NVIDIA GTC 2026: what was announced, what ships, and what is roadmap

NVIDIA's GTC conference ran March 16 to 19 in San Jose. Most of the headline news landed in the first-day keynote and a batch of same-day press releases. This analysis sorts what was announced into three groups: available now, announced with partner availability later in 2026, and roadmap. Nearly every performance figure comes from NVIDIA itself and has no independent test behind it yet.

The platform: seven chips, five racks

At the keynote, NVIDIA presented Vera Rubin as seven chips: the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch and a newly integrated Groq 3 LPU ^[1]. They are packaged into five rack types ^[1]:

• Vera Rubin NVL72: 72 Rubin GPUs and 36 Vera CPUs. NVIDIA claims large mixture-of-experts models train on one-fourth the GPUs of Blackwell, and up to 10x higher inference throughput per watt at one-tenth the cost per token. Only the training claim names Blackwell as the baseline ^[1].

• Vera CPU rack: 256 liquid-cooled CPUs. Each Vera CPU has 88 custom cores and up to 1.2 TB/s of memory bandwidth ^[2].

• Groq 3 LPX: 256 LPU processors with 128 GB of on-chip SRAM, claimed at up to 35x higher inference throughput per megawatt, with no baseline stated ^[1].

• BlueField-4 STX storage: a reference architecture for the key-value cache data that long-context agents generate, claimed at up to 5x more tokens per second than traditional storage ^[3].

• Spectrum-6 SPX networking: with co-packaged optics, claimed at up to 5x better optical power efficiency than pluggable transceivers ^[1].

The Rubin GPU itself, per NVIDIA's technical blog, has 336 billion transistors, up to 288 GB of HBM4 and 50 petaflops of NVFP4 inference compute ^[10]. The ratings are NVIDIA's and measured its way.

Demand and the "$1 trillion" line

Jensen Huang said he now sees at least $1 trillion in revenue from 2025 through 2027 ^[8]. Jon Peddie Research described the same figure as Blackwell and Vera Rubin purchase orders through 2027, double a $500 billion figure from a year earlier ^[9]. Those are different things, so treat the number as a statement of NVIDIA's expectation, not a reported result.

What ships, and what does not

The "full production" language deserves care. It describes NVIDIA's manufacturing, not what a customer can order. NVIDIA's own release says partner products arrive from the second half of 2026 ^[1], and one analyst write-up says Rubin "ships to customers later in 2026" ^[9]. Azure was named as the first hyperscale cloud to power up Vera Rubin NVL72 systems ^[8], but that is not general availability.

Other announcements

NVIDIA also announced DSX, a reference design and software for AI-factory power, claiming 30% more infrastructure in a fixed-power data center, and a claimed 100 gigawatts of unlocked grid capacity ^[1]. On the workstation side, DGX Station with GB300 offers 748 GB of coherent memory and models up to 1 trillion parameters, with first units delivered ^[8]. IGX Thor became generally available for industrial edge AI ^[8]. The press kit also lists open model families, robotics and autonomous-vehicle partnerships; we have not evaluated those ^[8].

What this means for buyers

1. Plan on cloud, not purchase orders. Early Vera Rubin capacity is aimed at hyperscalers and large AI clouds, so most businesses will meet it as instance availability and pricing.

2. Hold NVIDIA's multipliers loosely. Claims such as 10x and 35x have unstated or partial baselines, and NVIDIA's own release says features may be offered "when-and-if-available" ^[1].

3. Use the software now. Dynamo is available today ^[4], and agent runtimes such as OpenShell are worth evaluating, but check their maturity first.

4. Do not plan around 2027 and 2028 items. Kyber and Feynman are direction, not products.

For most businesses, the practical implication is to track provider pricing and instance availability through the second half of 2026 and to keep model and hosting choices portable, rather than committing to any single vendor's hardware roadmap.