Nvidia is reinforcing its leadership in artificial intelligence with the introduction of a new series of AI GPUs that aim to extend the frontiers of computational capability. Launched at the Nvidia GPU Technology Conference (GTC) 2025, this roadmap features the Blackwell Ultra GB300 (scheduled for shipment later this year), with the Vera Rubin (2026) and Rubin Ultra (2027) expected to follow, each offering significant performance enhancements.
With revenues skyrocketing to $2,300 per second, Nvidia’s data centre operations have overtaken its gaming GPU division, marking a major change in the company’s focus.
Nvidia founder and CEO Jensen Huang stated that “the industry now requires 100 times the resources we estimated just a year ago,” underscoring the relentless need for AI computational capacity.
Highlights
Blackwell Ultra: A Significant Mid-Cycle Update
The Blackwell Ultra GB300, which is to launch in the latter part of 2025, represents an enhancement of the existing Blackwell architecture rather than an entirely new model.
- AI Performance: 20 petaflops of FP4 inference (identical to Blackwell).
- Memory Capacity: 288GB of HBM3e memory (up from 192GB).
- Enterprise Cluster: DGX GB300 “Superpod,” delivering 11.5 exaflops of FP4 computing and 300TB of memory (increased from 240TB in Blackwell).
Vera Rubin & Rubin Ultra: The Future of AI Computing
Nvidia anticipates launching the Vera Rubin architecture in late 2026, followed by the Rubin Ultra in 2027.
- Vera Rubin: 50 petaflops of FP4 performance – 2.5 times quicker than Blackwell Ultra.
- Rubin Ultra: Combines dual Rubin GPUs to achieve 100 petaflops of FP4 alongside an impressive 1TB of HBM memory.
- NVL576 Rack System: Offers 15 exaflops of FP4 inference and 5 exaflops of FP8 training – delivering 14 times the performance of Blackwell Ultra.
These advancements have the potential to significantly minimise AI processing durations, with Nvidia asserting that the NVL72 cluster operating DeepSeek-R1 671B will yield responses in just 10 seconds – ten times faster than the 2022 H100 GPU.
New Hardware: The DGX Station and NVL72 Rack
For organisations seeking a single-unit AI processing solution, Nvidia has unveiled the DGX Station, a high-performance desktop that includes:
- One GB300 Blackwell Ultra GPU.
- 784GB of unified system memory.
- 800Gbps Nvidia networking.
- 20 petaflops of AI capability.
The NVL72 Rack is designed as a powerhouse within a single rack, providing 1.1 exaflops of FP4 computing, 20TB of HBM memory, and networking speeds of 14.4TB/sec.
Nvidia’s AI Leadership and Future Prospects
These latest announcements from Nvidia are set against a backdrop of reports indicating that the company has dispatched $11 billion worth of Blackwell GPUs in 2025, with major customers acquiring a total of 1.8 million units thus far.
Looking down the line, Huang announced that the 2028 GPU architecture would be named “Feynman,” in honour of the esteemed physicist Richard Feynman, highlighting Nvidia’s enduring dedication to innovation in AI.
