Talk with our engineers by joining the new Onehouse community Slack!
Deploying Onehouse
Run Onehouse on NeoClouds
Onehouse runs on any NeoCloud that provides two primitives: managed Kubernetes, and object storage that is S3-compatible or offers a Hadoop-compatible file system (HCFS) interface. With those in place, Onehouse deploys through bring your own Kubernetes and puts the full data platform next to your GPUs: ingestion, curation pipelines, agent traces, and post-training datasets, all as open tables in the NeoCloud’s object storage. If a NeoCloud lacks a primitive or offers only part of one, contact Onehouse and the support team will work through the deployment with you.
Many NeoClouds, including CoreWeave, Lambda, and Nebius, store training data on VAST Data. Onehouse reads and writes open tables directly on any S3-compatible storage system, including VAST, not only on the NeoCloud’s own object storage. Lakegres also runs vector search of the kind VAST DataBase offers, directly on open table formats at petabyte scale.
NeoClouds are GPU clouds built for AI training and inference. Their Kubernetes and storage services change quickly. The table below records what each one documents today.
What each NeoCloud provides
| NeoCloud | Managed Kubernetes | Object storage |
|---|---|---|
| CoreWeave | CoreWeave Kubernetes Service (CKS) | AI Object Storage, S3-compatible |
| Nebius | Managed Service for Kubernetes | Object Storage, S3-compatible |
| Crusoe | Crusoe Managed Kubernetes (CMK) | Object Storage, S3-compatible, reachable from the same location only |
| Lambda | Managed Kubernetes, on 1-Click Clusters only | S3 adapter over filesystems, partial S3 API |
| IREN | Bare-metal GPU clusters; no managed Kubernetes documented | File storage (WEKA); no object storage documented |
| Nscale | Nscale Kubernetes Service (NKS) | Object Storage, S3-compatible, reserved cloud only |
Full supportPartial or limitedNot documented
Checked against each provider’s public documentation on October 14, 2026. Follow the links for the current state; NeoCloud offerings change often.
Use cases on NeoClouds
Your business data usually lives in a lake on AWS, Google Cloud, or Azure, and the GPUs are on the NeoCloud. Onehouse runs on both, as projects under one account, so one platform carries training data from the lake to the GPUs:
- Curate in the lake. Quanton jobs on the hyperscaler crunch business data into open training tables.
- Version and export. Savepoints version each training set, and the export lands it in the NeoCloud’s object storage.
- Train next to the GPUs. Further curation runs on the NeoCloud and feeds its training runs.
On the NeoCloud, compute runs on its Kubernetes and tables live in its object storage. Use cases there include:
- Capture the data already there. If the NeoCloud runs databases or Kafka-like systems for your applications, Flows capture them into open tables on its object storage: database CDC, Kafka topics, and files. Where there is no Kafka, LakeLog carries the change log on object storage instead.
- Log agent traces as tables. Onehouse AI Gateway routes inference and tool calls for your agents, and the agent trace logging service writes every call as tables you can score, query, and replay. See agent traces as tables.
- Run curation pipelines. Quanton jobs join agent traces to the outcomes recorded in your operational data, filter and score the runs, and export validated, versioned training datasets. See AI training data pipelines on Onehouse and Training data pipelines.
- Post-train next to the GPUs. Hand the curated dataset to a training cluster on the same NeoCloud, or submit it to Baseten, Fireworks, or Together AI through the training data pipeline integrations. Savepoints keep the exact training set reproducible.
- Generate embeddings as data lands. Flows generate vector embeddings at ingest and re-embed only the rows that change, for documents as well as records.
- Serve context to agents. Lakegres serves the same tables to agents over the Postgres wire protocol, under the same access policies as every other query.
- Run Python on GPUs. Ray clusters run distributed Python across CPU and GPU capacity, reading the same tables.
Getting started
Deployment follows the bring your own Kubernetes onboarding: you provide the NeoCloud’s Kubernetes cluster and object storage, and Onehouse installs the data plane and operates it remotely from then on. Contact Onehouse to plan a NeoCloud deployment, including one where the NeoCloud lacks a primitive from the table above.