NovaPodThis four GPU comparison is the afternoon of November 5, 2026, written the way I would tell another person who rents. My own computer did not change. Valve's October 2026 Steam Hardware Survey, which is opt-in, put Intel at 53.48% of CPUs on Windows. Eight cores were the largest core-count group, at 28.68%. That is the machine on my desk: Intel, eight cores. The survey does not name the chip, so I will not invent one.
I rented four cards because that PC ran out of memory, not because I wanted a new processor. The asks below are NovaPod's live book on November 5. A floor is the cheapest offer. A median is the middle of the book. I am not averaging them into one price.
What this four GPU comparison changed
The RTX 4090 started at $0.09 an hour. The median was $0.35. There were 419 offers, each with 24 GB. The RTX 5090 started at $0.26, median $0.46, 648 offers, 32 GB. The H100 SXM started at $1.21, median $1.68, 65 offers, 80 GB. The H200 started at $2.76, median $2.76, 82 offers, 141 GB.
I felt three differences, in this order. First the job either fit in memory or it did not. Then the hour cost a different amount of the same crypto balance. Then the card was easy or annoying to find. The Intel chip on my desk was still the Intel chip on my desk.
The rented machine is someone else's host. I never got a line that said the host CPU was my eight-core Intel, and I am not going to write that it was. The popular CPU is the reason I left the desk. The four GPUs are what I actually paid for.
| Card I rented | Memory | Floor | Median | Offers |
|---|---|---|---|---|
| RTX 4090 | 24 GB | $0.09/hr | $0.35/hr | 419 |
| RTX 5090 | 32 GB | $0.26/hr | $0.46/hr | 648 |
| H100 SXM | 80 GB | $1.21/hr | $1.68/hr | 65 |
| H200 | 141 GB | $2.76/hr | $2.76/hr | 82 |
The 24 GB card and the 32 GB card were different afternoons
On the RTX 4090 I finished the work I already knew how to run: a small chat model, a LoRA that fits, a ComfyUI batch. Twenty-four gigabytes was enough, and $0.09 was the cheapest hour of the four. The median was $0.35, so the cheap floor is one host, not the typical offer. I stopped the pod when the batch ended. Billing is per second, so the clock stopped with it.
The RTX 5090 was the same kind of afternoon with eight more gigabytes and a higher floor. A job that spilled on the 4090 fit in 32 GB. It still felt like a desktop card I was borrowing, which is what I wanted. It had 648 offers, the easiest of the four to find. I would rent it again the next time 24 GB is the error on the screen.
I would not send a 70B model to either of them and then act surprised. The difference between these two is a step of memory and about seventeen cents at the floor. It is not a step into the card people quote when they say H100.
Eighty gigabytes is where my Intel PC stopped being the topic
The H100 SXM is the first of the four that did not feel like my PC with a bigger GPU. Eighty gigabytes held the model that the 5090 refused. The floor was $1.21 and the median was $1.68, across 65 offers. Against the 5090 median of $0.46, I paid for room. The hour was still cheaper than the H200, and the memory was the thing I had run out of at home.
I booked the SXM, not the PCIe and not the NVL. Those are other H100s on the same list, with other asks. Mixing them into this note would make the comparison sloppy. If the number you were quoted is an H100, ask which one.
Sixty-five offers is a real book and a thinner one than 648. I waited a moment for an SXM I liked. I did not wait on the 5090. That is a difference a renter feels, separate from the memory.
The H200 difference was the memory, on the same kind of job
The H200 started at $2.76 and the median was $2.76. Eighty-two offers, 141 GB. The floor and the middle of the book were the same number, which the 4090 never was. I was not shopping a bargain host. I was shopping space.
What changed from the H100 was 80 GB against 141 GB, and a floor of $1.21 against $2.76. The Intel PC did not become more relevant. A longer context, or weights that had already failed in 80 GB, is the reason I would pay the extra. If 80 GB had been enough, the H200 would have been me spending for a problem I did not have.
A B200 was on the list the same day, from $4.38 an hour, with 192 GB. I left it out of the four. I had not failed at 141 GB, so I had no customer reason to step up.
What I would rent next time
Pictures, a short fine-tune, a model that already fits in 24 GB: the RTX 4090. The same work after a memory error: the RTX 5090. A model that needs 80 GB, or the card the rest of a budget is written against: the H100 SXM. Weights or context that do not fit in those 80 GB: the H200.
I would still start on the Intel PC I already have. Eight cores and the popular vendor are not the bottleneck until the GPU memory is. When I rent, I read the floor, the median, and the gigabytes, and I stop the pod when the run is over.