November 20, 2026
placeholder

Written By Roman Chernin, Chief Business Officer at Nebius

GPUs are still the hottest commodity in the tech world. While it’s no longer impossible to get access to NVIDIA’s latest H100 and H200 cards through many cloud computing platforms, it can require an expensive, long-term commitment, paid in advance. Research shows that the majority of AI start-up funding is spent on compute. For smaller start-ups, open-source projects, and solo AI developers, the whole situation can be demoralising.

Things get even more complicated when you consider the shifting compute demands of training, building, and deploying AI applications. Training or fine-tuning a model requires lots of storage with high-speed data access and GPUs running at maximum capacity. On the other hand, prototyping and running inference have a far more varied demand. At some point, you might want four or eight GPUs running, while at others, you might need a single card operating—or even none. And, of course, until you’ve prototyped your application, you might not even know how many and which cards you need to get the performance you want.

This kind of flexibility just isn’t practical or economical with most compute providers. Once you are locked into a plan, you are paying for your compute resources whether you’re using them or not. To make the most of them, you need to accurately project what your future workloads will be. If your projections are too conservative, you will have to deal with extra costs or the possibility of not being able to access enough compute to meet your needs. If your projections are too optimistic, your expensive GPUs are going to sit idle, draining your bank accounts.

All these issues and complications are why I recommend AI practitioners find a compute partner that offers some kind of on-demand platform. While the headline per-hour GPU rates may be higher, the best prices that most compute providers advertise are only available to large customers on long-term deals. If you aren’t in a position to commit to a 12, 24, or even 36-month contract with dozens of GPUs, you aren’t going to get these sweetheart deals.

An on-demand cloud computing platform that fits your budget is thus a more affordable option. That way, you can maximise your compute spend by only paying for GPU access when you need it. At Nebius, for example, we offer the first 1,000 hours of NVIDIA H100 GPU usage for $1.50/hour, including CPU and RAM. We’ve found that this Explorer Tier allows our customers to best optimise the return on their compute investment as they only pay when they are actively training or running inference.

Perhaps as importantly, on-demand providers allow you to build and scale using the same platform. You can build and train your applications on a few cards and then spin up as many as you need for inference as you go to market.

Another advantage of an on-demand provider is that you can typically self-provision your compute resources. If you’re training a new model or otherwise need to spin up a few more GPUs, you don’t have to worry about long term commitments, speaking to a sales rep, or waiting in line. At Nebius, we’ve made it possible for anyone to sign up and provision up to eight GPUs instantly through an easy-to-use self-service platform. It doesn’t matter whether they need the resources for an hour or a month – they’re there and ready to go.

I feel that this is the kind of partner that most AI practitioners need, whether they’re launching a start-up, building an open-source application, conducting research, or anything else. I’m excited to see the whole market shift towards on-demand pricing. As AI practitioners ourselves, we know that what’s good for the community is also good for us.

 

About Nebius

Nebius has launched its new and accessible Explorer Tier for access to GPU compute. With prices starting at USD 1.50 per GPU hour for up to 1,000 GPU hours per month, Explorer Tier is designed to meet the needs of AI developers, startups and researchers.

Through Nebius’s self-service console, users can sign up and start using up to eight NVIDIA H100 GPUs on-demand high-performance GPUs immediately, with no long-term commitments, cumbersome contracts, or even the need to talk with a sales rep.