Jensen Huang opened GTC 2026 in San Jose on 16 March by launching Vera Rubin, NVIDIA’s next data centre platform, and by lifting his demand forecast sharply. “I see through 2027 at least $1 trillion,” he said of demand for Blackwell and Rubin systems, double the $500 billion figure he gave a year ago.
The trillion dollar line
Keynotes are built around one number, and this was it. A year ago Huang talked about $500 billion. Standing on the same kind of stage this week, he doubled it and stretched the horizon to the end of 2027.
The figure needs careful reading. It is Huang’s view of the demand he can see for two product generations, stated on a stage. It is not revenue NVIDIA has reported, and it is not a guidance number from an earnings call. Still, when the company at the centre of AI hardware says it sees that much demand coming, every cloud provider, chip rival and data centre builder takes note.
What Vera Rubin actually is
The flagship system is the Vera Rubin NVL72, a full rack that combines 72 Rubin GPUs with 36 Vera CPUs. NVIDIA designs it to be bought and deployed as one unit rather than as a pile of separate servers.
The claims are aggressive. NVIDIA says the NVL72 can train large mixture of experts models, the kind of architecture where only part of the network is active for each token, with “one-fourth the number of GPUs” that Blackwell would need, and at “one-tenth the cost per token”. Those are NVIDIA’s own comparisons, and I would treat them as a ceiling until customers publish their own results.
Alongside it came the Groq 3 LPX, a rack built specifically for inference. NVIDIA claims “up to 35x higher inference throughput per megawatt” on models with a trillion parameters. The choice of metric says a lot. Throughput per megawatt is the number that matters when the limit on your data centre is the power connection, not the budget for chips.
Availability is the practical part. NVIDIA says AWS, Google Cloud, Microsoft Azure and Oracle will offer Vera Rubin from the second half of 2026, and that Anthropic, Meta, Mistral AI and OpenAI are adopting the platform.
Huang also looked further ahead. The next platform, Feynman, is planned for 2028, and gamers got a look at DLSS 5.
By the numbers
| Item | Figure | Source |
|---|---|---|
| Blackwell and Rubin demand Huang sees through 2027 | at least $1 trillion | Tom’s Hardware |
| Same figure a year earlier | $500 billion | Tom’s Hardware |
| Rubin GPUs per NVL72 rack | 72 | NVIDIA Newsroom |
| Vera CPUs per NVL72 rack | 36 | NVIDIA Newsroom |
| GPUs to train large MoE models, vs Blackwell | one-fourth (NVIDIA claim) | NVIDIA Newsroom |
| Cost per token, vs Blackwell | one-tenth (NVIDIA claim) | NVIDIA Newsroom |
| Groq 3 LPX inference throughput per megawatt | up to 35x (NVIDIA claim) | NVIDIA Newsroom |
Why it matters
For most people, the relevant figure is not the trillion. It is cost per token. Every chatbot answer, every coding assistant suggestion and every AI feature inside an app is paid for in tokens, and that cost sets the price developers see on their API bills. If Rubin gets anywhere close to NVIDIA’s claim, the cost of running large models should keep falling, though how much of that reaches customers is up to the cloud providers.
The power angle is the one I would underline for readers in the Gulf. Saudi Arabia and the UAE are building large AI data centre projects, and those projects are planned and permitted in megawatts. A platform that squeezes more work out of each megawatt changes how much AI capacity a given site can deliver, which matters as much as the chip count.
There is also a timing point for local developers and startups. The four big clouds all say second half of 2026. Rubin capacity will likely reach the largest customers first, so for most teams in the region the practical effect this year will be indirect, through model providers that train and serve on it.
What to watch
- GTC continues until 19 March, and partner sessions should show how cloud providers plan to package Rubin.
- The first independent training and inference results once Vera Rubin ships in the second half of 2026.
- Whether the $1 trillion view shows up in the orders and commitments that customers disclose.
- Feynman, pencilled in for 2028, as the next marker on the roadmap.
Sources
- NVIDIA Newsroom, announcement of the NVIDIA Vera Rubin platform, March 2026, https://nvidianews.nvidia.com/news/nvidia-vera-rubin-platform
- Tom’s Hardware, NVIDIA GTC 2026 keynote live blog with Jensen Huang, 16 March 2026, https://www.tomshardware.com/news/live/nvidia-gtc-2026-keynote-live-blog-jensen-huang