Databricks Hits $190 Billion Valuation As CEO Ali Ghodsi Claims AGI Already Arrived

Databricks cofounder and CEO Ali Ghodsi claims artificial general intelligence (AGI) has already arrived — at least by the definition the industry used before 2022. Investors are now backing that thesis with $5 billion of fresh capital. The company announced that it closed strategic financing at a $190 billion post-money valuation after surpassing a $7 billion revenue run rate and growing more than 80% year over year. On July 16, the company had announced the round at $188 billion after signing a term sheet led by existing investor Coatue; Ghodsi says the higher final figure reflects the larger amount raised and additional shares issued.

Coatue led the recent round alongside Blackstone, MGX and T. Rowe Price, with Sixth Street Growth joining as a new investor. Databricks plans to direct the financing toward three areas it sees as essential to putting AI to work inside enterprises: Unity AI Gateway, which routes workloads across models and controls spending; Lakebase, its serverless Postgres database for agent-built software, which has now crossed a $100 million revenue run-rate; and Genie, which gives AI access to the context buried across an enterprise.

While Ghodsi had previously rejected persistent reports that the company was considering a funding round this summer, he says the decision to raise ultimately reflected the cost of expanding the company's AI business and the opportunity to invest more aggressively in hiring and acquisitions.

"There's a huge amount of interest in some of our AI technologies, especially since the token-maxing phenomenon has gone crazy," Ghodsi told me in an exclusive interview. “Unity AI Gateway lets you route all of your tokens through one system and set budgets for different groups or individuals. That gives companies a way to control their token spending and avoid simply maxing out tokens. We call it switching from token maxing to value maxing.”

Moreover, the platform is extending the same philosophy upward through Omnigent, its open-source meta-harness that sits above coding agents so customers can combine and switch between agents built on different models and frameworks. "Instead of being locked into a single harness, they can move between them, control how those agents operate and manage the costs across them," Ghodsi says. "This is one of our three major bets."

MORE FOR YOU

Ghodsi added that Databricks has open-sourced the gateway through MLflow, giving enterprises more flexibility to avoid vendor lock-in, while Unity Catalog and the gateway provide a common layer for governing data and AI assets. He pointed to 7-Eleven as one customer using Unity Catalog to manage governance across its data and AI environment.

"ARR growing at 50%+ and gross margin stabilizing at 70%+ over the next few years should justify the valuation," says Owen Lau, equity analyst at Clear Street. “The ROI debate at the application layer is still not settled. If the enterprises can't monetize these AI tools or increase productivity, they will likely cut back these data and AI investments.”

Enterprise Context Is The Biggest AI Bottleneck

Ghodsi's definition of AGI is narrower than the version of the term that now dominates the AI industry. He argues that if a system can perform the kinds of intellectual tasks humans perform and is smarter than most people most of the time, it meets the basic definition of AGI.

Moreover, he notes people are instead using AGI to describe superintelligence, a much more extreme concept in which AI could accomplish in seconds what the world’s researchers could achieve collectively, simulate global economies and process humanity's accumulated knowledge. "If that is your definition, then of course it is not here," he says. "It is also unclear whether what the industry is currently building will ever become that."

If AGI is here, why does so little inside a typical company look autonomous? Ghodsi’s answer is context. A model cannot fully reason about a business problem without access to the records, internal rules, permissions, and operational systems to act on what it knows. As a result, "the world remains largely unchanged, except that token spending is rising. There is a major gap between the intelligence AI possesses and the impact it is having."

He sees that gap primarily as an infrastructure problem, which is where Genie and Genie Ontology fit into Databricks' strategy. Ghodsi said the products are designed to connect information such as emails, meeting recordings, and internal data while preserving security and privacy, giving AI the organizational context needed to operate inside a business.

The thesis addresses a real disconnect between capability and business results — but it also conveniently expands the market for the infrastructure Databricks sells around the models. The harder test is whether the gap persists as AI models improve.

Rising AI Costs Are Making Model Routing Essential

Ghodsi pointed to a more immediate pressure building across Databricks’ enterprise customers: the rising cost of running AI agents. Companies are increasingly deploying coding agents across their organizations, but many still rely on powerful models for routine tasks, driving inference costs well beyond the value of the work being performed. He notes, "the resulting productivity gains are not keeping pace with the cost of generating and processing the tokens required to run those systems, raising concerns among CFOs and other executives about whether the economics of enterprise AI are sustainable."

That is the opening Databricks sees for its Unity AI Gateway. The company wants to give enterprises a central place to route AI traffic, establish budgets, compare providers, and move workloads between proprietary and open models rather than allowing every application or team to make its own spending decisions.

Ghodsi explained that companies want the ability to switch AI when a cheaper or better model appears, a point he says Databricks can see directly because more than a quadrillion tokens have passed through its gateway. The more interesting behavior may be at the high end of the market, where Ghodsi says "customers sometimes send the same important question to two models and pay twice to compare the answers."

Databricks is effectively betting that model proliferation will create a permanent need for a layer that decides which model should handle which task and keeps track of the financial consequences. There is still a neutrality problem. The company has financial relationships with major model providers while positioning itself as the layer above them. Ghodsi claims that the company "is multicloud and supports open-source alternatives, so customers are not being pushed toward a single vendor."

Building Lakebase For The AI Agent Database Boom

The database business is where the agent thesis becomes more tangible because AI-generated software behaves differently from software created by human developers. An AI coding agent can create several versions of an application, test them simultaneously, branch environments and discard most of the work within minutes. The infrastructure underneath that software has to keep pace.

Ghodsi estimates that "humanity could write more software in the next nine to 12 months than it has written throughout its entire history, based on the growth in software production he sees from AI coding systems". Every application needs a database, he argues, "and agents need databases that can start almost instantly, scale down when idle and support rapid experimentation".

The company reports more than 16 million Postgres database starts a day, a number that sounds enormous until the mechanics are explained. "They are database launches," he says. "We track them because agents can create databases much faster than traditional systems can, with Lakebase able to launch a database in less than one second compared with minutes for some competing systems."

But a large number of those environments never become long-lived production applications. "The starts figure is not intended to measure economic activity," he says. "Revenue measures economic activity and the start metric demonstrates the speed and throughput at which the system operates."

What makes Lakebase suited to agents is its branching architecture. A company can create a branch of a petabyte-scale database in about a second without making a full physical copy, because Lakebase tracks only the changes. That allows agents to spin up isolated environments, test different approaches and discard them without multiplying the underlying storage footprint. The harder question is whether the surge in agent-generated software will turn that technical advantage into a much larger and more durable database business.

Open Data Strategy Can Power The Next AI Stack

The Lakehouse — Databricks’ signature architecture, which combines a data warehouse’s structure with a data lake’s flexibility — is a victim of its own success. Its core idea, that customers should keep their data in open formats and under their own control rather than locked inside a vendor’s system, has been widely adopted across the data industry. Ghodsi acknowledged that competitors including Snowflake “are not quite there, but moving in the same direction.”

Lakebase and LTAP are the platform's attempt to bring operational databases and analytical systems onto the same underlying data. The company claims that the transactional and analytical engines can remain separate while working from one copy of the data, which Ghodsi claims distinguishes LTAP from earlier HTAP approaches.

Snowflake's constraint is that it is defending a powerful analytical franchise while moving into adjacent database territory. Likewise, SingleStore's constraint is that it can challenge the novelty of the architecture but still has to prove that its approach generates comparable production revenue at scale. Oracle starts from the opposite position: it owns the traditional operational database market and enterprise relationships Databricks is now trying to penetrate.

"Databricks can expand, but AI frontier labs can move downstream while hyperscalers can go upstream. Snowflake is also trying to offer services similar to what Databricks supports. Databricks could be sandwiched in the middle," Owen notes. "A key advantage for Databricks is its open format for data, allowing enterprises to move data in and out freely. This philosophy is attractive to new enterprises which don't have Snowflake."

Ghodsi’s proposed test is simpler than the product-marketing argument, saying, "The proof will be in Lakebase's performance." He went further, saying investors should hold him accountable if Databricks stops talking about Lakebase and should compare competitors based on the revenue attached to their database technology rather than the labels they use for it.

IPO Incoming For Databricks?

The recent financing also gives Databricks more flexibility to delay an IPO. Ghodsi says it is now "very unlikely" the company will go public before Anthropic or OpenAI, pointing to volatile markets, shifting valuations, changing interest rates and a wave of major technology offerings that could make the process more distracting and uncertain.

He stressed that the decision does not mean Databricks plans to remain private indefinitely. The company is already adopting many practices associated with public companies, including regular engagement with institutional investors. The goal, he said, is to remain prepared for an eventual listing while waiting for "a little more stability in the markets" and "calmer waters."

"The AI infrastructure model is more capital intensive than the traditional software model, but Databricks is a mix of the two," says Owen. "Longer term, investors would want to see the company generating positive free cash flow sustainably, with declining capex."