GPU & infrastructure / 2025–2026 SERIES
Blackwell Ultra: GB300 GPUs & Rack-Scale AI
The March 2025 milestone that put connected GPUs, networking, and serving software in the same conversation.
Milestone covered:
A single GPU specification tells only part of the AI performance story. Once an application uses several accelerators, the connections between them become part of the design.
Blackwell Ultra helped bring that system-level discussion into the foreground.
What NVIDIA announced
On March 18, 2025, NVIDIA introduced Blackwell Ultra. The announcement described GB300 NVL72 as a rack-scale system connecting 72 Blackwell Ultra GPUs and 36 Grace CPUs. It also emphasized networking and the Dynamo inference software announced alongside the platform. Partner products were scheduled for the second half of 2025. NVIDIA’s original announcement.
This is an important milestone near the start of the recent GPU cycle. It explains why newer announcements increasingly describe complete systems rather than isolated chips.
Why a business should care
Our architectural takeaway is to ask for the complete deployment configuration behind a performance claim. How many devices were used? Which model and software versions? What input lengths? What level of simultaneous traffic?
Without those details, a headline number is difficult to translate into a budget or a user experience.
Build a useful scorecard
For a customer-facing assistant, consider three measurements:
- Time until the first useful response appears.
- Time until the complete task is finished.
- Total cost per task that meets the quality requirement.
Run the comparison at normal traffic and at a realistic peak. Include failed requests in the cost calculation. If a test changes both the hardware and the serving software, report the result as a system improvement rather than attributing everything to the GPU.
Our practical recommendation
Ask providers for a small, reproducible workload trial before making a capacity commitment. Keep the test inputs, configuration, and scoring guide so another person can repeat it.
For many teams, the immediate benefit is a better way to evaluate infrastructure. You do not need to own an entire rack to insist on evidence about the service you plan to run.
Our Rubin overview follows how the same system-design theme continued in 2026.
Source note: launch facts are linked to the original announcements or documentation. Recommendations are Hydralogic’s analysis; this article does not report an independent product benchmark.
Explore the full collection ↗
