What is the difference between serverless GPU and instance pricing?
Instance pricing rents you a whole GPU by the hour for as long as you keep it running, idle time included. Serverless pricing bills while a worker runs, from when it starts until it shuts down: your requests plus its startup and a short idle wait. Workers set to scale to zero cost nothing between bursts. Serverless wins for spiky, low-duty workloads, while a dedicated instance wins for steady, high-utilization work.
In our latest check: The cheapest serverless H100 costs $2.50 for each hour a worker runs, against $1.80 for an always-on instance, so serverless is cheaper while its workers run less than 72% of the time. The cheapest serverless A100 costs $1.60 for each hour a worker runs, against $0.89 for an always-on instance, so serverless is cheaper while its workers run less than 56% of the time. The cheapest serverless L40S costs $1.20 for each hour a worker runs, against $0.96 for an always-on instance, so serverless is cheaper while its workers run less than 80% of the time. Worker time counts startup and the idle wait before scale-down, not only requests, and these compare compute rates alone.
With a rented instance you control the machine and pay a flat hourly rate whether it is busy or not. That is efficient when the GPU is loaded most of the time and predictable when you need a persistent environment, but you eat the cost of every idle minute.
Serverless platforms wrap the GPU behind a function or endpoint and charge per second while a worker runs, often with a per-request component. RunPod, for example, bills from when a worker starts until it fully stops, so startup and the idle wait before it scales down (five seconds by default) are billed as well as the requests. Workers that scale to zero cost nothing when no requests arrive, which is ideal for bursty traffic, occasional batch jobs, or a demo that sits quiet most of the day; workers kept always on bill around the clock, like an instance. The tradeoffs are cold-start latency when a worker spins up and a higher effective per-second rate.
The rule of thumb: if your workers run for a small share of the day, startup and idle wait included, serverless usually costs less overall; if you would keep an instance busy for hours at a stretch, the dedicated instance is cheaper. Estimate your requests per day and their duration to see which side you fall on.
Related questions
- Should I look at GPU price per month or per hour?
- What is the difference between spot and on-demand GPU pricing?
- What is a GPU hour?
Numbers on this page come from our verified snapshot of 21 November 2026. Full table on the homepage; method in the methodology.