Technology & AI

NVIDIA Vera Rubin Is Shipping: What AI Buyers Should Check Before Upgrading

Vera Rubin is shipping. Before you upgrade, find out where you can get capacity and whether it improves the work you actually do.

Illustration of AI server racks beside the headline NVIDIA Vera Rubin Is Shipping, with The Spokesdude Network logo.

After Jensen Huang stepped down, AI buyers had to ask whether NVIDIA could deliver on its roadmap. The company’s 17 November earnings release gives us something more useful than a promise: NVIDIA says its new Vera Rubin systems have shipped and are running at cloud partners. For South African businesses, the next question is more practical: can you get the capacity you need, where you need it, at a price that makes the work viable?

What NVIDIA confirmed

NVIDIA says Vera Rubin NVL72 systems are running at partners including Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure. Google Cloud and Oracle separately announced general availability for their Rubin offerings. Those announcements give buyers more options to investigate. They do not tell you whether the service is available in your preferred region, whether your account can reserve capacity or what your workload will cost.

NVIDIA reported US$115.8 billion in revenue for the quarter ended 25 October, before the CEO transition. Its US$130 billion guidance, plus or minus 3%, is for the following quarter, which includes an extra week. The first figure shows the scale of the business Colette Kress inherited; the second is management’s forecast, not a result. NVIDIA earnings release

What changed at the board

On the earnings call, board chair Harvey Jones said NVIDIA had hired Heidrick & Struggles to lead the search for a permanent CEO and did not intend to reinstate Huang. Kress remains interim CEO. The board also established a standing committee for national security and export controls. That committee changes who oversees those risks. It does not resolve the separate investigation.

What to check before you buy

Get a capacity commitment in writing. Ask your provider which region will run the service, when capacity can be reserved and what happens if it isn’t available at launch. If you buy a managed AI service rather than rent GPUs directly, ask about the service limits and commitments that apply to your account.

Run the same job on your current setup and the new option. Use a representative workload and the same quality threshold. For example, if one setup costs R120 to process 1,000 support tickets and all 1,000 pass review, that’s 12 cents per usable result. If another costs R98 but only 700 pass, it costs 14 cents per usable result. Compare turnaround time as well as cost.

Check what it would take to switch. Keep your evaluation set, workflow instructions and results exportable. Test whether you can rerun a small batch elsewhere before you commit to a contract you’ll struggle to unwind.

Rubin gives buyers a reason to ask for a trial. Upgrade when that trial shows a measurable improvement on your own work and the provider can commit to delivering it. Until then, keep the setup that works.

*Related reading: NVIDIA after Jensen Huang.*

Written by The Spokesdude NetworkDiscuss your reporting challenge