One-tenth. That's the headline number Dell and NVIDIA dropped at Tech World today. The new Dell PowerEdge XE9812 server, designed for agentic AI inference, delivers each token at 90% lower cost than current Blackwell-based systems.
The machine is built on NVIDIA Vera Rubin NVL72—Jensen Huang's next-generation AI chip platform, a year in the making. Vera is the CPU, Rubin the GPU, and NVL72 the rack-scale integration. The PowerEdge XE9812 is the first production server to use Vera Rubin.
What makes Vera Rubin so powerful
Jensen Huang on the main stage:
"We've now arrived at the era of useful AI, which is the reason why demand is going parabolic, utterly parabolic."
To back up the claim, he shared key metrics:
- Vera CPU: 1.2 TB/s memory bandwidth, 50% faster on agentic tasks than x86 processors
- Database queries: 3x speed on Starburst and DuckDB
- PowerEdge XE9880L / 9885L / 9882L series: Using HGX Rubin NVL8, packing 144 GPUs per rack, 5.5x performance over HGX B200, all 100% direct liquid cooled
The 144-GPU-per-rack density doubles the previous NVL72's 72 GPUs. Combined with liquid cooling, rack-level compute is over 5x that of the Blackwell era.
The 90% cost reduction comes partly from chip efficiency and partly from the tight coupling of Vera CPU and Rubin GPU—the CPU no longer bottlenecks the GPU, allowing token throughput to hit design limits.
Michael Dell's $3 trillion roadmap to 2030
Michael Dell laid out the financial picture:
- Global AI infrastructure spending will reach $3–$4 trillion by 2030
- Token consumption will grow 3,400% in the same period
He added:
"The rate of change has gone parabolic, and it's not slowing down."
Both he and Huang used "parabolic"—the curve is accelerating upward with no plateau in sight. For context, global data center capex in 2024 was around $300 billion. Ten times that is $3 trillion.
Dell's customer list already validates the forecast:
- Eli Lilly: 15 years on Dell infrastructure, now training AI models; LillyPod supercomputer has nearly 2 TB/s read bandwidth with 1,000+ GPUs
- Samsung: Using Dell for chip design R&D
- Honeywell: Running industrial AI and digital twins, targeting Middle East oil and gas
- Hudson River Trading: Algorithmic trading
These are not PPT customers. They are running production workloads on Dell on-premises solutions, and the PowerEdge XE9812 is an upgrade path—not a new customer acquisition.
An overlooked detail: liquid cooling is now standard
All new PowerEdge XE9880L, XE9885L, and XE9882L models are 100% direct liquid cooled. This signals the end of the air-cooled era. With 144 GPUs per rack and power density approaching 100 kW, air cooling is insufficient. Enterprises deploying Vera Rubin must retrofit their data centers for liquid cooling—a significant capital expense Dell did not elaborate on.
Dell's story today has two layers:
- Visible: Inference costs cut to one-tenth, token throughput up 5x
- Hidden: To use this hardware, data centers need a costly liquid cooling upgrade
The first is for sales; the second is what CIOs will be calculating tonight.
Sources: NVIDIA CEO Jensen Huang at Dell Technologies World: Demand Is Going Parabolic, Utterly Parabolic (NVIDIA Blog); CocoLoop; Dell Technologies World 2026: Enterprise AI Announcements This Week (Dell Blog)