I run VoltageGPU, which rents L40S GPUs by the hour, so this is a comparison written by one of the vendors in it. Our row comes from our public pricing feed (/api/pricing/snapshot, built at 07:45 UTC on 20 November 2026, 9 L40S on demand at that moment). Every other row comes from the provider's own pricing page, read on 19 November 2026 between 20:56 and 21:06 UTC; I read the prices again on 20 November at 07:49 UTC and none had changed.
What the L40S is
From NVIDIA's L40S page:
| NVIDIA L40S | |
|---|---|
| Architecture | Ada Lovelace |
| Memory | 48 GB GDDR6 with ECC |
| Memory bandwidth | 864 GB/s |
| Interconnect | PCIe Gen4 x16, 64 GB/s bidirectional |
| NVLink | No |
| Multi-Instance GPU (MIG) | No |
| Max power | 350 W |
| FP8 Tensor Core | 733 TFLOPS dense, 1,466 with sparsity |
For scale, NVIDIA's H100 page lists the H100 SXM at 80 GB and 3.35 TB/s. The L40S has 60% of that memory and about a quarter of the bandwidth (864 / 3,350 = 0.26).
Three things follow for planning:
- Bandwidth sets the pace of single-stream generation. Each new token means reading the model's weights from memory again, so with one request at a time, the 864 GB/s matters more than the TFLOPS. Batching several requests per pass is how you put the compute to work.
- No NVLink. A model split across several L40S cards exchanges data over PCIe. The simple case is one model replica per card.
- No MIG. You cannot cut one card into hardware-isolated slices; one workload per card.
What fits in 48 GB
Weights alone take parameters times bytes per parameter (GB = 10^9 bytes):
| Parameters | BF16 (2 bytes) | FP8 (1 byte) | 4-bit (0.5 byte) |
|---|---|---|---|
| 8B | 16 GB | 8 GB | 4 GB |
| 14B | 28 GB | 14 GB | 7 GB |
| 32B | 64 GB | 32 GB | 16 GB |
| 70B | 140 GB | 70 GB | 35 GB |
The KV cache, activations and the CUDA context need the rest, and the KV cache grows with context length and batch size. So 8B and 14B models fit in BF16 with room to batch; a 32B model needs FP8 or 4-bit; a 70B model fits only at 4-bit, with little room left for long contexts. The card supports FP8 in hardware, which makes the FP8 column a practical choice, not just a smaller number.
Prices on 20 November 2026
On demand, one GPU, US dollars per GPU-hour, before storage, egress and taxes. Sorted by price.
| Provider | Per GPU-hour | Comes with one GPU | Billing, as the page states it |
|---|---|---|---|
| RunPod, Community Cloud | $0.79 | not listed in what I read | page title: "Per-Second" |
| RunPod, Secure Cloud | $1.09 | 16 vCPUs, 94 GB RAM | page title: "Per-Second" |
| Koyeb | $1.20 | 15 vCPUs, 64 GB RAM, 160 GB disk | per second ($0.000333/s) |
| VoltageGPU (standard, no enclave) | $1.30 | varies by machine | first hour charged at start, then per second; unused time refunded on stop |
| Nebius, with Intel CPU | from $1.55 | 8 to 40 vCPUs, 32 to 160 GB RAM (range) | not read |
| DigitalOcean | $1.57 | 8 vCPUs, 64 GiB RAM, 500 GiB boot disk | per second, 5-minute minimum |
| Vultr, Cloud GPU | $1.671 | 16 vCPUs, 180 GB RAM, 1.2 TB storage | "on-demand and billed hourly" |
| OVHcloud, l40s-90 | $1.80 | 15 vCores, 90 GB RAM, 400 GB NVMe; region GRA11 | hourly rate shown; $1,296 a month listed |
| Nebius, with AMD CPU | from $1.82 | 16 to 192 vCPUs, 96 to 1,152 GB RAM (range) | not read |
| Modal | $1.95 | serverless functions; CPU billed separately per core-second | per second ($0.000542/s) |
Cheaper capacity that can be taken back: Nebius lists preemptible L40S from $0.74 (Intel CPU) and $0.90 (AMD CPU), at a flat rate rather than a moving spot price. On our own feed, the cheapest spot L40S was $0.76 (8 GPUs free), against $1.30 on demand. Our standard price follows the machines free at the moment: the evening before, our on-demand L40S was $0.90 and the cheapest spot one $1.06, above it.
Reading the table
- The spread is 2.5x between the cheapest and the dearest on-demand rate ($1.95 / $0.79). Over a full month of 730 hours, that is $577 against $1,424 for one card.
- The GPU is the same; the box around it is not. RAM per GPU runs from 64 GB to 180 GB in this table. If your job tokenizes or decodes data on the CPU, 8 vCPUs can become the bottleneck before the GPU does.
- Short jobs care about the minimum. DigitalOcean bills at least 5 minutes, and its FAQ says a powered-off GPU droplet is billed until you destroy it. Our first hour is charged up front and the unused part refunded on stop, so even a ten-minute test needs an hour of balance.
- Serverless is a different shape. Modal bills GPU time for its functions per second ("Pay per second for GPUs and CPUs"). Whether that beats a cheaper VM depends on how much of the hour your GPU is busy, which the table cannot show.
Not in the table
- AWS: its G6e page says G6e instances use the L40S, but carries no price; the EC2 pricing page loads prices by script, and I did not read it.
- Scaleway: its GPU pricing page answered HTTP 403 to my client.
- Lambda: its pricing page, as fetched, lists no L40S.
- Vultr bare metal: an 8x L40S server "starting from $0.848 /GPU/hour", a rate the page gives "for pre-paid contract terms", so it is not on-demand.
- Google Cloud, Azure, CoreWeave and others: not checked.
Check that you got an L40S
The L40 and the L40S both have 48 GB, and a listing can blur them. The PCI device ID does not. In the pci.ids database (version 2026.11.19), restricted to NVIDIA's vendor block:
curl -fsSLO https://raw.githubusercontent.com/pciutils/pciids/master/pci.ids
awk '/^10de/{f=1;next} /^[0-9a-f]{4} /{f=0} f' pci.ids | grep -E '^\s(26b5|26b9)\s'
Output on 19 November 2026 at 21:06 UTC (Windows 11, Git Bash):
26b5 AD102GL [L40]
26b9 AD102GL [L40S]
A plain grep on the whole file matches other vendors' devices with the same four digits, hence the awk step. On the rented machine, nvidia-smi gives the ID as device then vendor. Here is the query on my desk GPU, not an L40S, to show the format:
nvidia-smi --query-gpu=name,pci.device_id,memory.total,power.limit --format=csv
name, pci.device_id, memory.total [MiB], power.limit [W]
NVIDIA GeForce RTX 3060 Ti, 0x248910DE, 8192 MiB, 200.00 W
0x2489 is the device, 10DE is NVIDIA. On an L40S, expect 0x26B910DE. The ID is also more precise than the name: pci.ids lists 2489 as GA104 [GeForce RTX 3060 Ti Lite Hash Rate], a variant the name does not mention.
What this shows, and what it doesn't
It shows list prices on each provider's own page at one moment. It does not show:
- Stock. Only our row says how many GPUs were free; a price for a card you cannot get is worth nothing.
- Performance. I rented no L40S from the other providers for this post and ran no benchmark. The specs are NVIDIA's.
- Fine print I did not read: egress, storage, taxes, regional price differences, and the full terms behind "from" prices.
- Privacy. Our L40S runs as a standard GPU: no enclave, not confidential. We do not offer the L40S as a confidential VM.
Prices move. If you cite one, cite it with its date.
Drafted with AI assistance, then checked: every price was read on the provider's page on 19 November 2026, the prices in the table were read again on 20 November, and both commands were run on 19 November with their output copied as printed.
Top comments (0)