AI

Nvidia's GB300 GPU platform arriving soon

Will ship in volume before end of this year.

Paul Mah

Paul Mah

— 1 min read
Nvidia's GB300 GPU platform arriving soon
Photo Credit: ServeTheHome. B200 GPU Server from Asus

Nvidia last month unveiled their GB300 platform. For organisations looking to deploy the latest GPUs, here's what you'll get - and when it'll arrive.

Last week I wrote about Nvidia's next-gen AI offering.

The GB300 in a nutshell

If there are two key things to know about the GB300 platform, it would be:

  • It has 50% more FLOPS compared to the B200. And power is up just slightly, up to 1.4kW, making it preferably for those who can deploy it in their data centres.
  • In addition, GB300 offers greater customisation, which means hyperscalers and server makers can customise the main board, cooling, and others, for custom solutions.

Unsurprisingly, orders from hyperscalers have quickly shifted to the GB300, according to reports.

The rest of us

Hyperscalers and tech giants like OpenAI and Meta always get first dibs for the latest GPUs. It's easy to see why: They have earmarked 100's of billions on AI data centres.

As AI models continue to improve and organisations roll out AI for their use, academic institutions, second-tier cloud players, and enterprises are now looking to roll out their own AI infrastructure.

In general, these users access GPUs by purchasing GPU servers from Nvidia partners like Asus. They then deploy them within their data centres.

When will it arrive?

According to Asus, its finalised (L-11 ready) GB300 GPU server solution is expected to ship in 2H of 2025.

  • Small-scale shipments initially.
  • Fully integrated and ready for use.
  • Significant ramp-up expected in Q4 2025.

While it's not stated currently, I expect its B300 GPU server solution will be liquid-cooled (direct-to-chip) due to its power consumption.

Of course, non-hyperscalers can afford to deploy fewer GPU servers per rack, so an air-cooled option might increase its appeal due to greater deployment flexibility.

The cost? If you're asking, chances are you are not ready. But expect it to be more than US$300,000 it cost for the existing H200 GPU server.

Is your organisation deploying GPU servers for internal AI pilots?

Read more