The lowest cost GPU cloud for AI training on the market.
We’re helping companies train AI at the absolutely lowest cost—enabling you to save more.
50% cheaper training for 10% longer training runs
AI model training can be interrupted. We use that to our advantage to save you money.
Checkpoint your model training scripts to make them interruptible
Power our platform with cheap, renewable energy and pass on those savings to you
Renewable energy is expensive at peak hours
Because of this, we predictably pause our workloads for 10% of the day. With a slight increase in training time, we’re able to offer a cost of compute that other GPU clouds can’t compete with.
Our hardware
Short term reserved capacity.
All the tools your ML team needs
Get started with our low-cost GPUs in as little as 24 hours.
Data transfer support
Access your instances via SSH
Monitor your usage & billing
Slurm & Kubernetes coming soon
Boundless AI training, built from the ground up with the planet in mind.
We bring compute to where cheap, green energy happens
Companies & energy grids are scrambling to keep up with demand.
It will only get worse. We’re a solution.
FAQs
We only focus on optimizing our infrastructure for training and fine-tuning machine learning models, not inference. Competitors like CoreWeave, Lambda, and Crusoe require millions of dollars worth of backup power generation infrastructure to locate their data centers with remote oil fields and wind farms. We're able to shut down and place our data centers closer to the existing power infrastructure to optimize energy consumption and price.
If you spend upwards of 10k per month on pre-training or fine-tuning, we can help. Spend less than that? We'd love to help you scale up your training workloads.
All nodes in the cluster have CUDA and Python 3 installed. You can manage your training environment and install additional python packages as shown in our documentation.
We will assign you a bucket for data storage in our secure data solution. This bucket uses the same API as Amazon S3. See our documentation for more information.
We shut down during daily set time blocks (e.g. 5-8pm “shoulder period” when solar supply goes offline and demand jumps as people come home from work). We work with customers on checkpointing / saving their models ahead of these periods, so we can resume training, without replicating any work, when energy prices fall.
Over time, we will progress to a more dynamic AI-enabled model where we can take advantage of turning off our servers throughout high-cost periods within a day. Customers who would prefer speed>cost can pay a premium to keep their training workloads running during these periods.
To get a much better price! By making a 5–10% sacrifice on training speed, massive energy savings can be unlocked by not operating during high energy priced periods. You can use this trade off to conserve runway, train larger/more models, and to ultimately do it in a way that’s better for the planet.
We lower buildout cost by rapidly deploying our modular systems, and requiring less redundancy because workloads can be paused. For lowering ongoing cost, we use remote low cost renewable energy, shut down or raise prices to offset spiking power costs, & run our modular data centers remotely & autonomously.