NovaPodAfter the Rubin estimate, the list was still missing the machines people actually ask for when a single B300 is not the shape of the job. A Grace rack is not a B300 with a longer name. An MI355X is not a B300 either, even though both say 288 GB.
Same rule as the Rubin note: if I mark it as an estimate, I do not want you to treat it as a scraped floor.
What was already real
B300 was on the marketplace at about $6.95 an hour, 288 GB, fifteen offers, demand marked high. B200 was about $4.38 to start and about $6.30 in the middle, fifty-one offers. H200 was about $2.76, with dozens of offers. Those rows came from the pricing feed. I left them alone.
The October market notes still had Blackwell and H200 failing about a quarter of stock checks. So the hard cards were already here. The hole was everything above a single GPU, and the AMD side of 192 GB to 288 GB.
The estimates I added
GB300 NVL72 is seventy-two B300s and Grace CPUs. Reserved quotes in late October were still talking about a Q1 2027 start, and there was no public per-GPU index I trusted. I put the slice at $8.40 to start and $11.20 in the middle, per GPU, four offers. That is above the live B300 on purpose.
GB200 NVL72 is the same idea on B200: $5.25 to start, $7.70 in the middle, six offers, against a live B200 floor near $4.38.
MI355X is the AMD CDNA 4 card with 288 GB of HBM3e. I put it at $5.95 to start and $7.70 in the middle, under B300 and over H200. MI350X is the lower sibling at $5.25. MI325X, 256 GB, is $3.15. MI300X, 192 GB, is $1.96, which is where the previous AMD flagship belongs once H200 is $2.76.
| Card | Memory | Starting ask | Source |
|---|---|---|---|
| GB300 NVL72 | 288 GB per GPU, 72 GPUs | $8.40/hr | Estimate, per GPU |
| B300 | 288 GB | About $6.95/hr | Live offers |
| MI355X | 288 GB | $5.95/hr | Estimate |
| GB200 NVL72 | 192 GB per GPU, 72 GPUs | $5.25/hr | Estimate, per GPU |
| B200 | 192 GB | About $4.38/hr | Live offers |
| MI325X | 256 GB | $3.15/hr | Estimate |
| H200 | 141 GB | About $2.76/hr | Live offers |
| MI300X | 192 GB | $1.96/hr | Estimate |
Where I would actually spend
One GPU, model fits in 288 GB, no need for a Grace CPU: the live B300. One GPU, 192 GB is enough: the live B200. A job that wants one NVLink domain across a rack: GB300 NVL72, and I would re-read the ask before launching because that row is an estimate.
I would try MI355X when I want 288 GB and I do not need CUDA-only libraries. The con is the software. A stack that assumes NVIDIA kernels does not become faster because the memory label matches.
The annoying parts
A rack price per GPU hides the real bill. Eight GPUs at $8.40 is not a cute hourly number. Read the quantity before you start.
Estimated offers are not inventory. Two, four, or six on those rows means scarce, not reserved for you. And none of this changes the $100 deposit or the per-second clock. Stop the pod.