Head to head · verified Sep 5, 2026
H100 vs B200: do not upgrade a badge
B200 is a different machine: FP4, NVLink 5, ~1 kW, a catalog that often says waitlist. H100 is the hour you can actually start. I do not launder a Blackwell desire through an H100 page, and I do not buy B200 because the number in the name is bigger.
H100$1.73
VRAM80 / 180 GB
B200$3.75
The resultDefault: H100
The result
Default: H100
Pick: H100
Rent H100 this week unless the run is large enough that Hopper memory and NVLink 4 are the bottleneck and the B200 row is in stock. A datasheet is not a reservation.
In brief
What to know before you pick
- B200 is Blackwell: FP4, NVLink 5, ~1 kW. H100 is the Hopper hour you can usually start this week.
- A waitlist from-price is not cheaper. It is later.
- If your trainer does not hit FP4 kernels, you bought a heater with a prettier name.
Same fields
The comparison
Catalog facts first. The pick above is editorial. Confirm the live on-demand row before you provision.
| Field | H100 | B200 |
|---|---|---|
| VRAM | 80 GB | 180 GBWins this row |
| Memory | HBM3Tie | HBM3eTie |
| Bandwidth | 3,350 GB/s | 8,000 GB/sWins this row |
| Catalog from-price | $1.73Wins this row | $3.75 |
| Cheapest indexed on | Vast.ai | Packet.ai |
| On-demand rows in sample | 15 | 6 |
| In-stock rows | 10 | 2 |
| Architecture | hopper | blackwell |
Rent H100 if
The job fits 80 GB or an 8× Hopper node. You need to boot tonight. Fine-tunes, 70B serving, teams without a CUDA 12.8+ FP4 path.
Skip it if Context windows that blow the KV cache, multi-node jobs that need a reserved fabric, or anyone shopping a 4090 workload with Hopper money.
Rent B200 if
Frontier-ish training or long-context serving that will actually use FP4, with cooling and quota that survive 1 kW hours, and a live in-stock row.
Skip it if Anyone who needed an H100 this week. Fine-tunes that fit 80 GB. Shops without a Blackwell software path.
H100
- The SKU every catalog actually stocks. You can compare an hour, not a waitlist
- FP8 Transformer Engine is the reason this generation still trains at a sane $/step
- MIG exists when you want to split an 80 GB card instead of overpaying for a second one
B200
- HBM3e at ~8 TB/s is a different memory regime than Hopper
- NVLink 5 at 1.8 TB/s is why 8-GPU jobs stop looking like a network problem
- Native FP4 is the inference lever H100 does not have
Stock beats the datasheet
H100 is still the hour most catalogs will sell you tonight. B200 is a different machine and a thinner row. I do not skip Hopper because the badge says Blackwell. I skip Hopper when the job is blocked on Hopper memory or NVLink 4 and the B200 listing is actually in stock.
Compare on-demand, in-stock, same GPU count. Mixing a waitlist B200 with a live H100 is how vendors win the spreadsheet.
FP4 is software, not a slogan
Blackwell’s pitch is precision and fabric, not a bigger H100. If your stack is still FP16 on CUDA 12.4, the extra silicon is idle. Confirm the image, the kernel path, and whether anyone on the team has shipped FP4 without a week of archaeology.
Power is part of the product. ~1 kW hours need cooling and a host that is not a random marketplace chassis. Read TDP on the listing. Then price 720 hours, not the FLOPs slide.
When I actually rent B200
Frontier-ish training, long-context serving that will use the memory, a live in-stock row, and a quota that survives the power bill. Otherwise I rent H100 and ship.
PCIe Hopper and SXM Blackwell are not the same family just because both say NVIDIA. Confirm the SKU. Then confirm the live rate.
The trap
How this comparison usually goes wrong
Comparing a waitlist B200 from-price to an in-stock H100. Mixing sparse peak FLOPs with the hour you will pay. Treating PCIe Hopper and SXM Blackwell as the same SKU family.
What I would do
The move
If the B200 listing is waitlist, stop. Rent H100. If it is in stock, confirm GPU count, fabric, and whether your stack speaks FP4. Then price 720 hours, not the slide.
Catalog listings
H100 and B200 in this catalog
On-demand hours first. Spot and waitlists are labeled. From-prices are catalog samples, not a reservation.
ProvidersRegionBilling typeInterconnect/GPU/hr
ProvidersRegionBilling typeInterconnect/GPU/hr
Full review
The two reviews
Same product, full page. Specs, listings, who it is for. Confirm the live hour there.
Full review
NVIDIA H100
The default datacenter hour. 80 GB of HBM3 is enough for most 70B work if you manage the KV cache. It is not a bargain because the FLOPs slide is pretty. It is a bargain when the job already fits.
Read the H100 reviewFull review
NVIDIA B200
Blackwell for people who will actually use FP4 and a 1.8 TB/s fabric. 180–192 GB of HBM3e, ~1 kW, often a waitlist. Not a default H100 replacement.
Read the B200 reviewFAQ
H100 vs B200 questions
Should I skip H100 and wait for B200?+
Only if the job is blocked on Hopper memory or NVLink 4 and you can wait. A waitlist is not cheaper. It is later.
Is B200 always faster?+
On paper, in the precisions it was built for. Your trainer has to hit those kernels. Otherwise you bought a heater with a prettier name.
H100 vs B200 for a 70B fine-tune this week?+
H100, unless 80 GB is already the wall and B200 is in stock on-demand. A datasheet is not a reservation.
Is the B200 from-price comparable to H100?+
Only if both rows are on-demand and in stock. Waitlist B200 next to live H100 is marketing. We refuse to mix them.
Keep comparing