Netra Systems

Inference TCO

Cost transparency for routers, marketplaces and direct customers: our published list pricing, a public per-GPU-hour cost basis, and a break-even calculator you can run with your own throughput assumptions.

Published pricing

Model	Context	Input	Cached input	Output
Qwen3.8 27B	131,072	$0.20	$0.05	$0.80

USD per 1M tokens, billed on the token counts in each response's usage object. Full API details: netrasystems.ai.

Public cost basis

Per-GPU-hour economics below use the hardware registry published by SemiAnalysis's InferenceX (frontend bundle, read 2026-10-30). Two public tiers:

Owning at large-hyperscaler volume (costh) — amortized ownership $/GPU-hour.
Rent, 3-year commit (costr) — committed rental $/GPU-hour.

GPU	All-in kW/GPU	Owned $/GPU-hr	Rental $/GPU-hr
B200	1.71	1.73	3.70
B300	1.90	2.26	4.25
GB200 NVL72	1.87	1.86	4.00
GB300 NVL72	2.12	2.31	5.00
H100	1.37	1.17	2.00
H200	1.37	1.22	2.90
MI300X	1.39	0.95	1.30
MI325X	1.69	1.10	1.60
MI355X	2.09	1.50	2.90
RTX PRO 6000	0.975	0.68	0.52
Vera Rubin NVL72	3.30	3.61	8.50
Jalapeño	1.125	1.27	1.27
TPU v7	1.207	1.21	2.00

All-in kW includes host, networking and PUE as published by the source. These are third-party declared values, not Netra measurements.

Break-even calculator

Set a sustained per-GPU throughput assumption and a workload mix; the calculator derives revenue per GPU-hour at Netra's list prices and compares it against both cost tiers. Nothing here is a benchmark claim — the throughput is your input.

GPU

Sustained throughput (tokens/s per GPU)

Input share of tokens (%)

Cache-hit rate on input (%)

Input price ($/1M tokens)

Cached-input price ($/1M tokens)

Output price ($/1M tokens)

Revenue per GPU-hour at full utilization: —
Implied cost per 1M tokens at this throughput: —
Margin at full utilization (owned / rental): —
Owned break-even utilization: —

Rental margin is independent of utilization (you pay for time used); owned break-even utilization is the fraction of hours the GPU must be busy for revenue to cover the amortized ownership cost.

Methodology

Revenue per GPU-hour: 3600 × tput × [s·((1−c)·p_in + c·p_cached) + (1−s)·p_out] / 1e6, where s is the input share and c the cache-hit rate on input tokens.
Implied cost per 1M tokens: cost_per_gpu_hr × 1e6 / (3600 × tput).
Pricing: Netra list prices, published at netrasystems.ai (read 2026-10-30).
Cost basis: SemiAnalysis InferenceX hardware registry, public frontend bundle (read 2026-10-30).
Throughput, mix and cache-hit inputs are caller-supplied assumptions. Netra does not publish internal fleet measurements on this page.