Petaflare

B200 rental with crypto: NVIDIA Blackwell from $4.89 an hour

The fastest card we rent, paid only in crypto: 180 GB of HBM3e at 8 TB/s per GPU and Tensor Cores that run FP4. Whole cards, one to eight per machine.

Crypto-only payment, no card, no KYCUpdated 19 November 2026Prices checked 19 November 2026

B200 rental on Petaflare costs $4.89 per GPU-hour, $4.65 on the 120-hour Project block and $4.40 on the 480-hour Fleet block. The smallest purchase is 6 GPU-hours for $29.34. Each B200 has 180 GB of HBM3e at 8 TB/s, enough to hold a 70B model in BF16 on one card, and its Tensor Cores add FP4 to FP8 and BF16. You get 1, 2, 4 or 8 whole cards per machine with root SSH, paid in crypto.

B200 specs, per GPU

NVIDIA publishes the B200's figures for 8-card HGX and DGX systems. The numbers below divide those totals by eight: 1,440 GB of memory and 64 TB/s of bandwidth for the system, 72 PFLOPS of FP8 with sparsity.

Architecture
Blackwell
Memory
180 GB HBM3e
Memory bandwidth
8 TB/s
BF16 Tensor (dense)
2,250 TFLOPS
FP8 Tensor (dense)
4,500 TFLOPS
Board power
not published per GPU by NVIDIA
GPU link
NVLink 1.8 TB/s
Form factor
SXM (HGX board)

NVIDIA figures from the official B200 page. Tensor figures are without sparsity.

B200 price per hour across clouds

B200 prices shown on each provider's pricing page on 19 November 2026.

Sources: RunPod, Vast.ai public price feed (18 November median, cited with attribution), Lambda (plus sales tax), Massed Compute. Read on 19 November 2026. Every card and block is on cloud GPU pricing.
ProviderPrice typeConfigurationPer GPU-hour
PetaflareFixed, on demand1, 2, 4 or 8 cards$4.89 · $4.40 on Fleet
RunPod Community CloudThird-party hosts1 card$5.98
Vast.aiMarketplace, median of the day1 card$7.00
LambdaFixed, on demand8 cards (HGX) · 1 card$7.69 · $7.99
RunPod Secure CloudFixed, on demand1 card$7.99
Massed ComputeOn request8 cardsContact sales

Sources: RunPod, Vast.ai public price feed (18 November median, cited with attribution), Lambda (plus sales tax), Massed Compute. Read on 19 November 2026. Every card and block is on cloud GPU pricing.

Reading those B200 prices

On that day our $4.89 was the lowest B200 price in the table, 39% under RunPod Secure Cloud and Lambda's single-card rate, and below the cheapest marketplace and community offers too. That gap is wider than on older cards because B200 supply is still tight at most clouds. Prices for this card move faster than for the H100, so check the date on this page.

Paying for B200 hours in crypto

B200 hours start with a 6-hour block at $29.34, and the 480-hour block costs $2,112, or $4.40 an hour. For amounts this size, USDC or USDT on Ethereum keeps the price fixed in dollars and arrives in two to three minutes. Payment is in crypto only, never by card: see every coin and network.

What 180 GB on one card changes

At 2 bytes per parameter in BF16, 1 byte in FP8 and half a byte in FP4, the B200's memory opens jobs that needed two cards before:

  • A 70B model in BF16: about 140 GB of weights, with roughly 40 GB left for the cache. No other card we rent holds it alone; the H200 stops at 141 GB.
  • Llama 3.1 405B in FP4: about 203 GB, so two B200s with plenty of room for the cache, against four H200s in FP8.
  • DeepSeek-V3 or R1 (671B) in FP4: about 340 GB of weights, which fit on four B200s (720 GB). In FP8 they need all eight cards of a machine.
  • Fine-tuning a 70B model with LoRA in BF16: the base weights fit on one card, so you can tune without quantizing them.

For 8-card machines and what to check before distributed jobs, see GPU server rental.

FP4 and FP8 on Blackwell

Blackwell's Tensor Cores add 4-bit floating point (FP4, including NVIDIA's NVFP4 format) to the FP8 that Hopper introduced. Per GPU, dense throughput doubles at each step down: about 2.25 PFLOPS in BF16, 4.5 in FP8 and 9 in FP4.

FP4 is mainly an inference format: it halves the memory of FP8 weights and doubles peak math, at some cost in accuracy that depends on the model and on how it was quantized. Training still runs mostly in BF16 and FP8. If your serving stack does not support FP4 yet, the B200 still gives you twice the H100's FP8 throughput and 2.4 times its memory bandwidth.

B200 or H100: cost per unit of work

If a job keeps the card busy, what matters is the price of the work, not of the hour. Divide each card's hourly price by its throughput from NVIDIA's datasheets. Real speed-ups depend on the job; treat these as ceilings, not benchmarks.

Per-hour prices of 19 November 2026; dense FP8 Tensor figures from NVIDIA (H100 SXM and H200 SXM 1,979 TFLOPS; B200 72 PFLOPS sparse for 8 GPUs, so 4,500 dense per GPU). Lower is better in the last column.
CardPer hourDense FP8 TFLOPSMemory bandwidthDollars per hour per 1,000 FP8 TFLOPS
B200 180GB$4.894,5008 TB/s$1.09
H100 80GB SXM$2.391,9793.35 TB/s$1.21
H200 141GB$3.291,9794.8 TB/s$1.66

Per-hour prices of 19 November 2026; dense FP8 Tensor figures from NVIDIA (H100 SXM and H200 SXM 1,979 TFLOPS; B200 72 PFLOPS sparse for 8 GPUs, so 4,500 dense per GPU). Lower is better in the last column.

Software you need on Blackwell

The B200 needs a recent stack. CUDA 12.8 was the first release to support Blackwell, and PyTorch added Blackwell builds with its CUDA 12.8 wheels from version 2.7. Older containers built for CUDA 12.4 or earlier will not run kernels on the card.

When you connect, check the versions with nvidia-smi and python -c "import torch; print(torch.version.cuda)"; if you bring your own Docker image, start from a CUDA 12.8 or newer base. Serving frameworks such as vLLM and TensorRT-LLM ship Blackwell support in their recent releases; check the release notes of the version you pin. The fine-tuning guide lists the memory needs for each method.

B200 rental questions

How much does it cost to rent a B200?

$4.89 per GPU-hour, $4.65 on the 120-hour block and $4.40 on the 480-hour block. An 8-card machine costs $39.12 an hour at the list rate, or $35.20 on Fleet blocks.

How much memory does a B200 have?

180 GB of HBM3e per GPU in HGX and DGX systems, at 8 TB/s. Some sources quote 192 GB, the raw capacity of the memory stacks.

Is the B200 worth it over the H100?

For work that keeps the card busy, yes on paper: the B200 does about 2.3 times the H100's dense FP8 math for about twice the hourly price. For memory-bound serving of models that fit in 80 GB, the H100 is often enough.

Can I run my existing PyTorch code on a B200?

Yes, with PyTorch 2.7 or later built for CUDA 12.8 or later. Older builds do not include kernels for Blackwell.

Sources