L40S vs L4
The L4 rents for 46% less per hour today — $0.80 against $1.47 — with 24 GB against 48 GB of memory.
Prices collected 2026-09-18 · 15 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| L40S | L4 | |
|---|---|---|
| Median rental, per GPU-hour | $1.47 | $0.80 |
| Cheapest listing | $0.69 | $0.32 |
| Per month at 720 h | $1,058 | $575 |
| Providers pricing it | 30 | 19 |
| Memory | 48 GB GDDR6 | 24 GB GDDR6 |
| Memory bandwidth | 864 GB/s | 300 GB/s |
| GB of memory per $/hr | 32.7 | 30.0 |
| Architecture | Ada Lovelace | Ada Lovelace |
| Class | Datacenter | Datacenter |
| Released | 2023-08 | 2023-03 |
Same provider, both GPUs
The cleanest comparison: 15 providers price both cards, so the difference is the GPU and nothing else. The L40S is cheaper at 2 of them.
| Provider | L40S | L4 | L4 premium |
|---|---|---|---|
 RunPod Community Cloud | $0.79 | $0.44 | -44% |
 GPU.ai | $0.80 | $0.38 | -52% |
 Vast.ai | $0.80 | $0.33 | -59% |
 io.net | $0.89 | $1.05 | +17% |
 Runcrate | $0.97 | $1.05 | +8% |
 Seeweb | $0.98 | $0.44 | -55% |
 RunPod Secure Cloud | $1.09 | $0.49 | -55% |
 Koyeb | $1.20 | $0.70 | -42% |
| $1.20 | $0.57 | -52% | |
 OVHcloud | $1.80 | $1.00 | -44% |
 Modal | $1.95 | $0.80 | -59% |
 Cerebrium | $1.95 | $0.80 | -59% |
 Lightning AI | $2.14 | $0.79 | -63% |
 Qubrid | $2.92 | $1.50 | -49% |
 Amazon Web Services | $3.39 | $1.60 | -53% |
What each one runs, and what a million tokens costs
Weights + KV cache at 8K context + overhead against 90% of each card's memory. Where a model needs more than one card the cell says how many and prices the set at today's median. $/M tokens is the hourly median divided by estimated single-stream output, from memory bandwidth — an estimate for ranking, not a benchmark.
| Model | Bits | L40S | L4 |
|---|---|---|---|
| Kimi K3 | 4-bit | 37× GPU$54.39/hr | 74× GPU$59.14/hr |
| Kimi K3 | 16-bit | 130× GPU$191.10/hr | 260× GPU$207.79/hr |
| DeepSeek V4.1-Flash | 4-bit | 8× GPU$11.76/hr | 15× GPU$11.99/hr |
| DeepSeek V4.1-Flash | 16-bit | 26× GPU$38.22/hr | 52× GPU$41.56/hr |
| Llama 4 Maverick | 4-bit | 6× GPU$8.82/hr | 11× GPU$8.79/hr |
| Llama 4 Maverick | 16-bit | 19× GPU$27.93/hr | 38× GPU$30.37/hr |
| gpt-oss-120b | 4-bit | 2× GPU$2.94/hr | 4× GPU$3.20/hr |
| gpt-oss-120b | 16-bit | 6× GPU$8.82/hr | 11× GPU$8.79/hr |
| Llama 3.1 70B | 4-bit | 2× GPU$2.94/hr | 3× GPU$2.40/hr |
| Llama 3.1 70B | 16-bit | 4× GPU$5.88/hr | 7× GPU$5.59/hr |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~40 tok/s · $10.21/M | ✓ 21.2 GB~14 tok/s · $15.86/M |
| Gemma 4 31B | 16-bit | 2× GPU$2.94/hr | 4× GPU$3.20/hr |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~372 tok/s · $1.1/M | ✓ 19 GB~129 tok/s · $1.72/M |
| Qwen3 Coder 30B-A3B | 16-bit | 2× GPU$2.94/hr | 3× GPU$2.40/hr |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~46 tok/s · $8.88/M | ✓ 18.7 GB~16 tok/s · $13.88/M |
| Qwen3.8 27B | 16-bit | 2× GPU$2.94/hr | 3× GPU$2.40/hr |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~341 tok/s · $1.2/M | ✓ 13.7 GB~119 tok/s · $1.87/M |
| gpt-oss-20b | 16-bit | 2× GPU$2.94/hr | 3× GPU$2.40/hr |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~154 tok/s · $2.65/M | ✓ 6.9 GB~53 tok/s · $4.19/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~43 tok/s · $9.5/M | ✓ 18.4 GB~15 tok/s · $14.8/M |
Buy or rent?
Hours of renting at today's median that add up to the card's approximate list price. List prices are typical quoted figures, not tracked, and do not include the server, power or the person to run it — so the break-even is a floor, not a verdict.
Questions
Is the L40S or the L4 cheaper to rent?
On 2026-09-18 the L40S has a market median of $1.47 per GPU-hour across 30 providers and the L4 $0.80 across 19, so the L4 is 46% cheaper per hour. Among the 15 providers that price both, the L40S is cheaper at 2 and the L4 at 13. Per hour is only half the answer — the table below prices a whole model on each.
Which should I rent for a 70B model?
At 4-bit, Llama 3.1 70B needs about 44.5 GB. On the L40S that is 2 cards at $2.94/hr; on the L4 3 cards at $2.40/hr.
What is the real difference between them?
48 GB GDDR6 at 864 GB/s (Ada Lovelace) against 24 GB GDDR6 at 300 GB/s (Ada Lovelace). Memory decides what fits; bandwidth decides how fast it generates, because each token streams the active weights once. Compute matters for prompt processing and training, not for how fast text appears.