Inference

LLM Inference GPU Economics: Cost per Hour (2026)

Cheapest verified cloud GPUs for serving Llama-class models: observed hourly rates, VRAM fit, and the decode-bandwidth limit that decides tokens/sec.

Cheapest verified$0.34/hr
Reference modelLlama 3.3 70B Instruct
VRAM (INT4)40 GB
Candidates6 GPUs

The fast answer

Cheapest verified GPU: GeForce RTX 4090 at $0.34/hr on-demand (Vast.ai) among 6 candidate GPUs.

Observed On-Demand
Observed On-Demand RateHIGH
SourceObserved provider API rate (Vast.ai)
VerifiedSep 30, 2026
Value
0.34 USD/hr
Methodology

Lowest on-demand hourly row for this GPU's providers.json key; refreshed daily.

Refreshed daily from provider APIs and market scrapingSep 30, 2026

VRAM needed: Llama 3.3 70B Instruct at INT4 needs 48.1 GB full-stack (weights 35 GB + KV-cache + overhead) at 128,000 tokens.

Cost driver: Memory bandwidth per dollar: the cheapest GPU that fits the model at the target context usually leads, but a bandwidth-starved card stretches latency at load.

Candidate GPUs for LLM Inference

Fit = full-stack VRAM total for Llama 3.3 70B Instruct (INT4/FP16 at model context) ≤ GPU VRAM. Rates are observed rows from data/providers.json, refreshed daily.

GPUVRAMBandwidthINT4 fitFP16 fitOn-demandSpot
H200 SXM5141 GB4.8 TB/s✓ fitsOOM$2.79/hr$2.79/hr
H100 SXM580 GB3.35 TB/s✓ fitsOOM$1.89/hr$1.89/hr
B200 Blackwell192 GB8.0 TB/s✓ fits✓ fits$3.99/hr$3.99/hr
A100 80GB SXM480 GB2.0 TB/s✓ fitsOOM$1.59/hr$1.59/hr
L40S48 GB864 GB/sOOMOOM$0.69/hr$0.69/hr
GeForce RTX 409024 GB1.0 TB/sOOMOOM$0.34/hr$0.34/hr

Reference models for this workload

Methodology & provenance

Hourly rates are the lowest observed on-demand and spot rows for each GPU key in data/providers.json (refreshed daily by telemetry cron). VRAM fit uses the models-registry weight model (params × bytes-per-parameter) plus KV-cache sizing at the model's context window. Throughput ordering follows the decode-time bottleneck: token generation is memory-bandwidth bound, so ranking uses each GPU's spec-sheet bandwidth — no throughput figure is asserted without a benchmark row.

Rates: observed provider API rows, refreshed daily (UTC).VRAM: canonical VRAM engine (weights + KV-cache + overhead + headroom).Full methodology →

Next steps

Frequently Asked Questions

What is the cheapest GPU for LLM Inference?▾
GeForce RTX 4090 at $0.34/hr on-demand (Vast.ai) is the lowest observed rate among the candidate GPUs for this workload. Rates refresh daily from provider APIs.
How much VRAM does LLM Inference need?▾
Llama 3.3 70B Instruct as the reference model needs 40 GB at INT4 / 140 GB at FP16 for weights; full-stack totals including KV-cache are 48.1 GB (INT4) and 172.1 GB (FP16) at 128000 tokens context.
What drives cost for LLM Inference?▾
Memory bandwidth per dollar: the cheapest GPU that fits the model at the target context usually leads, but a bandwidth-starved card stretches latency at load. All rates on this page are observed provider rows — never estimates.