Enterprise AI’s Bottleneck Is Context, Not Smarter Models

Ali Ghodsi framed Databricks’ current opportunity around a deliberately contrarian claim: artificial general intelligence has already arrived. He grounded that claim less in a formal definition of AGI than in a workplace comparison: whether today’s frontier models are already smarter than most colleagues, most of the time. The same confidence ran through his comments on capital markets, where he described Databricks as able to rely on private funding while preparing, eventually, for public-market liquidity.

His evidence for the AGI claim was informal but pointed. He said he regularly asks audiences whether they believe AGI exists, and usually about 10% raise their hands. When he instead asks whether the frontier models they use are smarter than most of the people they work with most of the time, he said roughly 90% raise their hands.

For Ghodsi, that distinction changes the problem companies should be trying to solve. The frontier models are “already really plenty smart,” he said. The missing ingredient is access to the information that makes work intelligible: conversations, processes, business facts, and other organizational context.

That is also how he positioned Databricks. When Ed Ludlow interjected — “Data. Bricks. That’s what it is.” — Ghodsi agreed. His argument was that if companies could capture and feed their internal context to AI systems, the models could already become “extremely productive.” He contrasted that with what he sees as the industry’s fixation on superintelligence and continued scaling laws.

A Bloomberg Tech graphic shown during the exchange listed “Databricks Partnerships” with Gemini, OpenAI, and Anthropic. Ghodsi connected Databricks’ focus to Genie, which he described as a product for infusing enterprise data context into AI so it can answer and automate questions. In the commercial logic he laid out, Lakehouse holds and organizes the enterprise data, agents need that data to be useful, and Genie is the Databricks product meant to make that context available for question-answering and automation. “I think this is the most important question that people should focus on right now,” he said.

Ghodsi’s case for Databricks’ Lakehouse product rested on a simple premise: if AI makes software easier to produce, the systems behind that software do not disappear. “All software needs a database,” he said. In his view, the market is paying too much attention to whether software itself is being commoditized by AI and too little attention to the fact that every piece of software uses a database.

A Bloomberg Tech customer graphic shown during the discussion listed AT&T, Rivian, adidas, Mercedes-Benz, Unilever, Virgin, and BAYER under “Databricks Customers.” Ghodsi said Lakehouse is being tailored for agents: organizations can place KPIs in Lakehouse so agents can answer questions across those metrics.

Caroline Hyde pressed on whether agent adoption and productivity are actually there. Ghodsi qualified the claim. People are using agents, he said, but not anywhere near their full potential. He estimated that perhaps “1%” of that potential has been tapped, because agents are not yet autonomously collaborating with one another and with humans across organizational work.

Ghodsi’s example was Novo Nordisk, which he identified as a Databricks customer using Genie. He said the company can infuse information from clinical trials into AI systems and then ask questions such as how an obesity study is progressing. According to Ghodsi, that shortened the time needed to understand those studies from weeks to a few minutes.

Ed Ludlow put the software question in terms of Jensen Huang’s thesis: more compute leads to more tokens and more revenue, and agents themselves may become the ultimate users of software. Ludlow said he had found that concept hard to understand and asked Ghodsi to evaluate it.

Ali Ghodsi said Huang was “spot on” on both points. Databricks, he said, had done its own calculation and concluded that within roughly the next nine months — or, rounded out, within one or two years — more software will be written than in all of human history. “No matter how you back into it,” he said, the conclusion checks out.

His implication was not that software demand collapses. It was that the ecosystem around software expands. If much more software is written, there will be more demand for the “other things” around it. Databricks’ place in that ecosystem, as Ghodsi described it, is Lakehouse: a database product seeing strong demand as applications and agents need somewhere to work with data.

That also explained why he rejected a narrow version of the “AI commoditizes software” argument. Even if AI can produce software, software still depends on databases. In his telling, more agent-written and agent-used software increases demand for databases rather than making them less relevant.

At Databricks’ scale, Ali Ghodsi said the company expects to become public because it needs a market transaction mechanism for employees. He put the current employee count at about 10,000, and said that including prior employees the number is probably around 14,000. Public markets, in his view, provide that mechanism.

That rationale was distinct from any need to replace private capital. Ghodsi rejected the idea that a wave of other companies going public would starve Databricks of investor attention or funding. Since Databricks remains private, he argued, private-market investors will soon have fewer comparable companies available to back as more companies leave the private market. That leaves “a lot of capital available in private markets,” he said, and therefore no problem for Databricks.

He also separated the long-term expectation of going public from near-term timing. Ghodsi said this is “a terrible year” to go public because “there’s so much happening.” Asked how closely he read the SpaceX S-1 and how closely he would track an Anthropic IPO as a signal for Databricks’ own path, he said Databricks is not trying to time the market. The focus, he said, is “to really win the market in the long run.” He added that he had not read the SpaceX filing “word by word.”