If you have been following the AI industry at all in 2025, you have probably noticed that almost every major announcement — whether from OpenAI, Google, Meta or any of the cloud providers — eventually traces back to the same company: NVIDIA. Its latest chip, the Blackwell Ultra (B300), has just reinforced that position in a way that competitors will struggle to respond to quickly.
Why GPU Chips Matter So Much for AI
Training and running large AI models requires an enormous amount of parallel mathematical computation. Graphics Processing Units (GPUs), originally designed to render video game graphics, turned out to be exceptionally well-suited to this kind of parallelism. NVIDIA recognised this early and built an entire software ecosystem (called CUDA) around its GPUs. NVIDIA now controls roughly 80% of the market for AI training chips.
What Is Blackwell Ultra?
The Blackwell Ultra (B300) is NVIDIA’s latest generation AI accelerator, beginning to ship to hyperscale customers in Q2 2025. The headline numbers are extraordinary:
- Up to 20 petaflops of AI performance (compared to around 4 petaflops for the H100)
- 192 GB of HBM3e memory — significantly more than the H100’s 80 GB
- 8 TB/s memory bandwidth — faster data movement between memory and the processor
- New FP4 precision support, which enables running models at lower precision for inference tasks
“Blackwell Ultra is not an incremental improvement over Hopper — it’s a generational leap. The economics of running frontier AI models change significantly with this hardware.” — Jensen Huang, NVIDIA CEO, GTC 2025
What This Means for AI Companies
The practical impact of Blackwell Ultra is that frontier AI models become cheaper to run. This matters because inference cost (the cost of generating each response from an AI model) is one of the biggest constraints on how widely AI can be deployed in commercial products.
What About the Competition?
NVIDIA’s dominance has attracted serious challengers: AMD’s MI300X is a genuinely competitive chip for inference workloads. Google’s TPU v5 is optimised for Google’s own models. Amazon has Trainium and Inferentia. None of them have broken NVIDIA’s hold on the market.
The Geopolitical Dimension
NVIDIA chips are at the centre of one of the most significant geopolitical stories of the decade: US export controls on advanced semiconductors to China. Blackwell Ultra is subject to the same restrictions as H100. Huawei’s Ascend 910C is the most advanced Chinese AI chip currently in production — analysts generally place it roughly 18-24 months behind NVIDIA’s current generation.