Tensorpool

TensorPool is a GPU cloud for machine learning training and inference. Users create multi node GPU clusters on demand from a command line tool, with commands to create a cluster, attach shared storage and push training jobs through a Git style interface. Normal SSH access is also offered. The company schedules users onto clusters it operates in partnership with compute providers including Nebius, Lightning AI, Lambda Labs, Google Cloud and Microsoft Azure. It also announced nb.dev, an ML focused notebook with a GPU attached, snapshots and branching. The service targets ML engineers and research teams training foundation models who want to avoid managing infrastructure. Clusters use Infiniband networking, and shared storage is engineered for distributed training with stated peak performance of up to 300 GB per second reads and 150 GB per second writes. Pricing is published. The Standard plan is a sign up in seconds with on demand access to up to 128 GPUs and support through Slack. Hourly rates are $4.99 for B200 SXM, $2.99 for H200 SXM, $5.49 for B300 SXM, $1.99 for H100 SXM, $1.49 for L40S and $0.015 for CPU. Shared storage is $100 per TB per month, and S3 compatible object storage is $50 per TB per month plus $0.005 per 1,000 requests, with no egress or ingress charges. Enterprise pricing is available by contacting the company, and reservations and committed discounts are offered.