The check isn’t there yet. But the valuation is real. Databricks announced Thursday it has secured funding at an $188 billion valuation.
It is a massive jump. Coatue led the round. The money will likely land in their accounts later this summer. TechCrunch reports the raise totals roughly $3 billion. Usually, companies wait for the ink to dry before shouting about price tags. Databricks didn’t. So many firms wanted in, keeping it secret was pointless. The deal is solid.
This isn’t the first time. Far from it. Databricks has been running through its funding rounds like a sprinter with nothing left to lose. Or gain, depending how you look at the alphabet. People are already meme-ing about when “Series AA” hits.
Five months ago. February. A $5 billion Series L raised the valuation to $134 billion. Before that, September 2025 brought $1 billion for a $100 billion price tag. December 2024? A record-setting $10 billion raise at $62 billion.
The math is wild. But the reason for the cash rush makes sense. Databricks successfully shed its “big data utility” skin. Back in BC times (Before ChatGent, Before ChatGPT, Before the chaos), it was just another SaaS play for cloud analytics. Fast and secure for enterprise storage. Boring? Maybe. Effective? Yes.
Then the AI wave hit.
Companies wanted artificial intelligence. They also wanted governance. They didn’t want their trade secrets leaking to chatbots. Databricks already held the keys. Their clients stored the sensitive data there. Naturally, Databricks positioned itself as the safe harbor for AI integration.
From Big Data to AI Governance
The pivot worked. The company launched product after product. Lakebase became the database for AI agents. Unity acted as the gateway. Then came Omnigent, a “meta-harness” to manage multiple agents talking to each other.
But here is where Databricks got interesting. Cost. Everyone worries about compute burn. Databricks doubled down on open-weight models. Specifically, Chinese-based ones. This fits the 2026 trend of enterprise cost control. They champion Z.ai’s GLM 5.2.
Why? Because it codes. And it’s cheaper.
CEO Ali Ghodsi decided to test this himself. Last week, he published internal benchmarks covering 3,000 of their own software engineers. Real work. Not synthetic tests. The results backed their stance. Open models like GLM 5.2 handled top-tier coding tasks. The cost? Significantly lower than proprietary giants like Anthropic or OpenAI.
But Ghodsi found a surprise. The model choice is only half the story. The harness matters just as much.
Choosing the Right Agentic Harness
Think of a harness like the wrapper around a model. Tools like Codex or Claude Code. They manage context and instructions. Databricks’ data showed that the wrapper impacts cost heavily. Sometimes more than the brain inside it.
They tested Pi, an open-source harness. It won on two fronts:
1. Excellent context management for every prompt.
2. Low cost without sacrificing quality.
This breaks the conventional wisdom. You don’t just need the best brain. You need the best interface to manage it. Native harnesses aren’t automatically better. Expensive proprietary models aren’t always necessary for the highest difficulty tasks if you wrap them in the right tooling.
The lesson from the Databricks blog was clear: Model choice is one piece of the puzzle. The ecosystem matters more.
Does this mean every startup will drop OpenAI tomorrow? No. Enterprise software moves slowly. Governance requirements are sticky. But the pressure is on. When you can save millions on code generation with a Chinese open-weight model and a smart open-source wrapper, the ROI argument gets very loud.
Databricks bet on infrastructure over hype. It bet on data safety over speed. And now it’s worth $188 billion based on the assumption that enterprises will keep spending on the pipe, not just the water.
For now, they are waiting on the rest of that $3 billion check. The market is voting with its wallet. Databricks seems to be the place they want it to go. Whether the $188 billion tag holds once the money actually clears remains the real test.
