Nvidia is doubling down on its dominance in artificial intelligence, unveiling a new lineup of AI GPUs designed to push the limits of computing power. Announced at the Nvidia GPU Technology Conference (GTC) 2025, the roadmap introduces the Blackwell Ultra GB300 (shipping later this year), followed by the Vera Rubin (2026) and the Rubin Ultra (2027), each promising exponential performance gains.
With profits soaring to $2,300 per second, Nvidia’s data centre business has now surpassed its gaming GPU segment, signalling a seismic shift in the company’s priorities.
“The industry needs 100 times more than we thought we needed this time last year,” said Nvidia founder and CEO Jensen Huang, emphasising the insatiable demand for AI computing power.
Blackwell Ultra: A Powerful Mid-Cycle Refresh
The Blackwell Ultra GB300, set to launch in the second half of 2025, serves as an upgrade to the original Blackwell architecture rather than an entirely new design. • AI Performance: 20 petaflops of FP4 inference (same as Blackwell). • Memory Expansion: 288GB of HBM3e memory (up from 192GB). • Enterprise Cluster: DGX GB300 “Superpod” with 11.5 exaflops of FP4 computing and 300TB of memory (up from 240TB in Blackwell).
Vera Rubin & Rubin Ultra: The Future of AI Processing
Vera Rubin, Nvidia’s next major GPU architecture, is expected to debut in late 2026, with Rubin Ultra arriving in 2027. • Vera Rubin: 50 petaflops of FP4 performance - 2.5x faster than Blackwell Ultra. • Rubin Ultra: Dual Rubin GPUs fused into one, delivering 100 petaflops of FP4 and a staggering 1TB of HBM memory. • NVL576 Rack System: 15 exaflops of FP4 inference and 5 exaflops of FP8 training - 14x the performance of Blackwell Ultra.
These upgrades could dramatically reduce AI processing times, with Nvidia claiming the NVL72 cluster running DeepSeek-R1 671B will now generate responses in 10 seconds - 10x faster than the 2022 H100 GPU.
New Hardware: The DGX Station and NVL72 Rack
For companies looking for single-unit AI processing, Nvidia introduced the DGX Station, a high-performance desktop featuring:
• A single GB300 Blackwell Ultra GPU. • 784GB of unified system memory. • 800Gbps Nvidia networking. • 20 petaflops of AI performance.
The NVL72 Rack will be a single-rack powerhouse offering 1.1 exaflops of FP4 computing, 20TB of HBM memory, and 14.4TB/sec networking speeds.
Nvidia’s AI Dominance and the Road Ahead
Nvidia’s latest announcements come amid reports that the company has already shipped $11 billion worth of Blackwell GPUs in 2025, with top buyers purchasing 1.8 million units so far.
Looking further ahead, Huang confirmed that Nvidia’s 2028 GPU architecture would be named “Feynman,” after the legendary physicist Richard Feynman, signalling Nvidia’s long-term commitment to AI innovation.