Databricks Hits $188B While Defaulting to Chinese AI for Code

Databricks signed a term sheet on July 16 for a new strategic funding round that values the company at $188 billion, with Coatue leading roughly $3 billion in fresh capital. The deal is expected to close later this summer, according to the company's own newsroom post and reporting from TechCrunch and SiliconANGLE.

What got less attention is what Databricks did eight days earlier. On July 8, a blog post signed by co-founder and CTO Matei Zaharia and four colleagues announced that GLM-5.2, an open-weight coding model from China's Z.ai, had become the default engine for coding tasks across the company, ahead of Claude Opus 4.8. The reasoning behind that switch is arguably the more consequential story, and it explains exactly what the new $188 billion is meant to protect.

TL;DR

Databricks signed a term sheet for a $3 billion round led by Coatue at a $188 billion valuation, up from $134 billion in February and $62 billion in December 2024.
The raise lands eight days after Databricks made GLM-5.2, a Chinese open-weight model from Z.ai, the default coding engine for its own engineers.
Internal testing found GLM-5.2 matched Claude Opus 4.8 on quality at $1.28 per coding task versus $1.94, a 34% cut, according to the blog post from CTO Matei Zaharia's team.
Coinbase says it cut AI spending by nearly half after defaulting to GLM-5.2 and Kimi K2.7, and Snowflake found similar cost parity in its own tests against Opus.
Chinese-origin models have held more than 30% of weekly token volume on OpenRouter since February 8, up from an 11% average the year before.

The Term Sheet

Databricks has now been repriced four times in under two years. A $10 billion Series J round closed in December 2024 at a $62 billion valuation. Nine months later, in September 2025, a $1 billion raise pushed that to over $100 billion. By February 2026, a combined $5 billion in equity and $2 billion in new debt capacity brought it to $134 billion. The new term sheet adds another 40% on top of that, in five months.

Date	Valuation	New Capital	Lead Investor
Dec 2024	$62B	$10B (Series J)	Thrive Capital
Sep 2025	$100B+	$1B	a16z, Insight, MGX
Feb 2026	$134B	$5B equity + $2B debt	Goldman Sachs, QIA
Jul 2026	$188B	~$3B	Coatue

CEO Ali Ghodsi framed the raise around what he called a shift in enterprise buying behavior. "Enterprises are moving from tokenmaxxing to valuemaxxing," he said in the announcement. "They don't want to burn expensive tokens on the smartest model for every task. They want the best outcome per dollar."

Ali Ghodsi, co-founder and CEO of Databricks Databricks CEO Ali Ghodsi has framed the funding as a bet on "valuemaxxing" over raw model capability. Photo by Michel Edens. Source: commons.wikimedia.org (CC BY-SA 4.0)

Where the New Money Goes

The company says the capital will expand three products: Unity AI Gateway, a governance layer that routes requests across models and tracks their cost; Genie, an AI assistant that lets employees query business data in plain language; and Lakebase, a serverless Postgres database built to store what AI agents remember between tasks. Databricks has also been acquisitive, paying roughly $1 billion for the serverless database startup Neon in 2025, and the newsroom post signals more deals are coming. More than 20,000 organizations use the Databricks platform today, including 70% of the Fortune 500.

The Benchmark That Started It

Zaharia's team built its own coding benchmark instead of relying on public suites like SWE-Bench, arguing that published results leak into training data and rarely resemble a real company's codebase anyway. They pulled real, human-authored pull requests from Databricks' own multi-million-line repository, spanning Python, Go, TypeScript, Scala, Rust, Java, Protobuf, and Bazel, then filtered for recency and self-contained changes with solid test coverage. Roughly a quarter of the tasks were rated low complexity, about 60% medium. Scoring came from whether the code actually passed its tests, not from another model acting as judge.

GLM-5.2 came out statistically tied with Claude Opus 4.8 on quality, both landing in the 82% to 90% pass-rate range. On cost, they weren't close: $1.28 per task for GLM-5.2 against $1.94 for Opus 4.8, a 34% gap.

"The evidence shows it's time to start deploying these as daily drivers for coding," Zaharia's team wrote.

Close-up of a data storage server rack with status lights The benchmark ran on Databricks' own multi-million-line codebase rather than a public leaderboard, using real pull requests and pass/fail test results instead of an LLM judge. Source: unsplash.com

The Harness Matters as Much as the Model

One detail buried in the post complicates the tidy cost story: the tool wrapping the model changed results by more than 2x even when the underlying model stayed the same. A harness called Pi sent roughly a third the context per turn that Claude Code or Codex-style harnesses did, and that alone moved the price. Databricks' conclusion wasn't "GLM-5.2 wins" so much as "model choice is only one variable, and most enterprises are only tracking one."

Coinbase and Snowflake Ran the Same Experiment

Databricks isn't alone. Coinbase CEO Brian Armstrong said the exchange now defaults its engineers to open-weight models, specifically GLM-5.2 and Moonshot AI's Kimi K2.7, through its internal gateway, cutting AI spending by nearly half.

"You may want a frontier model for planning, but not for execution where they can be overkill," Armstrong said.

Snowflake ran a comparable test against Claude Opus 4.7 and landed at roughly the same conclusion: near-tied quality, a fraction of the price. GLM-5.2's own published pricing, $1.40 per million input tokens and $4.40 per million output tokens, undercuts Opus 4.8's $5 and $25 by a wide margin.

The OpenRouter Numbers Behind the Shift

Candlestick chart on a screen showing an upward trend Databricks' valuation has climbed from $62 billion to $188 billion in under two years, tracking the same curve as enterprise adoption of cheaper open-weight coding models. Source: unsplash.com

None of this is happening in isolation. Chinese-origin models have held more than 30% of weekly enterprise token volume on OpenRouter every week since February 8, according to the-decoder's analysis of the router's public data, peaking near 46% at points this year. A year earlier, that average sat at 11%. Z.ai and DeepSeek are named as the platforms driving the shift, at prices researchers describe as 60% to 90% cheaper than comparable US offerings. Zhipu AI first opened GLM-5.2's weights under an MIT license in June, a deliberate bet that developers would pay in adoption what the company gave up in licensing fees.

Databricks is the loudest enterprise name to say the quiet part out loud: the company raising at a record valuation to build AI infrastructure just told its own engineers to stop paying full price for the AI doing the building. Whether Anthropic and OpenAI can hold enterprise coding budgets against a model that costs a third as much and tests the same now looks like the more urgent question than any single funding round.

Sources: