L40S Price, 48GB VRAM & Local LLMs
Price alert
We'll email you when the eBay (US) price reaches your target.
NVIDIA's L40S (2023) is a 48GB VRAM card built for servers that answer many users at once, and it runs 30B-class MoE models at about 131 tokens/s. It has no fan, and an ordinary workstation can't keep it cool.
eBay
L40S price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
L40S VRAM & full specs
- Memory48 GB · GDDR6
- Bandwidth864 GB/s
- FP16 / BF16362.1 TFLOPS
- FP8 / INT8733 / 733 TOPS
- Bus width
- 384-bit
- Architecture
- Ada Lovelace (AD102)
- Cores
- 18,176
- Tensor cores
- 568
- Ecosystem
- CUDA
- Interface
- PCIe 4.0 x16
- NVLink
- Not supported
- Power
- 350 W
- Power connector
- 16-pin
- Slots
- 2
- Length
- 267 mm
- Type
- Data center
- Launch date
- Aug 8, 2023
Best L40S deals
Powered by VRAMGlass real-time price tracking — handpicking the best-value hardware deals across the web
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
UsedNVIDIA L40S 48GB GPU Deep Learning Graphics Card PN: 699-2G133-0242-L00$7,800+ shippingBuy on eBay
UsedNEW BULK NVIDIA L40S 48GB GPU Deep Learning Graphics CardTop Rated$8,799Buy on eBay
UsedNvidia L40S - Passive PCle 48GB PN: 900-2G133-0080-000$8,950+ shippingBuy on eBay
NewP67869-001 NEW SEALED NVIDIA L40S 48GB PCIe Accelerator$8,950Buy on eBay
NewS2L70C NEW SEALED NVIDIA L40S 48GB PCIe Accelerator$8,950Buy on eBay
UsedDell Nvidia L40S 48GB GPU AcceleratorTop Rated$8,999+ shippingBuy on eBay
As an Amazon Associate I earn from qualifying purchases. The price you pay stays the same.
What LLMs can a single L40S run?
A single L40S fits up to 35B at 8-bit, but bandwidth is modest: 27B dense at 4-bit gets only 30-odd tokens/s, while 30B-class MoE holds 100 at 8-bit.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 12B12B | 4.7 GB | Runs | 113.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 135.7 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 51 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 60.3 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 146.5 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 49.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 156.9 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 7.1 GB | Runs | 80.2 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 113.3 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 36.8 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 37.4 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 128.4 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 33.2 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 131.4 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 12B12B | 12.7 GB | Runs | 47.8 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 90.1 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 22.2 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 21.9 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 101.7 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 19.3 | 202K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 105.6 | 256K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can multiple L40S cards run?
Four L40S cards (192GB) run DeepSeek V4 Flash at 4-bit at about 64 tokens/s, and two suit gpt-oss 120B. They usually come in fours or eights, with no NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 2 cards | $20,254 | 96 GB | 106.9 | 700 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | $40,508 | 192 GB | 86.8 | 1,400 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | $40,508 | 192 GB | 41.4 | 1,400 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | $40,508 | 192 GB | 70.4 | 1,400 W |
| GLM 5.3744B (40B active) | 253.9 GB | 8 cards | $81,016 | 384 GB | 36.4 | 2,800 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 8 cards | $81,016 | 384 GB | 27.5 | 2,800 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 8 cards | $81,016 | 384 GB | 32.5 | 2,800 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 8 cards | $81,016 | 384 GB | 43.9 | 2,800 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 8 cards | $81,016 | 384 GB | 43.8 | 2,800 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 4 cards | $40,508 | 192 GB | 90.7 | 1,400 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 4 cards | $40,508 | 192 GB | 64.3 | 1,400 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | $81,016 | 384 GB | 31 | 2,800 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | $81,016 | 384 GB | 45.6 | 2,800 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 4 cards | $40,508 | 192 GB | 62.4 | 1,400 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | $81,016 | 384 GB | 66.7 | 2,800 W |
| Hy3295B (21B active) | 317.7 GB | 8 cards | $81,016 | 384 GB | 21.4 | 2,800 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 8 cards | $81,016 | 384 GB | 29.5 | 2,800 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- You want 27B to 31B dense models at 8-bit
Qwen3.8 27B at 8-bit can use the full 256K, at about 22 tokens/s
- You already have a GPU server that takes four to eight cards
Four cards total 192GB, enough for DeepSeek V4 Flash at 4-bit
- You want FP8 in vLLM
Ada supports FP8 compute, so FP8 models get a real speedup
Not a good fit
- Going into a regular desktop or tower workstation
It has no fan and overheats without server-grade airflow
- Driving a monitor as your main card
The four DP ports are disabled from the factory, and you need NVIDIA's tool to switch modes
- Running 2026's big MoE models on one card
48GB can't hold even Qwen3.8 Flash Next at 2-bit, so you need at least two
Before you buy
Servers only
The L40S has no fan and relies on strong front-to-back airflow through a server chassis. On the NVIDIA developer forums, someone who put one in a Dell Precision 7875 workstation saw it hit 100°C at idle, and replies agreed it isn't suited to workstations. Before buying, check server vendors' compatibility lists; Lenovo's SR650 V3 and SR675 V3, for example, are both on them.
No power cable in the box
Power comes through a single 16-pin connector, at 350W under full load. The card doesn't come with a power cable; Lenovo's product guide specifies that the cable is ordered separately, with a length matched to the server model (from 200 to 450 mm). If you're buying a card pulled from a server, be especially sure your server has the matching cable.
Video output is disabled by default
The card has four DP ports, disabled at the factory; you need to switch display modes with NVIDIA's displaymodeselector tool before you can connect a monitor. Using it purely as a compute card isn't affected.
Don't mix it up with the L40
The L40 and L40S share the same chip and the same 48GB, but the L40S has more compute and draws 350W. Used listings often mix the two up, so look for the S in the model name.
Compared with similar GPUs
L40S vs RTX 6000 Ada, A100 80GB PCIe, and RTX PRO 5000 48GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | L40S | RTX 6000 Ada | A100 80GB PCIe | RTX PRO 5000 48GB |
|---|---|---|---|---|
| Memory | 48 GB | 48 GBSame | 80 GB67% more | 48 GBSame |
| Bandwidth | 864 GB/s | 960 GB/s11% higher | 1,935 GB/s124% higher | 1,344 GB/s56% higher |
| BF16 / FP16 compute | 362 TFLOPS | 364 TFLOPSSame | 312 TFLOPS14% lower | 258 TFLOPS29% lower |
| INT8 compute | 733 TOPS | 729 TOPSSame | 624 TOPS15% lower | 516 TOPS30% lower |
| Qwen3.8 27B speed (2-bit) | 60.3 t/s | 66.3 t/s10% faster | 122.2 t/s103% faster | 89.5 t/s48% faster |
| Largest dense model (4-bit) | 79B | 79BSame | 136B72% more | 79BSame |
| Power | 350 W | 300 W14% lower | 300 W14% lower | 300 W14% lower |
| GPU interconnect | Not supported | Not supported | NVLink 600 GB/s | Not supported |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 5.0 x16 |
| Released | 2023-08 | 2022-09 | 2021-06 | 2025-03 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
L40S rental price per hour
On-demand rental of the L40S starts at $0.79 per hour (RunPod), about $192/month at 8 hours a day. Buying one (eBay median asking price $10,127) pays off in about 9.5 years, 4.8 years or 1.6 years at 4, 8 or 24 hours a day.
| Option | Price | 4 h/day | 8 h/day | 24 h/day | View |
|---|---|---|---|---|---|
 RunPodRentUpdated 19 min. ago | $0.79/hr | $96/mo | $192/mo | $577/mo | Rent |
 Vast.aiRentUpdated 19 min. ago | $0.8/hr | $98/mo | $195/mo | $585/mo | Rent |
 eBayUpdated 2 hr. ago | $10,127 | Payback ~9.5 yrsPower $8/mo | Payback ~4.8 yrsPower $15/mo | Payback ~1.6 yrsPower $46/mo | Buy |
Break-even uses the US residential average ($0.18/kWh) and 350 W rated power, excluding resale value and the rest of the build.
Rental links are referral links: we may receive a commission or credit if you sign up through them. As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Amazon.jp
Xianyu