Daily billing for GPU cloud servers

From today, GPU cloud servers with NVIDIA RTX 4090, RTX 5090 and L40S cards can be billed by the day. A short training run, a model idea to test, or extra capacity for an event no longer means paying for a whole month.

Why daily billing

Over the past year we have talked with many customers building with AI, and one pattern kept coming up: GPU demand comes in waves. A team might fine-tune for four days straight this week and only run the odd inference next week; at the end of the month they need four cards for two days to demo to a client, and none the rest of the time.

With monthly plans only, they had two bad choices: rent a whole month and watch the machine sit idle, or piece together short rentals elsewhere and move data and environments every time. For young teams, one burns money and the other burns time.

When it fits

• Fine-tuning or evaluating a 7B–30B model, which usually takes a few days.
• Load tests and inference latency and throughput comparisons before a launch.
• AI video, digital humans or batch image generation with heavy work on particular days.
• Research or competitions, with a burst of compute before a deadline.

Three cards, how to choose

The RTX 4090 offers the best value; its 24 GB is enough for fine-tuning and inference on most 7B–13B models and suits image and video generation well. The RTX 5090 has more memory and a newer architecture for mid-sized models that need headroom. The L40S is a data-centre card with 48 GB, built for long, steady runs, with fuller virtualisation and monitoring.

How billing works

A day is 24 hours from when the server starts. Choose 1, 2, 4 or 8 cards; system and data disks are billed per day as well. Delete the server when you are done — there are no unused monthly days. Prices on the product page are live and match checkout, and growth-programme balance and cashback can pay daily invoices too.

A tip: keep your data off the machine

The point of daily rental is starting and stopping freely, so keep datasets, weights and checkpoints in object storage rather than on the server's local disk. Delete the machine today, start another tomorrow, attach the same bucket and carry on. Keeping GPUs and storage in the same region makes reads fastest and avoids cross-region traffic.

When monthly is better

For round-the-clock inference, monthly plans — or data-centre cards such as A100 and H100 — remain the better fit. Overseas GPU servers, like cloud servers, include 1 Gbps of shared bandwidth. Not sure which to choose, or want to mix daily and monthly? Talk to us. Many customers end up experimenting by the day and moving to monthly inference once the model settles.