B200 vs H200 vs H100 for LLM training in 2026: which GPU to rent
On the Lium Secure tier an 8x H100 node costs $2.90 per GPU-hour and 8x H200 $4.06; on Spot, 8x H100 costs $1.49 and 8x B200 $5.50. On Spot, a B200 only saves money if your job runs at least 3.7x faster than on H100. Prices checked 9 November 2026.
Lium team5 min read
gpu-pricing
training
b200
h200
h100
comparison
Short answer: rent 8x H100 for most training jobs, 8x H200 when the model does not fit in 640 GB of GPU memory, and B200 only when your training stack runs at least 2.1x faster on it than on H100 (Secure tier; on Spot the bar is 3.7x).
On Lium's Secure tier, a whole 8x H100 node costs $2.90 per GPU-hour, 8x H200 costs $4.06 and B200 starts at $6.10. On Spot, 8x H100 costs $1.49 and 8x B200 $5.50. (Prices from the public nodes feed, generated 9 November 2026 at 15:05 UTC.) What matters for training is not the hourly price but the price of finishing the job, so this post compares the three per useful hour.
Lium prices on 9 November 2026
Cheapest whole 8-GPU node that was free to rent, per GPU-hour:
¹ No whole Secure 8x B200 node was free at that moment; single Secure B200 GPUs started at $6.10.
Read from the public nodes feed, generated 9 November 2026 at 15:05 UTC. Secure is a dedicated machine, and you are credited if the provider interrupts your rental. Spot can be reclaimed, so save checkpoints off the pod. Lium is a marketplace where providers set the price, so check Pricing and the GPU list for today's numbers, and Stats for how many GPUs are listed and rented.
For reference, Lambda's pricing page lists its 8x instances at $4.59 per GPU-hour for H100 and $7.69 for B200, effective 16 November 2026 (read 9 November 2026, 15:06 UTC). It lists no H200 instance.
What you get per GPU
From NVIDIA's spec pages:
Sources: H100 and H200 SXM columns; B200 is the DGX B200 system total (1,440 GB, 64 TB/s, 72 PFLOPS FP8) divided by its 8 GPUs.
Two things follow:
- H200 has the same compute as H100. It adds 76% more memory and 43% more bandwidth. A job that already fits on H100 and is limited by compute will not finish much sooner on H200.
- B200 has about 2.27x the peak FP8 compute of H100. NVIDIA states that DGX B200 delivers "3X the training performance" of the previous generation, measured on a 32,768-GPU cluster. That is a vendor claim at a scale you are unlikely to rent. Your speedup depends on your model, your precision (FP8 needs a stack that supports it) and your code.
Price per useful hour
The useful number is the break-even speedup: how much faster a GPU must finish your job to cost the same as H100 on the same tier. It is the price of the faster GPU divided by the price of H100.
² Single Secure B200 GPUs; no whole Secure 8x B200 node was free at the time of the snapshot.
If your job runs faster than the break-even speedup, the bigger GPU is cheaper overall. If it runs slower, H100 is cheaper even though each hour costs less on the bigger GPU.
Worked example: a 100-hour training job
Assume a job that takes 100 hours on one 8x H100 node and fits in its 640 GB of memory.
On the Secure tier:
- 8x H100: 100 h x $23.20 per node-hour = $2,320.00.
- 8x H200: same compute, so about the same 100 hours: 100 h x $32.48 = $3,248.00. You pay 40% more for memory you do not need.
- 8x B200: no whole Secure 8x B200 node was free at the time of the snapshot, so there is no Secure node price to quote. Single Secure B200 GPUs started at $6.10, which puts the break-even at about 2.1x.
On the Spot tier, where a node can be reclaimed:
- 8x H100: 100 h x $11.92 = $1,192.00.
- 8x B200: $44.00 per node-hour. At 1.5x faster it costs $2,933.33, at 2.27x $1,938.33. The break-even is 3.7x.
For reference, 8x H100 on Lambda at $4.59: 100 h x $36.72 = $3,672.00.
Storage and data transfer are not included. On Lium a pod is billed per second while it exists, and network volumes are billed for storage. See Billing.
When H200 is worth it
H200 pays off when memory, not compute, is the limit. Say a model and its optimizer state need about 1 TB of GPU memory. On H100 that means two 8-GPU nodes, 16 x $2.90 = $46.40 an hour, plus the slower links between the two machines. One 8x H200 node holds 1,128 GB for $32.48 an hour. The same is true if H100 forces you into offloading or very small batches that slow every step.
How to choose
- Measure first. Rent the cheapest GPU that fits, run a few hundred steps, and note the time per step. Per-second billing means a 20-minute test on 8x H100 costs about $7.73.
- Run the same test on the bigger GPU. Divide the two step times to get your real speedup.
- Compare with the break-even table. Pick the GPU where your speedup is above the break-even, or stay on H100.
- Use Secure for long runs. A multi-day job should not be reclaimed halfway. Use Spot for tests and for jobs that checkpoint often.
To start, open the GPU list, filter by H100, H200 or B200, and rent a whole node. Pricing shows today's prices for each model. No quota request or contract is needed. For a price comparison across clouds, see Cheapest H100 and H200 cloud in 2026.
Run it on Lium
Billed per second, only while your pod runs.
Get new posts by e-mail: Subscribe to the Lium newsletter
