud Infrastructure and Vultr. (storagenewsletter.com)) NVIDIA is positioning STX as a reference architecture co‑designed with storage OEMs (Dell, NetApp, HPE, IBM, VAST, Weka and others) rather than a direct‑sell appliance, while company channel leadership says Groq LPX and Vera racks will be pushed through partners “over time.” (storagereview.com)) The systems were unveiled at GTC on March 16, 2026, and NVIDIA has indicated Groq LPX, Vera CPU racks and BlueField‑4 STX will be available or shipping in the second half of 2026. (blogs.nvidia.com)) Public reporting previously put Apple in the market for a large AI server purchase—an analyst‑cited ~$1 billion NVIDIA order reported in 2025—which intersects with NVIDIA’s push to extend from GPUs into inference, CPU and storage infrastructure. (9to5mac.com)) NVIDIA’s strategic moves are sizable: the company’s Groq acquisition and integration were reported as a ~$20 billion bet, and NVIDIA has publicly sketched a pathway to roughly $1 trillion in AI‑related revenue by 2027, metrics that materially change vendor leverage in accelerator and data‑infrastructure sourcing decisions. (forbes.com))
