GPU prices · updated 2026-11-21 15:00 UTC

B200 vs H200 vs H100

Lowest on-demand list prices on 2026-11-21: B200 $6.13, H200 $3.60, H100 SXM $3.56 per GPU-hour. We measured all three on the same model, on the same day and under the same latency target. At these prices, a million output tokens costs $0.217 on the B200, $0.249 on the H200 and $0.276 on the H100 SXM.

Price per hour and cost per million tokens

On-demand list prices from provider APIs, normalized per GPU-hour, next to output tokens per second we measured within the latency target (p99 TTFT ≤ 2 s, p99 TPOT ≤ 50 ms) on Qwen3-8B BF16, vLLM 0.33.0, one GPU each. Cost per million output tokens = price per GPU-hour ÷ (tokens/s × 3,600) × 1,000,000.

GPUMemoryBandwidthLowest
$/GPU-hour
Median
$/GPU-hour
OffersOutput tok/s
(measured)
$/M output tokens
at lowest price
$/M output tokens
at median price
B200180 GB8.0 TB/s$6.13$7.7967,834$0.217$0.276
H200141 GB4.8 TB/s$3.60$4.8074,014$0.249$0.332
H100 SXM80 GB3.35 TB/s$3.56$4.8483,580$0.276$0.376

Which offers: H100 SXM, the 8 on-demand offers listed as SXM, the version we measured (all 20 H100 offers, which also include PCIe cards and listings that don't state the form: lowest $1.99, median $4.19, as on the H100 page); B200, all on-demand offers as on its price page, whatever form they list (3 SXM, 3 unspecified); H200, all on-demand offers as on its price page, whatever form they list (1 SXM, 6 unspecified).

What the numbers say

Speed. On Qwen3-8B BF16, the B200 delivered 7,834 output tokens per second within the latency target, 2.2× the H100 SXM's 3,580. The H200 delivered 4,014, 12% more than the H100 SXM, although it has 141 GB of memory against 80 GB: an 8B model fits on any of these cards with room to spare, so at this size the extra memory doesn't buy much.

Per token at 2026-11-21 prices. At the lowest list prices, the B200 costs 1.7× as much per hour as the H100 SXM and delivered 2.2× the tokens, so a million output tokens costs $0.217 on the B200, 21% less than on the H100 SXM ($0.276). Against the H100 SXM, the H200's lowest price is 1% higher and its throughput 12% higher, so a million output tokens costs $0.249, 10% less. At median prices: B200 $0.276, H200 $0.332, H100 SXM $0.376.

At the prices of the nodes we measured, set by their hosts on Lium on November 10, 2026 (B200 $8.88, H200 $5.30, H100 SXM $2.50 per GPU-hour), the same throughput cost $0.315, $0.367 and $0.194 per million output tokens. Which card is cheapest per token depends on the price you pay.

When the H200's and B200's memory matters

The H200 has 141 GB and the B200 180 GB of GPU memory, against 80 GB on the H100. It matters when a model's weights and KV cache don't fit comfortably in 80 GB: larger models, long contexts, and many concurrent requests, each of which keeps its KV cache in GPU memory. On a larger model, Qwen3-30B-A3B in FP8 (measured on RunPod in October 2026), the H200 served 18% more requests than an H100 SXM at a 50 ms per-token target and 29% more at 30 ms: H100 vs H200 for LLM inference.

Limits

  • One node per GPU type, rented on Lium's Secure tier on November 10, 2026 and chosen by the same rule written down before renting (most reliable listing, not lowest price). The H100 and B200 nodes came from the same host.
  • Quick mode: two-minute windows, two repeats. It can't separate differences of a few percent.
  • One small model (Qwen3-8B BF16), one chat-like workload (median 1,024 input and 256 output tokens) and one engine (vLLM 0.33.0, default settings). Larger models, other precisions and other engines may give different ratios.
  • Prices are on-demand list prices collected 2026-11-21 15:00 UTC and change daily; they exclude storage, egress, reservations and discounts. Cost per token assumes the card you rent performs like the one we measured: check yours.

Questions

Is B200 worth it over H100?

On an 8B model, at 2026-11-21 prices: the B200 delivered 2.2× the H100 SXM's output tokens per second (7,834 against 3,580; Qwen3-8B BF16 on vLLM 0.33.0, p99 TTFT ≤ 2 s, p99 TPOT ≤ 50 ms), and its lowest on-demand price was 1.7× the H100 SXM's ($6.13 against $3.56 per GPU-hour). A million output tokens cost $0.217 on the B200 and $0.276 on the H100 SXM at the lowest prices, $0.276 and $0.376 at the median. This is one node of each on one small model.

H200 vs B200: which is cheaper per token?

At the lowest on-demand prices on 2026-11-21, the B200 is cheaper per token: $0.217 per million output tokens against $0.249 on the H200. At median prices, the B200 is cheaper per token: $0.276 per million output tokens against $0.332 on the H200. On Qwen3-8B BF16 the B200 delivered 2.0× the H200's output tokens per second (7,834 against 4,014) at 1.7× its lowest hourly price ($6.13 against $3.60).

How much does a B200 cost per hour?

As of 2026-11-21, on-demand NVIDIA B200 list prices ranged from $6.13 to $8.50 per GPU-hour across 6 offers, median $7.79. Prices exclude storage, egress and committed-use discounts.

Sources

Measurements: five GPU marketplace nodes measured (write-up) and the per-node reports with raw JSON: B200 · H200 · H100 SXM. Method: how we measure.

Price per hour by GPU: B200 · H200 · H100 · all GPUs and providers. Related: H100 vs H200 for LLM inference · H100 SXM vs PCIe vs NVL.

Prices are list prices at collection time from the providers on our GPU price page; they exclude storage, egress, reservations and discounts. Data.