NVIDIA Blackwell Ultra: The $40,000 GPU Powering the AI Revolution

NVIDIA has unveiled the Blackwell Ultra B300, the company's most powerful data center GPU, at a price point that underscores the enormous capital requirements of modern AI. At $40,000 per unit — and typically deployed in clusters of thousands — the B300 represents both the pinnacle of AI hardware and the growing divide between companies that can afford to train frontier AI models and those that cannot.

Performance Specifications

The numbers are staggering:

• AI Performance: 20 petaflops of FP4 compute (4x the H100)
• Memory: 288GB of HBM3e at 12 TB/s bandwidth
• Interconnect: NVLink 6.0 with 3.6 TB/s GPU-to-GPU bandwidth
• Power: 1,200W TDP (up from 700W for the H100)
• Process Node: TSMC 3nm

The Power Problem

Perhaps more notable than the GPU itself is the infrastructure it demands. A single rack of 8 B300 GPUs requires 10kW of power and advanced liquid cooling. NVIDIA's new DGX B300 system, housing 8 GPUs, costs $500,000 and requires dedicated cooling infrastructure that many data centers don't have.

"We're at a point where the limiting factor for AI isn't software or algorithms — it's power and cooling," said Jensen Huang at the launch event. "The companies that solve the infrastructure problem will win the AI race."

Market Impact

Despite the price, demand is overwhelming. NVIDIA reports a 12-month order backlog for the B300, with major cloud providers, AI labs, and sovereign AI initiatives competing for allocation. The company's data center revenue is projected to exceed $200 billion in fiscal 2027.

The concentration of AI compute in the hands of a few large companies has drawn scrutiny from regulators. The EU's proposed AI Infrastructure Act would require major cloud providers to offer subsidized compute access to startups and researchers, addressing concerns about AI becoming a "rich company's game."