RTX 4090 vs L40S
The RTX 4090 rents for 59% less per hour today — $0.60 against $1.47 — with 24 GB against 48 GB of memory.
Prices collected 2026-09-18 · 9 providers price both
Index median per GPU-hour, on-demand, per snapshot.
Side by side
| RTX 4090 | L40S | |
|---|---|---|
| Median rental, per GPU-hour | $0.60 | $1.47 |
| Cheapest listing | $0.30 | $0.69 |
| Per month at 720 h | $432 | $1,058 |
| Providers pricing it | 13 | 30 |
| Memory | 24 GB GDDR6X | 48 GB GDDR6 |
| Memory bandwidth | 1,008 GB/s | 864 GB/s |
| GB of memory per $/hr | 40.0 | 32.7 |
| Architecture | Ada Lovelace | Ada Lovelace |
| Class | Consumer | Datacenter |
| Released | 2022-10 | 2023-08 |
Same provider, both GPUs
The cleanest comparison: 9 providers price both cards, so the difference is the GPU and nothing else. The RTX 4090 is cheaper at 9 of them.
| Provider | RTX 4090 | L40S | L40S premium |
|---|---|---|---|
 io.net | $0.30 | $0.89 | +197% |
 RunPod Community Cloud | $0.34 | $0.79 | +132% |
 Packet.ai | $0.39 | $0.92 | +136% |
 GPU.ai | $0.60 | $0.80 | +33% |
 Runcrate | $0.66 | $0.97 | +47% |
 Beam | $0.69 | $0.76 | +10% |
 Vast.ai | $0.72 | $0.80 | +11% |
| $0.72 | $0.96 | +33% | |
 RunPod Secure Cloud | $0.74 | $1.09 | +47% |
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 | RTX 4090 | L40S |
|---|---|---|---|
| Kimi K3 | 4-bit | 74× GPU$44.40/hr | 37× GPU$54.39/hr |
| Kimi K3 | 16-bit | 260× GPU$156.00/hr | 130× GPU$191.10/hr |
| DeepSeek V4.1-Flash | 4-bit | 15× GPU$9.00/hr | 8× GPU$11.76/hr |
| DeepSeek V4.1-Flash | 16-bit | 52× GPU$31.20/hr | 26× GPU$38.22/hr |
| Llama 4 Maverick | 4-bit | 11× GPU$6.60/hr | 6× GPU$8.82/hr |
| Llama 4 Maverick | 16-bit | 38× GPU$22.80/hr | 19× GPU$27.93/hr |
| gpt-oss-120b | 4-bit | 4× GPU$2.40/hr | 2× GPU$2.94/hr |
| gpt-oss-120b | 16-bit | 11× GPU$6.60/hr | 6× GPU$8.82/hr |
| Llama 3.1 70B | 4-bit | 3× GPU$1.80/hr | 2× GPU$2.94/hr |
| Llama 3.1 70B | 16-bit | 7× GPU$4.20/hr | 4× GPU$5.88/hr |
| Gemma 4 31B | 4-bit | ✓ 21.2 GB~46 tok/s · $3.62/M | ✓ 21.2 GB~40 tok/s · $10.21/M |
| Gemma 4 31B | 16-bit | 4× GPU$2.40/hr | 2× GPU$2.94/hr |
| Qwen3 Coder 30B-A3B | 4-bit | ✓ 19 GB~434 tok/s · $0.38/M | ✓ 19 GB~372 tok/s · $1.1/M |
| Qwen3 Coder 30B-A3B | 16-bit | 3× GPU$1.80/hr | 2× GPU$2.94/hr |
| Qwen3.8 27B | 4-bit | ✓ 18.7 GB~53 tok/s · $3.14/M | ✓ 18.7 GB~46 tok/s · $8.88/M |
| Qwen3.8 27B | 16-bit | 3× GPU$1.80/hr | 2× GPU$2.94/hr |
| gpt-oss-20b | 4-bit | ✓ 13.7 GB~398 tok/s · $0.42/M | ✓ 13.7 GB~341 tok/s · $1.2/M |
| gpt-oss-20b | 16-bit | 3× GPU$1.80/hr | 2× GPU$2.94/hr |
| Llama 3.1 8B | 4-bit | ✓ 6.9 GB~179 tok/s · $0.93/M | ✓ 6.9 GB~154 tok/s · $2.65/M |
| Llama 3.1 8B | 16-bit | ✓ 18.4 GB~50 tok/s · $3.33/M | ✓ 18.4 GB~43 tok/s · $9.5/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 RTX 4090 or the L40S cheaper to rent?
On 2026-09-18 the RTX 4090 has a market median of $0.60 per GPU-hour across 13 providers and the L40S $1.47 across 30, so the RTX 4090 is 59% cheaper per hour. Among the 9 providers that price both, the RTX 4090 is cheaper at 9 and the L40S at 0. 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 RTX 4090 that is 3 cards at $1.80/hr; on the L40S 2 cards at $2.94/hr.
What is the real difference between them?
24 GB GDDR6X at 1,008 GB/s (Ada Lovelace) against 48 GB GDDR6 at 864 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.