NVIDIA’s groundbreaking hardware technologies and AI are unlocking unprecedented computational power. At the NVIDIA GTC 2025, NVIDIA unveiled its Blackwell Ultra GPU designed for the “Age of Reasoning” at its 2025 GPU Technology Conference (GTC). AI accelerators like GPUs are well suited for AI training and inference due to parallel processing, which allows for many calculations to be performed simultaneously. Only 30% of the top 500 supercomputers relied on accelerated computing; today, 80% do. The Green 500 ranking of supercomputers by energy efficiency shows an even more pronounced trend.
Blackwell Ultra GPUs for the Age of Reasoning
AI reasoning models emulate how the brain thinks to render a conclusion, popularized by OpenAI’s o1, Google’s Gemini 2.0 Flash Thinking and DeepSeek’s R1 A1 models. Reasoning models improve responses to queries and more powerful GPUs improve the performance of these models. Blackwell Ultra GPUs are the next generation of the evolution of the GB200 bolstered by more inference power horsepower, packing 50% FLOPS at 1.1 exaFLOPS of FP dense compute.
At the NVIDIA GTC 2025, in his March 18 presentation titled, “The Next Frontier of AI Supercomputing: Efficiency With Unprecedented Capability”, NVIDIA’s Vice President of Hyperscale and HPC Computing, Ian Buck, stated, “Blackwell Ultra takes GB200’s 40x data center revenue opportunity to 50x”, citing faster token serving and higher throughput ideal for post-training for models like DeepSeek, which chomp through 100 trillion tokens.
NVIDIA GB300 NVL72 Unleashes Inference Horsepower
NVIDIA’s GB300 superchip combines two Blackwell Ultra GPUs with one Grace CPU. Blackwell Ultra GPUs can be used in the NVL72 rack server, which integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs. The NVIDIA GB300 NVL72 has a fully liquid-cooled rack-scale design. AI factories achieve 50X higher output for reasoning model inference with the NVIDIA GB300 NVL72 compared to the NVIDIA Hopper platform when used with the NVIDIA Quantum-X800 InfiniBand or Spectrum-X Ethernet paired with ConnectX-8 SuperNICS.
Blackwell Ultra’s Silicon Photonics Slashes Power Consumption by Up to 77%
NVIDIA’s Blackwell Ultra GPUs use co-packaged optics with silicon photonics, which integrates optical and silicon components onto a single substrate. This reduces power consumption by eliminating the need for external lasers and pluggable transceivers to achieve a significant reduction in power from 39 watts to 9 watts. Buck said that silicon photonics “… gives you that benefit from going from 30 watts of power down to only 9 watts of power for the same number of ports, and that’s huge. It doesn’t sound like 39 sounds a lot. But if you get 400,000 GPUs in an AI supercomputer, there’s like 24 megawatts of lasers like so that’s a lot of laser light that could be optimized and made more efficient.”
Join thousands of investors who trust I/O Fund’s expert stock analysis on AI, semiconductors, cryptocurrency, and adtech — sign up for free! Click here!
Beth Kindig, Lead Analyst at the IO Fund, pointed out in her “AI Power Consumption: Rapidly Becoming Mission-Critical” blog article that, “In my analysis last month on the Blackwell architecture, I made the argument these estimates are too low and that my firm expects we will see a $200 billion data center segment by end of CY2025 propelled forward by the B100, B200 and GB200, including the following points: “Taiwan Semi’s CoWos capacity, which is essential for Blackwell’s architecture, is estimated to rise to 40,000 units/month by the end of 2024, which is more than a 150% YoY increase from ~15,000 units/month at the end of 2023. Applied Materials has boosted its forecast for HBM packaging revenue from a prior view for 4X growth to 6X growth this year.””
