Head to head · verified Sep 5, 2026
B200 vs H200: do not skip the Hopper that boots
B200 is Blackwell: more HBM3e, more bandwidth, FP4 and a ~1 kW envelope. H200 is Hopper with a bigger tank and a catalog that actually has stock. I rent H200 when the job is memory-bound on Hopper software. I rent B200 when the row is live and my stack uses what Blackwell sells.
B200$3.75
VRAM180 / 141 GB
H200$3.00
The resultDefault: H200
The result
Default: H200
Pick: H200
Start on H200 unless FP4 and Blackwell memory are the point and the B200 row is in stock on-demand. A waitlist B200 is not cheaper than an H200 you can start tonight.
In brief
What to know before you pick
- B200: ~180 GB HBM3e, ~8 TB/s, FP4 and NVLink 5. H200: 141 GB HBM3e, 4.8 TB/s, Hopper.
- If your trainer never hits FP4 kernels, the B200 premium buys headroom, not speed.
- A B200 that is waitlist is later, not cheaper.
Same fields
The comparison
Catalog facts first. The pick above is editorial. Confirm the live on-demand row before you provision.
| Field | B200 | H200 |
|---|---|---|
| VRAM | 180 GBWins this row | 141 GB |
| Memory | HBM3eTie | HBM3eTie |
| Bandwidth | 8,000 GB/sWins this row | 4,800 GB/s |
| Catalog from-price | $3.75 | $3.00Wins this row |
| Cheapest indexed on | Packet.ai | Koyeb |
| On-demand rows in sample | 6 | 10 |
| In-stock rows | 2 | 6 |
| Architecture | blackwell | hopper |
Rent B200 if
Frontier-ish training or long-context serving that will use FP4 and NVLink 5, a live in-stock B200 row, and a facility and quota that survive 1 kW per GPU.
Skip it if Anyone who needed an H100 this week. Fine-tunes that fit 80 GB. Shops without a Blackwell software path.
Rent H200 if
Memory-bound Hopper work: 70B-class serving with a fat KV cache, full fine-tunes whose Adam states spill 80 GB, teams on CUDA 12.4-era stacks who need to boot today.
Skip it if Batches that already fit H100, budgets that only work at Vast.ai floor H100, or anyone who needed NVLink 5 and FP4.
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
H200
- 141 GB is the difference between one GPU and a sharded headache for many 70B stacks
- 4.8 TB/s HBM3e moves KV cache without pretending bandwidth is free
- Same Hopper software path as H100, fewer “new architecture” surprises
Stock decides the comparison
H200 is the Hopper SKU most catalogs can sell you this week. B200 is a thinner row with more waitlists and quotes. I do not compare their from-prices until both rows are on-demand and in stock.
A cheaper-looking H200 that is limited and a B200 that is waitlist is not a price comparison. It is a stock comparison.
Memory and bandwidth are the real delta
B200 roughly adds 40 GB over H200 and about 1.7x the memory bandwidth. That matters when KV cache, long context or optimizer states are the bottleneck. It does not matter if you already fit and are compute-bound.
H200 already fixed the H100 memory wall for many 70B jobs. If 141 GB holds, ask what the B200 premium is buying besides a newer badge.
FP4, NVLink 5 and a 1 kW bill
Blackwell’s pitch is lower-precision throughput and a faster fabric. Both are software and facility questions. Confirm the image, the kernels and the cooling before you assume the datasheet shows up in your step time.
Read TDP and interconnect on the listing. Then decide. The catalog table is the map, not the reservation.
The trap
How this comparison usually goes wrong
Comparing a waitlist B200 from-price with an in-stock H200. Treating peak FP4 as the throughput your trainer will see. Ignoring power and cooling on the Blackwell row.
What I would do
The move
Filter on-demand and in-stock. If B200 is waitlist, rent H200. If B200 is live, confirm GPU count, interconnect and whether your stack has an FP4 path. Then price 720 hours, not the slide.
Catalog listings
B200 and H200 in this catalog
On-demand hours first. Spot and waitlists are labeled. From-prices are catalog samples, not a reservation.
ProvidersRegionBilling typeInterconnect/GPU/hrOpen provider
ProvidersRegionBilling typeInterconnect/GPU/hrOpen provider
Full review
The two reviews
Same product, full page. Specs, listings, who it is for. Confirm the live hour there.
Full 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 reviewFull review
NVIDIA H200
Hopper with a bigger tank. Same SM family as H100, 141 GB of HBM3e at 4.8 TB/s. You rent it when memory is the bill, not because the badge says 200.
Read the H200 reviewFAQ
B200 vs H200 questions
Is B200 better than H200?+
On paper, yes: more memory, more bandwidth and FP4. In practice the better hour is the one that is in stock and fits your stack. H200 is often that hour.
B200 vs H200 for 70B inference?+
H200 if 141 GB holds the model and cache. B200 if you need the extra bytes or want FP4 serving and the row is live on-demand.
How much more does B200 cost than H200?+
Check the from-prices in the table above. The gap is only meaningful if both rows are in stock on-demand. Waitlist rows are labeled and not ranked as cheaper.
Should I wait for B200 instead of renting H200?+
Only if the job is blocked on memory H200 cannot hold. A waitlist is a delay, not a plan.
Keep comparing