Buried under the valuation is a real cost lever: Databricks' own benchmarking shows the coding harness you wrap a model in affects total cost as much as the model choice itself, something most teams aren't optimizing yet.
- Databricks announced a new funding round Thursday valuing the company at $188 billion, led by Coatue, though the money hasn’t closed yet.
- This extends a rapid climb: $134 billion in February, $100 billion in September 2025, and $62 billion in December 2024.
- CEO Ali Ghodsi published internal benchmarking results last week from managing AI costs across 3,000 software engineers.
- Databricks found open models, particularly GLM 5.2, now handle even the hardest coding tasks at lower total cost than Anthropic or OpenAI’s proprietary models.
- More surprising: the coding harness wrapping a model, not just the model itself, equally affects total cost, with the open-source harness Pi performing best on cost without sacrificing quality.
Databricks confirmed the round’s size on Thursday: it raised $5 billion at a $190 billion valuation, higher than the roughly $3 billion earlier reports had estimated and above the $188 billion valuation first disclosed in July. The round was led by Coatue with Blackstone, MGX, T. Rowe Price accounts, and new investor Sixth Street Growth among about two dozen participants.
CEO Ali Ghodsi said the round ended up far larger than planned: Databricks wanted to raise $1 billion, but after The Information reported the fundraise mid-conference, investor interest reached $15 billion. Ghodsi said the company raised more to avoid turning away long-term backers, and to fund heavy AI costs, multibillion-dollar cloud commitments, a 100-person AI research team, and ongoing M&A. Databricks reports a $7 billion annualized revenue run rate, growing 80% and cash-flow positive.
What Happened
Databricks announced a new funding round Thursday valuing the company at $188 billion, led by Coatue, TechCrunch reported. Databricks says the money hasn’t closed yet, with other outlets reporting the raise itself at roughly $3 billion.
The valuation extends a rapid 18-month climb: $134 billion in February, $100 billion in September 2025, and a then-record $10 billion raise at $62 billion in December 2024. Founded in 2013 as a big data analytics company, Databricks has rebuilt its image around AI products including Lakebase, an AI agent database, and Omnigent, a tool for managing multiple agents.
Last week, CEO Ali Ghodsi published internal benchmarking results from managing AI costs across the company’s 3,000 software engineers. Databricks found open models, particularly GLM 5.2, now handle even the hardest coding tasks at a lower total cost than proprietary models from Anthropic or OpenAI.
The more surprising finding: the coding harness, the agentic tool wrapping a model like Codex or Claude Code, affected total cost as much as model choice itself. Databricks found the open-source harness Pi managed context most efficiently, delivering the lowest cost without sacrificing output quality.
Why It Matters
The harness finding is the more actionable detail buried under the valuation headline. Most teams evaluating AI coding costs focus entirely on which model to use, but Databricks’ own benchmarking suggests the wrapper managing that model’s context is an equally significant, and far less discussed, cost lever.
Databricks’ AI transformation has clearly paid off in valuation terms, but this is now the second major voice this month, after Microsoft’s Nadella, publicly framing proprietary AI costs as a problem serious companies need to actively manage. That pattern, enterprise leaders openly discussing AI cost optimization rather than just capability, is itself a signal worth tracking regardless of Databricks’ own valuation story.
Bottom Line
Watch whether other engineering-heavy companies publish similar harness-versus-model cost breakdowns. If this becomes a recurring theme, it confirms the cost optimization conversation is shifting beyond model selection alone.
For SaaS engineering teams running coding agents at scale, benchmark your current harness against alternatives like Pi, not just your model choice. Databricks essentially ran this test for free, its findings are a reasonable starting point for your own evaluation.
