At GTC 2026, NVIDIA CEO Jensen Huang delivered a keynote predicting $1 trillion in chip order backlog by year-end while unveiling multiple products that mark NVIDIA's expansion beyond GPUs into broader computing territory.
NVIDIA Groq 3 LPU Debuts
NVIDIA launched the Groq 3 LPU—an inference accelerator developed from last year's Groq acquisition. Each Groq 3 LPU integrates 500 MB SRAM with 150 TB/s bandwidth (far exceeding HBM's 22 TB/s), optimized for low-latency, high-throughput token delivery in multi-agent systems.
The Groq 3 LPX rack packs 256 LPUs, delivering 128GB SRAM and 40 PB/s inference bandwidth—boosting agent-to-agent communication from 100 to over 1,500 tokens per second when paired with Rubin platforms.
New 88-Core Vera CPU Challenges AMD/Intel
NVIDIA unveiled detailed specs for its 88-core Vera data center CPU—marking its direct entry into CPU sales. Built on NVIDIA's custom Arm v9.2-A Olympus cores, it delivers 1.5x IPC throughput, 50% higher performance than standard CPUs, 3x per-core memory bandwidth, and 2x better efficiency.
The 88-core, 144-thread design features 1.2 TB/s mesh memory bandwidth, 80 GB/s per-core throughput, and 1.8 TB/s NVLink-C2C interconnect—far exceeding PCIe 6.0 speeds. The Vera CPU rack architecture integrates 256 liquid-cooled CPUs with 300 TB/s total memory throughput and 6x CPU throughput improvement. Meta, Alibaba, and Oracle have already signed on, with Dell and Lenovo servers shipping in H2 2026.
Vera Rubin Space Module & More
NVIDIA also unveiled the Vera Rubin space module (25x H100's orbital inference capability), IGX Thor for edge environments, Jetson Orin for satellites, and the RTX PRO 6000 Blackwell server GPU (100x performance boost).
ICgoodFind: NVIDIA's GTC unveilings mark its evolution beyond GPUs into a comprehensive computing powerhouse, reshaping the competitive landscape.
