What Does Ali Ghodsi Know And How Databricks is Winning The AI Game | Rajkumar Sen, Arcion Labs

Why Databricks keeps winning the AI infrastructure race

Rajkumar Sen sold his data replication company Arcion to Databricks for $100m, after an $18m offer he says would have been the outcome without Delta Lake and Iceberg. From inside Databricks he explains why the company keeps winning. Ali Ghodsi pairs technical and business judgement, and Databricks buys companies for talent rather than revenue. The wider argument is that every computing shift pays in data management, and the next wave belongs to agents and inference.

Takeaways

• Agents will demand fast, exact databases. Sen expects agents to write software on the fly, and unlike people they will not tolerate slow or wrong answers. Databases will have to answer in milliseconds and be accurate every time. He traces the pattern back through earlier computing shifts, each of which paid off in data management.
• Databricks buys companies for talent and vision. Ghodsi bought MosaicML when its revenue was barely $10m, and repeats the pattern with Tabular, Arcion, Neon and others. Arcion arrived with 20 customers and reached a thousand inside Databricks, and generous RSUs made 90% of its employees millionaires.
• Delta Lake and Iceberg shaped Arcion's exit. Arcion held an $18m offer from another database company for migrations. Sen says that without Delta Lake and Iceberg that would have been the outcome. The open table formats gave it a place in the lakehouse stack Databricks was assembling, and Databricks paid $100m.
• Inference can grow bigger than cloud. Sen compares inference to the early internet, when one Craigslist served everyone and Airbnb and DoorDash came later. Companies such as Baseten and Together AI exist because an application running 700 to 800 calls per second needs setups OpenAI and Anthropic will not build.
• The next model frontier is business data. Sen says the use cases beyond BI are ten times larger, and the data is trapped in ERP systems from Oracle and SAP. He gives an oil and gas rig as the example. Models that already answer questions about documents can reason about business questions too.
• Ghodsi pairs technical and business depth. Sen says he has seen that combination only in Larry Ellison and Thomas Kurian, and that Ghodsi works 20 hours a day. Databricks turned down a reported $40-50bn offer from Microsoft, and Sen expects the company to reach $1tn.

Arcion's path from Blitz to the Databricks deal

1. It started as Blitz, and its first use case came from a Bay Area cybersecurity firm running a 400-node MongoDB cluster that needed a real-time replica for forensics.
2. The first check was $48,000 in 2019, after the customer initially offered $10,000, and Dell became the second customer at $100,000.