Petaflare

H200 rental with crypto: 141 GB per GPU from $3.29 an hour

The Hopper chip of the H100 with 76% more memory and 43% more bandwidth, rented by the hour and paid only in crypto. Rent it when 80 GB is what stops your model, not compute.

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

H200 rental on Petaflare costs $3.29 per GPU-hour, $3.13 on the 160-hour Project block and $2.96 on the 640-hour Fleet block; the smallest purchase is 8 GPU-hours for $26.32. Each card has 141 GB of HBM3e at 4.8 TB/s, and a machine holds 1, 2, 4 or 8 of them with root SSH. That memory is the reason to pick it: a 70B model in FP8 with a 128,000-token context fits on one H200 but not on one H100.

H200 specs

Same Hopper compute as the H100 SXM, so FP8 and BF16 throughput are the same. The difference is the memory: 141 GB of HBM3e instead of 80 GB, at 4.8 TB/s instead of 3.35 TB/s.

Architecture
Hopper
Memory
141 GB HBM3e
Memory bandwidth
4.8 TB/s
BF16 Tensor (dense)
989 TFLOPS (SXM)
FP8 Tensor (dense)
1,979 TFLOPS (SXM)
Board power
up to 700 W (SXM) · 600 W (NVL)
GPU link
NVLink 900 GB/s (SXM)
Form factor
SXM or NVL (PCIe), shown before you connect

NVIDIA figures from the official H200 page. Tensor figures are without sparsity.

H200 price per hour across clouds

One H200, as shown on each provider's pricing page on 19 November 2026. Lambda listed no H200 instance that day.

Sources: Massed Compute, RunPod, Vast.ai public price feed (18 November median, cited with attribution), Silicon Data (H200 reading). Read on 19 November 2026. All our blocks and prices: what a GPU-hour costs.
ProviderPrice typeH200 versionPer GPU-hour
PetaflareFixed, on demand141 GB, SXM or NVL$3.29 · $2.96 on Fleet
Massed ComputeFixed, on demandNVL 141 GB$3.62
RunPod Secure CloudFixed, on demandH200$5.29
RunPod Community CloudThird-party hostsH200$3.59
Vast.aiMarketplace, median of the dayH200 · H200 NVL$3.50 · $3.17
Silicon Data indexComposite of neo-clouds and marketplacesH200$3.76

Sources: Massed Compute, RunPod, Vast.ai public price feed (18 November median, cited with attribution), Silicon Data (H200 reading). Read on 19 November 2026. All our blocks and prices: what a GPU-hour costs.

Reading those H200 prices

Our $3.29 is below every on-demand H200 price in the table and below the marketplace median for the SXM card. Only Vast.ai's H200 NVL median, $3.17, is lower, and the Fleet block brings ours to $2.96. Marketplace cards belong to independent hosts, with storage and bandwidth billed on top.

Paying for H200 hours in crypto

An 8-hour H200 block costs $26.32, and 640 hours cost $1,894.40, or $2.96 an hour. Bitcoin payments are credited after 2 confirmations, about 20 minutes, and USDT on Tron is faster. We take crypto only, without a card or an identity check; the details are on GPU server with Bitcoin.

Why 141 GB matters: the cache sum

When you serve an LLM, the weights are only part of the memory. Every token in every conversation also keeps its keys and values in the KV cache, and with long contexts the cache can outgrow the weights.

Take Llama 3.1 70B: 80 layers, 8 key-value heads of 128 dimensions each. One token needs 2 (key and value) × 80 × 8 × 128 values, or 163,840 values. At 2 bytes per value that is about 0.33 MB per token, so one 128,000-token conversation needs about 42 GB of cache, or about 21 GB if the cache is stored in FP8.

  • On one H100 (80 GB): the FP8 weights take about 70 GB, leaving about 10 GB, enough for roughly 60,000 tokens of FP8 cache shared by all users.
  • On one H200 (141 GB): the same weights leave about 70 GB, enough for three full 128,000-token conversations with an FP8 cache.

These are upper bounds before the serving framework's own overhead. The LLM inference guide has the same sums for other model sizes.

One H200 or two H100s

If a model needs more than 80 GB, the alternative to one H200 is splitting it across two H100s. Splitting works, but it costs more per hour and adds traffic between the cards.

Hourly cost is the GPU-hour rate times the number of cards. One H200 costs 31% less per hour than two H100s; four H200s cost 31% less than eight H100s.
SetupGPU memoryPer hour at our list pricePer hour on Fleet blocks
1 × H200141 GB$3.29$2.96
2 × H100160 GB$4.78$4.30
4 × H200564 GB$13.16$11.84
8 × H100640 GB$19.12$17.20

Hourly cost is the GPU-hour rate times the number of cards. One H200 costs 31% less per hour than two H100s; four H200s cost 31% less than eight H100s.

Large models on 4 and 8 H200s

  • Llama 3.1 405B in FP8: about 405 GB of weights. Four H200s hold 564 GB, which leaves around 160 GB for the cache; with H100s you would need all eight cards of a machine.
  • DeepSeek-V3 and R1 (671B parameters, FP8 weights): about 700 GB including extra layers. They fit on eight H200s (1,128 GB) and not on eight H100s (640 GB).
  • 70B in BF16 with long contexts: 140 GB of weights. Two H200s leave around 140 GB for the cache, where two H100s leave about 20 GB.

Machines with 4 and 8 cards, and what to check before distributed jobs, are on the GPU server page.

SXM or NVL

NVIDIA makes the H200 as an SXM module for HGX boards (up to 700 W, NVLink at 900 GB/s between all cards on the board) and as the H200 NVL, a PCIe card limited to 600 W that links to its neighbours with an NVLink bridge. Both have 141 GB at 4.8 TB/s; the NVL version runs at a lower power limit and gives less peak Tensor throughput. Your dashboard shows which one your machine has before you start working. Need more compute per card as well as memory? The B200 has 180 GB and roughly twice the FP8 throughput.

H200 rental questions

How much does it cost to rent an H200?

$3.29 per GPU-hour, $3.13 on the 160-hour block and $2.96 on the 640-hour block. The smallest block is 8 GPU-hours for $26.32.

Is the H200 faster than the H100?

For compute-bound work they run the same Tensor Cores. The H200 is faster when memory is the bottleneck, such as serving LLMs with long contexts or large batches, thanks to 4.8 TB/s of bandwidth and 141 GB.

Can one H200 run a 70B model?

Yes in FP8, with plenty of room for the cache. In BF16 the weights alone take about 140 GB, so you need two cards.

Where is the cheapest H200 rental?

On 19 November 2026 the only lower median we found was Vast.ai's H200 NVL marketplace at $3.17. Among fixed on-demand prices, our $3.29 was the lowest in our table.

Sources