Databricks announced a new funding round on Thursday that values the company at $188 billion, led by Coatue in what sources peg as roughly a $3 billion raise. But the number that matters more than the valuation arrived quietly alongside it: CEO Ali Ghodsi published internal benchmark data showing that an open-weight model from Chinese AI lab Z.ai matched proprietary models from Anthropic and OpenAI on his own company's coding tasks, at a fraction of the cost. That is the real story behind the raise, and it explains why the market is willing to pay 1.4x the company's February valuation despite frothy public market conditions.
Databricks disclosed the round before the money actually landed in its bank account, an unusual move that speaks to how oversubscribed the deal is. The company has been on a staggering fundraising trajectory, going from $62 billion in December 2024 to $100 billion in September 2025 to $134 billion in February 2026 and now $188 billion. That nearly tripling in 18 months is not driven by revenue multiples alone. It is driven by a thesis that Databricks has become the infrastructure layer on which enterprise AI runs, and that its bet on open models positions it to win the cost war that is just beginning.
The Benchmark That Changes the Cost Calculus
Ghodsi published findings from internal testing conducted across Databricks' 3,000 software engineers, comparing AI models on the actual programming tasks the company's developers perform day to day. The results were striking: Z.ai's GLM 5.2 handled even the highest difficulty coding tasks at a quality level matching Anthropic's Claude and OpenAI's GPT-5.6. The difference was cost. Open-weight models deployed on Databricks' own infrastructure cost a fraction per token compared to proprietary API calls, and the company found that the choice of agentic harness wrapping the model mattered as much as the model itself.
The harness finding is the detail that founders should pay closest attention to. Databricks tested tools like Codex, Claude Code, and the open-source harness Pi, and found that Pi managed context windows and prompt sequencing more efficiently than the proprietary alternatives. The lesson, as Ghodsi's team wrote, is that model choice is only one piece of the puzzle. The agentic orchestration layer the harness that routes prompts, manages conversation history, and determines when to call tools can produce bigger cost swings than switching from Claude to GLM. For a company with thousands of engineers, those savings compound into eight-figure annual reductions.
Why Open Models Are Databricks' Best Moat
The conventional narrative about AI competitive advantage holds that proprietary models are the moat. Databricks is betting the opposite: that owning the infrastructure on which open models run is a better long-term position than owning the model itself. The company's product lineup Lakebase for AI agent databases, Unity as an AI gateway, and Omnigent for multi-agent orchestration all sit underneath the model layer, not on top of it. By being model-agnostic, Databricks captures value regardless of which foundation model wins, and it benefits when open models improve because they drive more enterprises to run inference on its platform.
This logic mirrors what happened in the cloud computing era. Amazon Web Services did not try to own the operating system or the application layer. It owned the abstraction below both, and that proved to be the most durable position in the stack. Databricks is executing the same playbook for AI. The $188 billion valuation reflects investor belief that the company has successfully completed this transition from a data warehouse vendor to the underlying infrastructure that connects enterprise data to any AI model, open or closed, at any scale.
What This Means for Founders Building on AI
Ghodsi's published benchmarks carry a direct implication for every founder running an AI-native company: the model premium is collapsing faster than most realize. If Databricks with its 3,000 engineers can switch to open-weight models and save millions without sacrificing quality, most startups can do the same. The cost and latency advantages of open models deployed on your own infrastructure versus paying per-token to API providers are widening as Chinese labs like Z.ai, Moonshot AI, and DeepSeek continue to release frontier-competitive models at zero licensing cost.
The strategic question for startups is no longer which model performs best on a leaderboard. It is whether you own enough of your own infrastructure to benefit when the model layer commoditizes. Companies that are entirely dependent on a single API provider face margin compression as open alternatives improve. Companies that build on top of a platform like Databricks's or on their own infrastructure can capture the difference as profit. The $188 billion round is a bet that the second group is where the next generation of billion-dollar AI companies will emerge.
The fundraising frenzy also carries a subtler signal about market timing. Databricks announced before closing because demand was so high that the company had no fear of the deal falling through. In a market where late-stage rounds are taking longer and facing more scrutiny, an oversubscribed round at a 40 percent premium to the previous valuation in five months is a vote of confidence in the infrastructure thesis. For founders raising their own rounds, the Databricks story offers a template: if your company owns a defensible position in the AI stack that is independent of any single model provider, the market is still willing to pay richly. The window is not closed. It is just more discriminating about what counts as a real moat.
