Databricks Hits $188 Billion Valuation on a Chinese AI Model Bet

Databricks signed a term sheet Thursday valuing the data and AI company at $188 billion, a 40% jump from the $134 billion mark it hit just five months earlier. Coatue Management is leading the round. The cash isn’t in hand yet; Databricks expects the deal to close later this summer.

Eight days earlier, the same company published a very different kind of announcement. Its engineers had spent weeks benchmarking AI coding tools against Databricks’ own multi million line codebase, and the model that came out on top for cost wasn’t from Anthropic or OpenAI. It was GLM 5.2, an open weight model (one whose underlying code is published for anyone to use and modify) built by a Chinese lab called Z.ai.

Coatue Leads a Deal That Isn’t Closed Yet

Databricks said in a statement that it will use the new money to fund its Unity AI Gateway and Lakebase build out, along with Genie, the company’s AI coworker product, and to bankroll acquisitions. The Wall Street Journal first reported the round would total roughly $3 billion; Reuters and Bloomberg have since cited the same figure. Reuters also describes Databricks as one of the world’s most valuable privately held companies, grouped by analysts alongside OpenAI and Anthropic as a likely future listing.

The pace has become a running joke in venture circles. One founder joked about turning on alerts for a Series AA round, a nod to Databricks having already burned through most of the alphabet. Databricks hasn’t said what letter this one carries. The February round was Series L.

Databricks Spent a Decade Preparing for This

Founded in 2013, Databricks built its name in the big data era, selling software that let companies dump enormous volumes of information into the cloud and still run fast analytics on top of it. That history mattered once the AI boom hit. Because Databricks already sat on troves of governed enterprise data, it was positioned to sell AI tools with the same security and compliance enterprises already expected.

The financials back up the pivot. Benzinga reported that Databricks SQL has reached $1.5 billion in annual recurring revenue, growing more than 100% year over year, and that the company expects to exit the first half of its 2027 fiscal year near $6.9 billion in annual recurring revenue, up more than 80%. Last month it acquired cybersecurity startup Panther Labs for an undisclosed sum, after CEO Ali Ghodsi argued at the company’s Data + AI Summit that AI has sped up how quickly attackers turn software bugs into real intrusions.

It is not cheap to run. Even cooling fixes that ease AI’s water bill do little to shrink the electricity these models burn through, which is precisely why cost has become its own product category, not a footnote.

Why Is Databricks Using a Chinese AI Model?

Databricks ran its coding agents through a task by task cost test on its own codebase and found that Z.ai’s open weight GLM 5.2 matched Anthropic’s Opus 4.8 on quality while costing about a third less per finished task, so engineers began routing everyday coding work to it by default.

The test avoided public benchmarks like SWE-Bench on purpose. Databricks argued those tasks leak into training data over time, so it built its own suite from real, human written pull requests across a codebase spanning Python, Go, Scala, Rust and more, then sealed the git history during each run so agents couldn’t just look up the answer.

The evidence shows it’s time to start deploying these as daily drivers for coding.

Databricks’ engineering team wrote that in the post that found GLM 5.2 tied with Opus 4.8 on quality, a post whose listed authors include co-founder and chief technology officer Matei Zaharia. The results split into three tiers: a top group scoring 82% to 90% that included Opus 4.8, GLM 5.2 and GPT 5.5 in some configurations; a middle tier around 71% to 82%; and a bottom tier near 51% to 60%. Databricks’ own chart put GLM 5.2 at $1.28 per completed task against $1.94 for Opus 4.8.

Zaharia later posted the benchmark’s cost findings on X, writing that many models, including open source ones, are truly competitive now. The harness mattered almost as much as the model. Pi, an open source harness, sent roughly three times less context to the model than Claude Code in one comparison, cutting the cost of an Opus 4.8 task by more than half at similar quality.

Coinbase and Lindy Made the Same Bet

Databricks isn’t alone in defecting from the priciest US models for everyday work. Reports on the shift point to a small cluster of companies making similar calls in recent months.

• Coinbase – chief executive Brian Armstrong has said the exchange cut its AI spending in half after shifting engineering tasks to GLM 5.2 and Moonshot AI’s Kimi K2.7.
• Lindy – founder Flo Crivello moved the startup’s entire API traffic off Anthropic’s Claude and onto DeepSeek v4, saving millions after AI costs outgrew payroll.
• Snowflake – Databricks’ closest rival ran its own comparison of GLM 5.2 against Opus and found the two nearly tied on quality at a fraction of the price.

Each case follows the same logic Databricks laid out in its own post: token price and finished task cost are not the same number, and the gap between them is where open models have started winning.

A Five-Week Chain Reaction

The timing lines up with a separate fight in Washington, one that had nothing to do with Databricks at first.

1. June 12, 2026: The Commerce Department orders Anthropic to cut off its new Fable 5 and Mythos 5 models for every foreign national, including the company’s own foreign employees, citing national security.
2. June 13, 2026: Z.ai, the Beijing lab formerly known as Zhipu AI, begins rolling out GLM 5.2, which it says trained entirely on Huawei chips rather than Nvidia hardware.
3. Within days: GLM 5.2 climbs to the top of open model leaderboards and Z.ai’s valuation clears HK$1 trillion, about $128 billion.
4. June 30, 2026: The Commerce Department lifts the export controls after weeks of closed door talks with Anthropic.
5. July 8, 2026: Databricks publishes its internal benchmark, naming GLM 5.2 a daily driver at roughly a third less cost than Opus 4.8.
6. July 16, 2026: Databricks signs its own term sheet at a $188 billion valuation.

Anthropic said the export order’s disable Fable 5 and Mythos 5 worldwide requirement left it no choice but to shut both models off for every customer to stay compliant, even though it disagreed with the government’s reasoning. CNBC later tied the shutdown to a swift rise in Chinese open source models that were proving almost as capable and far cheaper than the most powerful US systems, a shift that worried tech executives who felt Chinese developers had just been handed valuable time.

The Center for Strategic and International Studies flagged something else: the directive leaned on a statutory authority never used this way before, meaning no regulation yet exists to define its limits. That leaves every US model maker guessing whether the same kill switch could be pulled again.

Anthropic and OpenAI Feel the Squeeze

Opus 4.8 remains in Databricks’ top performance tier. Its problem is price. Reporting on the benchmark has pegged Opus 4.8’s per token cost at several times higher than GLM 5.2’s, a gap that gets harder to justify once an internal test shows equivalent output on real work.

The premium shows up elsewhere too. Anthropic’s Claude’s rupee pricing still tops rivals in India even after the company finally localized its rates there this year, a market where cheaper regional and Chinese options compete hardest on cost.

Both labs have their own listing ambitions tangled up in this. Fortune reported that Anthropic confidentially filed paperwork for a public listing in June, around the same time it was fighting the export order. Reuters groups Databricks, OpenAI and Anthropic together as the next likely wave of AI companies to go public. CNBC has already put the underlying tension in a headline: “Why OpenAI and Anthropic may be rushing to IPO amid fears of AI premium fading.”

Databricks Eyes an IPO as Early as Next Year

Ghodsi told investors the company remains on track to go public, potentially as soon as next year, according to Benzinga. Thirteen years after founding, and after four valuation setting rounds in about nineteen months, Databricks still hasn’t taken that step, choosing private capital and secondary sales instead.

A financial newsletter, Finimize, put the risk plainly: “A $188 billion valuation doesn’t stay private for long.” Whatever number Databricks eventually lists at will have to clear a bar the private market just set for it, built in part on tools that made it easier for its own customers to walk away from the most expensive AI money can buy.