NVIDIA GTC 2026 — Jensen Huang Unveils Vera Rubin GPUs, Groq 3 LPU, Kyber Racks, and the First Data Center in Space
NVIDIA's GTC 2026 keynote reveals the Vera Rubin GPU platform with 10x performance per watt, the Groq 3 LPU for 35x inference gains, and a $1 trillion Blackwell-to-Rubin opportunity.
The Biggest GTC Yet
NVIDIA CEO Jensen Huang took the stage at the San Jose Convention Center on March 16 for what may be the most consequential GPU Technology Conference keynote in the company's history. Over the course of two hours, Huang laid out a technology roadmap that stretches from next-generation GPU silicon to the first data center in orbit — and backed it all with a staggering revenue projection: purchase orders between Blackwell and Vera Rubin are expected to reach one trillion dollars through 2027.
That number — double NVIDIA's previous $500 billion projection — reflects the explosive demand for AI compute infrastructure from hyperscalers, sovereign AI initiatives, and enterprise deployments. It also underscores why NVIDIA's GTC has become the most important event on the AI industry calendar.
Vera Rubin: 10x Performance Per Watt
The centerpiece announcement is the Vera Rubin GPU platform, scheduled to ship later this year. Built on a next-generation architecture, Vera Rubin delivers 10 times more performance per watt than its predecessor, Grace Blackwell. For data center operators spending millions on electricity, that efficiency gain translates directly into lower operating costs and higher density deployments.
Huang also showed off a prototype of Kyber — NVIDIA's next rack architecture leap after Rubin. Kyber integrates 144 GPUs in compute trays arranged vertically instead of horizontally, representing a fundamental rethink of how AI compute infrastructure is physically organized. The Kyber design will ship as part of Vera Rubin Ultra, expected in 2027.
Groq 3 LPU: Inference at Unprecedented Speed
In one of the keynote's biggest surprises, Huang unveiled the NVIDIA Groq 3 Language Processing Unit — the company's first chip derived from the startup it largely acquired through a $20 billion asset purchase in December 2025. The Groq 3 LPX rack holds 256 LPUs and can increase the tokens-per-watt performance of Rubin GPUs by 35 times, making it a purpose-built inference accelerator that complements NVIDIA's training-focused GPU lineup.
For companies deploying large language models at scale, the combination of Rubin for training and Groq 3 for inference creates an end-to-end platform that addresses both halves of the AI compute equation.
Space-1: AI Compute in Orbit
Perhaps the most audacious announcement was Vera Rubin Space-1 — NVIDIA's plan to build the first data center in space. While details remain limited, Huang confirmed that the company has dedicated engineering teams working on the project, targeting applications in satellite communications, Earth observation AI, and space-based compute that reduces latency for global AI services.
Sources: CNBC (March 16, 2026), Tom's Hardware (March 16, 2026), NVIDIA Blog (March 16, 2026), TechCrunch (March 16, 2026)