Databricks' $190 Billion Round Prices the Whole AI Stack, and the Data Layer Wins

Databricks closed a $5 billion round this week at a $190 billion valuation, per CNBC. Annualized revenue passed $7 billion, and the company's last comparable mark was $62 billion, set less than two years ago. Run the numbers and investors just paid roughly 27 times revenue for a private company that sells data infrastructure.

To see why that number is strange, look at what happened to the rest of the AI stack in the same news cycle. Meta released a frontier-class agent model under an open license, one that reportedly runs on a single consumer graphics card, per the New York Times. The Financial Times reported OpenAI and Anthropic in an open API price war, with cuts on some developer rates reported near 80 percent. The model layer, the part of the stack everyone spent two years calling the moat, is deflating toward free from both directions at once: open weights from below, price cuts from above.

The application layer offers no shelter either. Every SaaS vendor has now shipped a copilot, most of them wired to the same handful of models, and buyers have noticed that features built in a quarter get matched in a quarter. The two layers everyone can see, the models and the apps, are the two layers losing pricing power in public.

Data has gravity and models have none

So the largest private print of the season went underneath both. That is the bet in plain terms: whichever model wins and whoever's agents end up doing the work, the data they run on sits in the same place, governed and metered by the same vendor. Databricks is priced on the observation that data has gravity and models, on this month's evidence, have none.

The company's own numbers carry the record. A valuation tripling in under two years isn't, by itself, remarkable in this market. Tripling while revenue passes $7 billion is different; the multiple held while the base compounded. And a round this size is functionally pre-IPO capital. At $5 billion, the raise is the listing minus the disclosure documents, which means investors with public-market exit math signed off on the price.

ZoomInfo and Zoom gave up the screen

The instructive contrast is a GTM data vendor that moved the opposite direction the same week. ZoomInfo announced that its data graph, roughly 500 million contact records by its own count, is now reachable inside Microsoft Copilot, Dynamics 365, Excel, and Word through an MCP connector. A decade of building a destination, and the destination just became a feature of somebody else's surface. Read charitably, the data is valuable enough to live anywhere. Read plainly, the interface stopped being defensible, and ZoomInfo chose to keep the data business and surrender the app. Either reading concedes the same point: the durable asset was never the screen. It was the records.

ZoomInfo isn't an outlier; the pattern showed up twice in the same stretch of news. Zoom's Revenue Accelerator shipped an MCP server of its own, exposing its revenue intelligence to third-party AI platforms, and restructured the product's pricing from seats to consumption in the same release. A conversation-intelligence vendor just agreed that its insights will be consumed inside surfaces it doesn't control, and repriced onto usage because the seat stopped describing how the value flows. Two GTM vendors, one month, the same concession.

The protocol choice is the detail worth pausing on. Both companies wired themselves into other surfaces through MCP, an open standard any vendor can implement. When the plumbing between data and interface becomes a commodity protocol, the interface loses its last structural defense, which is friction. What remains defensible is the thing on the far end of the pipe. That is the same verdict the Databricks round priced, delivered from the opposite end of the market: the labs' investors and the GTM vendors' product teams converged on it independently, in the same month.

Here is the gap between what the market says and what it does. Every vendor deck this year leads with an AI strategy. Capital is paying for data gravity, not AI strategies, because the copilots got cheap and the thing they need to be useful, clean proprietary data with governance attached, didn't. The stack's scarce input moved down a layer, and the valuations followed it.

The counter-case has two honest parts. First, 27 times revenue in a private round is a negotiated number, not a market price; late-stage rounds carry structure, and the multiple may flatter the headline. Second, part of what Databricks sells is not data infrastructure at all but a decade-old enterprise sales machine, the kind the frontier labs spent this week trying to hire into existence. Both objections stand, and neither changes the direction. Structured or not, $5 billion moved, and it moved into the data layer during the same week the model layer cut prices by most of their value. Direction is the signal even when the multiple is noisy.

Data compounds while models deflate and interfaces churn

For a GTM leader, the exercise is a renewal-book audit, one question per vendor: does this company own a data asset, or is it an interface? A data asset means proprietary records, a graph, or exhaust the vendor collects and compounds. An interface means a workflow over a rented model and your own data. The test is simple to run: imagine the vendor's model costs going to zero for everyone, because they roughly are. If nothing about the vendor's pricing power changes, you are paying for data. If the whole margin depends on the markup, you are paying for a wrapper, and someone will undercut it before the contract renews.

The second half of the audit points inward. The one data layer every GTM organization already owns, the calls, the CRM history, the product usage, the engagement exhaust, is the only asset in the stack that appreciates while models deflate. The vendors know this, which is why the fine print about how customer data feeds their graphs and models keeps getting longer. Most teams manage that asset the way they manage office furniture. Consolidating ownership of it, knowing what feeds whose graph and on what terms, is worth more over the next three years than any copilot rollout on the roadmap.

The stack has been priced now, in public, by people spending real money. Models deflate, interfaces churn, and data compounds. The companies writing $5 billion checks have decided which layer survives every model release, the vendors on your renewal calendar haven't all gotten the memo, and the space between those two facts is where your next negotiation lives.

By Business of GTM