GPU rental versus API pricing: how much traffic it takes to break even
A back-of-envelope check: how many output tokens per second a rented H100 must sustain to match current API prices, using published list prices.
Teams weighing self-hosted models against pay-per-token APIs often skip the arithmetic. It is short, and it shows why most workloads belong on an API unless traffic is steady and high.
The calculation
Lambda lists on-demand H100 SXM (80 GB) instances at $3.99 per GPU-hour on an 8-GPU instance, rising to $4.29 for a single GPU, and B200 at $6.69 to $6.99 [1]. Lambda's page shows no pricing date, so treat these as the list prices at the time we checked. Taking the lowest, $3.99:
Derived formula: break-even tokens per second = ($3.99 ÷ 3,600 seconds) ÷ (API price per million output tokens ÷ 1,000,000) = 1,108 ÷ price in dollars. For example, at $1.20 per million output tokens: 1,108 ÷ 1.20 = about 924 tokens per second.
The result is the throughput one GPU must sustain, around the clock, to cost the same as buying those output tokens from the API. All figures in the chart are derived this way, not measured.
Derived: sustained output tokens per second needed to match API price
One H100 SXM at $3.99 per hour, 100% utilization
View data
| Break-even tokens/sec | |
|---|---|
| GPT-5.6 Terra ($12) | 92 |
| DeepSeek V4 Pro ($3.96) | 280 |
| Gemini 3.7 Flash ($3.75) | 296 |
| GPT-5.6 Luna ($1.20) | 924 |
| DeepSeek V4 Flash ($0.28) | 3958 |
Prices used: Terra $12 and Luna $1.20 per million output tokens [2]; Gemini 3.7 Flash's $3.75 [3]; DeepSeek V4 Pro (US$3.96, version V4-Pro-0813) and V4 Flash (US$0.28) as attributed by The Standard to Artificial Analysis [4]. Gemini's $3.75 is described as an introductory price with no end date given [3]. DeepSeek has said it will raise API prices and add peak and off-peak rates, with no new figures given [4].
How to read it
Cheap tiers set a high bar. To beat a $1.20 output price, a single H100 would need to hold roughly 924 output tokens per second continuously. Whether a given model and serving stack reaches that depends on model size, batching and context length; this post does not claim a figure.
Utilization decides everything. The break-even assumes the GPU is busy all hour. At 25% utilization the required throughput quadruples.
Model size can rule out one GPU. heise reports DeepSeek V4-Pro at 1.6 trillion parameters [6], and Moonshot claims 2.8 trillion for Kimi K3 [5]. Assuming 8-bit weights (one byte per parameter), 1.6 trillion parameters is about 1,600 GB (derived) against 80 GB of memory on an H100 [1], before any working memory; lower-precision formats shrink this but not by a factor of 20.
Costs left out. Engineering time, monitoring, redundancy, power or colocation if you buy hardware, and the input-token side of the bill.
Implication for businesses: start on an API, record monthly token volume and peak concurrency, and revisit renting or buying GPUs only when traffic is steady enough to keep them busy or data-residency rules require it.
Sources
- GPU Cloud pricing (on-demand instances) Lambda. lambda.ai/pricing
- OpenAI cuts GPT-5.6 Luna and Terra prices Let's Data Science, 2026-08-01. letsdatascience.com/news/openai-cuts-gpt-56-luna-and-terra-p
- Google launches Gemini 3.7 Flash model Thurrott, 2026-08-13. thurrott.com/a-i/340392/google-launches-gemini-3-7-flash-mod
- DeepSeek launches V4 Pro at prices up to 14 times higher than V4 Flash The Standard (Hong Kong), 2026-08-14. www.thestandard.com.hk/innovation/article/339952/DeepSeek-la
- Moonshot unveils Kimi K3, the world's largest open AI model The Next Web, 2026-07-17. thenextweb.com/news/moonshot-kimi-k3-largest-open-model
- DeepSeek V4: new generation with 1.6 trillion parameters heise online, 2026-04-24. heise.de/-11271628
