Databricks is growing at a speed most software companies would envy, but AI agents are making that growth more expensive to serve.
Databricks has the kind of revenue number you notice. As CNBC reported on June 16, the company has reached $6.9 billion in annualized revenue, growing more than 80% year over year, with AI products alone now running at $1.7 billion. For a private software company already valued in the tens of billions, that should be the clean part of the story.
It isn't.
The more interesting number is 74%. That is where gross margins have slipped, down from north of 80% not long ago, according to CNBC's report. CEO Ali Ghodsi told reporters he expects them to fall further, though he didn't give a target. You don't usually want margin compression sitting next to IPO speculation, but that is exactly where Databricks now finds itself.
The culprit, Ghodsi says, is AI agents. Not AI as a slogan. Not some vague future of work pitch. Agents create more queries than people do. A person using Databricks' Genie product might ask one question about corporate data and wait for the answer. An agent can keep working through the same system, checking data, running workflows, triggering actions and asking again. Multiply that across a large company and the cost curve changes quickly.
That is the part every founder selling AI into the enterprise should pay attention to. If your software was priced around human behavior, agents can break the math without breaking the product. The customer may be getting more value. The vendor may be paying more cloud and compute cost to deliver it.
Also read: Snap's $2,195 Specs bet everything on AR being the next smartphone • CriteriaBot Is a Universal Content Classifier That Works for Any Category You Give It • Huawei's chip breakthrough makes the case that US export controls built the rival they were meant to stop
AI revenue is arriving with a bill attached
The timing is not accidental. The Wall Street Journal reported on June 16 that Databricks is launching Genie One, an agentic coworker for business teams in finance, marketing and sales. The company also announced Genie Agents, Genie App Builder and Genie Code, all built around helping companies use their own corporate data in more active AI systems.
That is a stronger pitch than asking companies to dump another chatbot onto employees. Databricks already sits close to the data. It was founded in 2013 by creators of Apache Spark, and its whole business has been built around organizing and analyzing enterprise data in the cloud. If AI agents need context to be useful, Databricks has a credible claim to being one of the places that context already lives.
But credible does not mean cheap to operate. The Journal reported that early users include Albertsons and Rivian, with use cases ranging from merchandising decisions to performance tracking and forecasting. Those aren't toy workloads. They are exactly the kind of repetitive, data-heavy jobs where an agent can keep asking the system for more until the bill shows up somewhere.
Frankly, this is a better test of enterprise AI than another demo video. The question isn't whether an agent can produce a useful answer once. The question is whether the economics still work when the agent is used all day by real teams inside real companies.
The IPO story now has a margin question
Databricks is still in an enviable position. CNBC's reported $6.9 billion annualized revenue figure is enormous for a private software company, and the AI run rate rising from $1.4 billion to $1.7 billion in four months shows customers are not just experimenting. They are paying.
Still, public market investors don't only buy revenue growth. They buy a belief about what that revenue becomes later. Software companies have spent years being rewarded for high gross margins because the old promise was simple: once the platform is built, the next customer is relatively cheap to serve. AI agents complicate that promise.
Look at Snowflake for the comparison. The Journal noted that Snowflake's shares jumped in late May after strong earnings and rising demand for AI tools. Data infrastructure companies are being pulled into the center of the AI trade, but the market will eventually separate companies that can price the usage properly from companies that subsidize it.
Databricks can still solve this. It can change pricing, move more customers toward consumption models, optimize the agent workloads or pass more of the cost through to buyers who are getting more automation out of the product. None of those options is painless. Customers love more usage until the invoice explains what more usage means.
Ghodsi's margin warning is useful because it says the quiet part plainly. AI agents may lift revenue and pressure margins at the same time. If you're building or buying enterprise AI, that is the trade you need to watch, not the launch language. Databricks has shown the demand is real. Now it has to show the business model can keep up with the agents it wants customers to unleash.
