Google Cloud spent October 21 putting a calendar date on Nvidia’s next big box. The A5X instance, built around the Vera Rubin NVL72, is supposed to hit general availability next month. The capacity target written next to it is 960,000 GPUs.
What “generally available” will mean
In cloud language it means a customer can order the instance without being in a private preview. It does not mean every region has slack, or that a startup without a committed-spend contract gets the same yes as a lab. Rubin is Nvidia’s follow-on to the Blackwell generation. NVL72 is the rack-scale system, not a single card.
Nine hundred and sixty thousand is a planning number from Google, not a census of chips already plugged in. Read it as intent. The interesting comparison is Meta talking about a path to a million GPUs, and Amazon writing eleven-digit checks for European halls. The buyers are publishing targets because the supply argument is now public.
If you are the person who has to book training
Ask which region, which quota, and whether your current reservation converts. A new instance family often launches beside the old one, with the old one staying cheaper and easier to get for months.
And do not rewrite a training plan around a chip you have not profiled. Rubin’s value shows up in tokens per dollar on your model, not in the name on the slide.