NVIDIA's monster Blackwell GPU takes the stage

Jensen Huang walked onto the arena stage and pulled a new chip out of his jacket pocket. Held side by side with the previous Hopper architecture, the new Blackwell processor made its predecessor look tiny.

At a packed SAP Center in Silicon Valley, Huang declared that general-purpose computing has officially run out of steam. To handle trillion-parameter models, Nvidia is betting everything on its new Blackwell platform.

A single rack with exaflop power

The Blackwell architecture brings serious upgrades across the board. Per chip, it delivers 2.5 times the FP8 training performance of Hopper and a massive 5 times the FP4 inference performance.

To keep all those cores fed, Nvidia introduced a fifth-generation NVLink that doubles the speed of Hopper and can link up to 576 GPUs. The company even designed a custom NVLink Switch chip pushing 1.8 terabytes per second.

The real centerpiece is the GB200 Grace Blackwell Superchip, combining two Blackwell GPUs and a Grace CPU over a 900 GB/s interconnect.

Scale that up into the liquid-cooled GB200 NVL72 rack, and you get a 3,000-pound machine built from 600,000 parts. It delivers 1.4 exaflops of FP4 inference performance in a single cabinet. Beyond that, the new DGX SuperPOD system scales compute up to 11.5 exaflops.

Assembling teams of AI agents

Nvidia did not stop at hardware. Huang introduced Nvidia NIM, which stands for inference microservices, as a completely new model delivery system.

Huang argued that future software will rarely be written from scratch in Python. Instead, developers will orchestrate specialized teams of AI models using standardized APIs.

NIM packages enterprise models with acceleration libraries so companies can deploy them out of the box or fine-tune custom copilots with internal data.

Nvidia also revealed that its computational lithography platform, cuLitho, is officially entering production with TSMC and Synopsys, accelerating chip manufacturing workloads by up to 60 times.

Why it matters

For builders and enterprise teams, a 5x leap in inference throughput fundamentally alters the economics of running generative AI. Trillion-parameter multimodal models that were once impossibly expensive to query are suddenly within practical reach.

Across the semiconductor landscape, Nvidia is widening its moat by turning the entire data center into a single product. Rivals targeting standalone accelerator chips now face a competitor selling liquid-cooled, switch-integrated mega-systems bundled with an end-to-end software delivery layer.

We think this full-stack dominance is brilliant for keeping hyperscalers locked in, but it creates operational bottlenecks. Deploying a 3,000-pound liquid-cooled rack demands massive facility retrofits that many legacy facilities simply cannot handle right now.