CoreWeave has named the data layer it wants AI labs to build on
CoreWeave's new partner network names only two data-services vendors, ClickHouse and LanceDB. For the cloud that says most leading foundation model providers train on its platform, that shortlist is a distribution decision, and it hands a small private startup the industry's sharpest channel.
Richard Tang · 3 min read
On September 30, CoreWeave−3.00% — CoreWeave, down 3.00 percent today used its Fully Connected conference to launch the CoreWeave Partner Network: vendors whose integrations it has tested under production load on its cloud 1. It spans the whole stack, anchored by VAST Data in storage, CrowdStrike−0.84% — CrowdStrike, down 0.84 percent today in security and ClickHouse, a columnar-database maker, in data services 1. In that last category the release names only one other vendor, LanceDB 1. That shortlist is the news.
CoreWeave says nine of the ten leading foundation model providers train on its platform 1. It has now put a validation mark on the data layer it expects them to run.
Curation is a distribution decision
CoreWeave is explicit that the list is earned: partners join only when customers are already running them or asking for them 2. The pitch is scale without rebuilding, keeping the tools a team already has 2. Chief marketing officer Jean English frames the strategy the same way, "so much beyond the GPU," a full-stack cloud completed by partners 3. For a company that listed on Nasdaq in March 2025 1, this is a retention play layered onto a business that keeps compounding.
CoreWeave's quarterly revenue has more than doubled year over year since listing
Data
| Revenue | |
|---|---|
| Q3 '24 | $0.58B |
| Q4 '24 | $0.75B |
| Q1 '25 | $0.98B |
| Q2 '25 | $1.21B |
| Q3 '25 | $1.37B |
| Q4 '25 | $1.57B |
| Q1 '26 | $2.08B |
| Q2 '26 | $2.58B |
The bottleneck moved to the curate step
The same day, CoreWeave announced Forge, a development layer that runs the AI loop: run, observe, curate, improve, evaluate, repeat 5. Compute is the abundant part of that loop. What limits teams now is moving multimodal data, meaning text, images, video and audio handled together 6, into training fast, and giving agents memory that survives a session.
Lance, the open-source format under LanceDB, targets exactly that, and the company publishes benchmarks against Parquet, Delta Lake and Iceberg, the table formats data teams already store analytics and training data in 7. The traction is documented rather than promised: ByteDance's cloud unit Volcano Engine rebuilt its AI stack on Lance and cut a seven-day pipeline to one day, and the agent framework CrewAI now ships LanceDB memory by default across 12 million monthly downloads 7.
Who gets paid
CoreWeave gets stickier labs. LanceDB, led by chief executive Chang She, gets the channel: a validated seat in front of the labs already training there, built on a $30 million Series A led by Theory Ventures, with CRV, Y Combinator and Runway~$5B — Runway, private, latest valuation ~$5B 6. Databricks$201.64B — Databricks, private, latest valuation $201.64B Ventures also invested 6, and its parent company, Databricks, develops Delta Lake 8, one of the formats Lance is fighting.
The losers show up as absences: the release names no rival vector database and no Parquet, Delta or Iceberg vendor in data services 1. What no source shows yet is conversion, whether labs on CoreWeave rebuild on Lance the way Volcano Engine did. That is the test to watch. The payoff is shifting to moving production data back into training fast, and that job currently has a shortlist of two.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
