NovaPodA friend asked if we finally had the newest NVIDIA rack, the one that showed up on Oracle Cloud this week. We did not have a scraped price. Public trackers through late October still had no on-demand Vera Rubin rate. I did not want the page to stay blank, so I put an estimate on the list and I am writing down how I got there.
Treat the dollar figures below as the ask I would start a conversation with, not as a receipt.
What actually shipped this week
On October 27, 2026, NVIDIA and Oracle said Vera Rubin NVL72 and the Vera CPU were generally available on Oracle Cloud. A few days earlier NVIDIA said racks were already operating at CoreWeave, Google Cloud, Microsoft Azure, Oracle, and Nebius. Google's own A5X instances were announced as coming the following month. AWS had no public price.
That is a real product, and it is still a hard one to book. The October rental notes had a B300 rentable in about three of four stock checks. Rubin was behind that: the same notes had no live on-demand Rubin offer to average.
The estimate I put on the list
The live B300 on this marketplace was about $6.95 an hour when I looked, 288 GB of HBM3e, fifteen offers. Rubin keeps 288 GB and moves the memory to HBM4 at up to 22 TB/s. I did not apply NVIDIA's tokens-per-watt claim as a price multiplier. I put one Rubin GPU at $12.60 to start and $16.80 in the middle, about 1.8 times the B300 floor, with three offers so the row looks as scarce as it is.
A Vera Rubin NVL72 slice is a different product: 72 GPUs on one NVLink domain. I priced that per GPU at $14.70 to start and $18.90 in the middle, two offers. Both rows say they are estimates. If a scraped offer shows up later, the scraped offer should replace them.
| Row | What it is | Starting ask | How sure I am |
|---|---|---|---|
| Vera Rubin | One Rubin GPU, 288 GB HBM4 | $12.60/hr | Estimate. No public on-demand index. |
| Vera Rubin NVL72 | Rack slice, priced per GPU | $14.70/hr | Estimate. Hardest row on the list. |
| B300 | Blackwell Ultra, 288 GB HBM3e | About $6.95/hr | Live offers already on the marketplace. |
What I like about it
The memory capacity did not move. The bandwidth did. If a job is already inside 288 GB and still waiting on memory, Rubin is the step that is supposed to matter. The NVL72 rack is the other reason: one NVLink domain instead of stitching eight-GPU nodes together.
Putting a number on the page is more useful than a pre-order badge that never says what the hour costs. I would rather argue with $12.60 than with a blank.
What I would tell you before you click it
This is not a vast.ai scrape. Three offers and two offers are placeholders so the row sorts with the scarce cards. They are not a promise that a machine is sitting idle.
If the model fits on a B300, paying toward $17 an hour to say you used Rubin is a story. Stop the machine when the run ends. The deposit floor is still $100, and the token still has to match the chain on the payment page.