Baseten

Baseten is an AI inference cloud. Instead of general-purpose servers, it runs machine learning models in production. It has four main products. Dedicated Inference serves custom or open-weight models on GPUs you pick, from T4 up to B200. Model APIs give OpenAI-compatible, per-token endpoints for open models such as DeepSeek, GLM, Kimi and GPT OSS 120B. Training covers fine-tuning, including the Loops SDK. Sandboxes give AI agents isolated runtimes. Grid is a newer control layer for routing and governing model access. You package a model with a config file, a Python model class or a custom Docker server, deploy it with the Baseten CLI, and promote it through development, staging and production environments. Autoscaling can scale replicas down to zero. Pricing is pay as you go with no contract. GPU time is billed per minute, from about $0.63 an hour for a T4 to $6.50 for an H100 and $9.98 for a B200. CPU instances, sandbox runtime and storage are metered too. Larger customers can negotiate volume discounts, reserved capacity and SLAs. Logs, metrics and traces export to Datadog, Prometheus, Grafana and New Relic. Enterprise options include single-tenant clusters, a self-hosted data plane in your own VPC, and private networking through AWS PrivateLink or GCP Private Service Connect. Named customers include Cursor, Notion, Abridge and OpenEvidence. The main limitation is scope. Baseten only does inference and training, so you need another provider for databases, storage and general app hosting. Dedicated deployments also bill for idle replicas and cold-start minutes, which can make costs hard to predict.