one year on
Nvidia unveils Blackwell GPU platform, promising up to 30x inference leap over H100
With a 208-billion-transistor chip and new software subscription, Nvidia aims to lock in AI customers as the era of trillion-parameter models begins.
SAN JOSE — Nvidia chief executive Jensen Huang took the stage at the SAP Center on Monday to unveil Blackwell, a new GPU architecture the company is positioning as the engine for a coming wave of trillion-parameter AI models. The first chip, the GB200 Grace Blackwell Superchip, pairs two B200 GPUs with an Arm-based Grace CPU and is expected to ship later this year.
Nvidia says the GB200 NVL72 provides up to a 30x performance increase compared with the same number of H100 Tensor Core GPUs for LLM inference workloads, while reducing cost and energy consumption by up to 25x. “Hopper is fantastic, but we need bigger GPUs,” he told the packed arena. The GPU packs 208 billion transistors and is manufactured as a single chip from two dies using TSMC’s 4NP process.
Alongside the hardware, Nvidia introduced NIM (Nvidia Inference Microservice), a software layer designed to let customers run AI models on any Nvidia GPU with minimal reconfiguration. NIM will be part of Nvidia’s enterprise software subscription, priced at $4,500 per GPU per year. The company is positioning itself as a platform provider akin to Microsoft or Apple, not just a chip vendor.
Major cloud providers including AWS, Google Cloud, Microsoft Azure, and Oracle plan to offer Blackwell-based instances. AWS is co-developing Project Ceiba, a cluster of 20,000 GB200 chips for Nvidia’s own AI research. Nvidia said the NVL72 system combining 72 Blackwell GPUs can handle a 27-trillion-parameter model — far beyond today’s largest known models. The announcement comes during GTC, which runs through March 21.
The record
Described Blackwell as offering massive performance leaps that will accelerate delivery of leading-edge models.
Stated there is 'nothing better than NVIDIA hardware for AI' currently.
Said Meta looks forward to using Blackwell to train open-source Llama models and build next-gen AI products.
Committed to bringing GB200 to Microsoft datacenters globally, building on a history of optimizing Nvidia GPUs for Azure.
Highlighted a 13-year collaboration and noted AWS will offer Blackwell instances, co-developing Project Ceiba with Nvidia.
One year later — open only if you can handle spoilers
Blackwell became Nvidia's fastest-ramping product in history, driving its market cap above $3 trillion by mid-2024. The 'the more you buy, the more you save' catchphrase became a meme among AI engineers, while Nvidia's software push with NIM gradually expanded its platform moat. Yet tight supply persisted for months, leaving some customers frustrated.