Petaflare

L40S rental with crypto: an NVIDIA L40S 48GB for $0.87 an hour

A datacenter card built for inference and graphics at once, paid only in crypto: 48 GB of memory, FP8 Tensor Cores, ray-tracing cores and three AV1 video encoders. Whole cards, one to eight per machine.

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

L40S rental on Petaflare costs $0.87 per GPU-hour, $0.83 on the 500-hour Project block and $0.78 on the 2,000-hour Fleet block; the smallest purchase is 30 GPU-hours for $26.10. The L40S has 48 GB of GDDR6 with ECC, 733 TFLOPS of dense FP8 and, unlike the A100 or H100, ray-tracing cores and hardware video encoders. Pick it for image and video generation, 3D rendering and serving models up to about 30B parameters.

L40S specs

Ada Lovelace generation. Besides the Tensor Cores, the L40S keeps the graphics hardware of NVIDIA's workstation cards: 142 third-generation RT cores (212 TFLOPS of ray-tracing performance) and three NVENC and three NVDEC engines with AV1 encode and decode.

Architecture
Ada Lovelace
Memory
48 GB GDDR6 with ECC
Memory bandwidth
864 GB/s
BF16 Tensor (dense)
362 TFLOPS
FP8 Tensor (dense)
733 TFLOPS
Board power
350 W
GPU link
PCIe Gen4, no NVLink
Form factor
PCIe

NVIDIA figures from the official L40S page. Tensor figures are without sparsity.

L40S price per hour across clouds

One L40S, as listed on each provider's pricing page on 19 November 2026.

Sources: Massed Compute, RunPod, Vast.ai public price feed (18 November median, cited with attribution). Read on 19 November 2026. Every block is on price per GPU-hour.
ProviderPrice typePer GPU-hour
PetaflareFixed, on demand$0.87 · $0.78 on Fleet
Massed ComputeFixed, on demand$0.97
RunPod Secure CloudFixed, on demand$1.09
RunPod Community CloudThird-party hosts$0.79
Vast.aiMarketplace, median of the day$0.53

Sources: Massed Compute, RunPod, Vast.ai public price feed (18 November median, cited with attribution). Read on 19 November 2026. Every block is on price per GPU-hour.

Reading those L40S prices

Among fixed on-demand prices, our $0.87 is the lowest in the table, and the Fleet block at $0.78 matches RunPod's Community Cloud. Vast.ai's marketplace median is lower at $0.53: the L40S is a common card among independent hosts, who set their own price and bill storage and bandwidth apart. If the lowest hourly price matters more than a fixed one, compare it on the cheapest cloud GPU page.

Paying for L40S hours in crypto

The smallest L40S block, 30 GPU-hours, costs $26.10, and 2,000 hours cost $1,560, or $0.78 an hour. Litecoin arrives in about 15 minutes and USDT on Tron in about one. We accept crypto only, with no card on file: see paying for a GPU with crypto.

Built for images, video and rendering

Most AI cards drop the graphics hardware. The L40S keeps it, which matters for three kinds of work:

  • Image and video generation. 48 GB holds Flux or SDXL with their text encoders and room for large batches, and the FP8 Tensor Cores speed up quantized diffusion models. Video models that need more than the 24 GB of an RTX 4090 fit here. The Stable Diffusion GPU guide lists the memory per model.
  • Encoding the results. Three NVENC engines with AV1 encode video on the card instead of the CPU, so a render or generation pipeline does not stall on encoding. The A100 and H100 have decoders but no encoder.
  • Ray-traced rendering. Blender Cycles with OptiX, Redshift, Octane and V-Ray GPU use RT cores to trace rays. GPU rendering in the cloud compares the L40S with the RTX 4090 and 5090 for that job.

L40S vs A100, H100 and RTX 5090

The four cards people weigh the L40S against, on NVIDIA's published figures and our prices of 19 November 2026.

NVIDIA does not publish dense Tensor figures for GeForce cards in the same form, so the RTX 5090's BF16 and FP8 throughput is left blank rather than estimated.
CardMemoryDense FP8Dense BF16RT coresVideo encodersPer hour
L40S48 GB GDDR6 ECC733 TFLOPS362 TFLOPSYes3 NVENC (AV1)$0.87
A100 80GB80 GB HBM2eNo FP8312 TFLOPSNoNone$0.99
H100 80GB SXM80 GB HBM31,979 TFLOPS989 TFLOPSNoNone$2.39
RTX 509032 GB GDDR7SupportedNot publishedYes3 NVENC (AV1)$0.49

NVIDIA does not publish dense Tensor figures for GeForce cards in the same form, so the RTX 5090's BF16 and FP8 throughput is left blank rather than estimated.

Choosing between them

  • L40S over A100: for inference and image work, the L40S runs FP8, which the A100 cannot, and costs less per hour. Choose the A100 when you need 80 GB or BF16 training with NVLink.
  • L40S over H100: the H100 does about 2.7 times the FP8 math for 2.7 times the price, with 80 GB and NVLink. For serving models that fit in 48 GB, or for anything that needs RT cores or encoding, the L40S does the job for far less.
  • L40S over RTX 5090: the 5090 is cheaper per hour and faster on many render and generation jobs, but it has 32 GB and no ECC memory. Pick the L40S when the model or scene needs more than 32 GB, or when you want ECC on long runs.

What the L40S does not do

  • No NVLink. Cards in a multi-GPU machine talk over PCIe Gen4 at 64 GB/s, so splitting one large model across several L40S cards is slower than on SXM A100s or H100s. Data-parallel jobs, where each card holds a full copy of the model, are not affected much.
  • No MIG. The card cannot be split into isolated slices; you always get the whole GPU anyway.
  • Not for double-precision computing. It is built for FP32 and lower precisions. Scientific codes that need FP64 run far better on an A100 or H100.

L40S rental questions

How much does it cost to rent an NVIDIA L40S?

$0.87 per GPU-hour on Petaflare, $0.83 on the 500-hour block and $0.78 on the 2,000-hour block. On 19 November 2026 other clouds charged between $0.53 (marketplace median) and $1.09 (RunPod Secure Cloud).

Is the L40S good for LLM inference?

Yes for models that fit in 48 GB: 8B to 14B models in BF16 or FP8 with large batches, and 30B-class models in 8-bit. For 70B models, 4-bit weights fit but leave little room for the cache; an H100 or H200 is the better choice.

Can I use several L40S cards together?

Yes, 2, 4 or 8 per machine. They communicate over PCIe rather than NVLink, which suits data-parallel training and serving several models better than splitting one model across cards.

Does the L40S support FP8?

Yes. Its fourth-generation Tensor Cores run FP8 at 733 TFLOPS dense, twice its BF16 rate.

Sources