Part IThe Story
The Trial Run
The news reached him sideways, which is how it usually travels in small companies: secondhand, through the grapevine, in fragments. In 2015 the board of Databricks was interviewing people to run the company, and Ali Ghodsi, its vice president of engineering and product management and one of its seven co-founders, kept hearing about the interviews. He heard how they had talked to that guy, and how they had talked to this one. He knew the identities of at least two candidates. Nothing he heard suggested that he was one of them. "So as far as I knew," he said years later on Sequoia's Long Strange Trip, "I was like, they're not going to even pick me."
The company had a strange problem. Its open-source project, Apache Spark, was a worldwide sensation. Downloads were climbing, the conference was packed, and engineers at banks and telecoms and web companies were rewriting their pipelines around it. Databricks, the company its creators had founded to commercialize it, recorded about $1.5 million in US GAAP revenue that year. "So the vibe wasn't great," Ghodsi said. "And I'm like a glass-half-empty guy. So I was like, oh, I'm probably going to leave and we're all probably going to leave."
So he did the sensible thing and built himself an exit. He applied for a faculty position at the University of California, Berkeley, where he had arrived six years earlier as a visiting scholar who meant to stay for one year. Some retellings say he had already accepted the professorship. In his own account he was still choosing, and the choice, when it came, was between two doors. One led to the job he had wanted his whole adult life, a professor's office at one of the great computer science departments in the world. The other led to a job he had never wanted, chief executive of a company whose most successful product was one it gave away for free.
He chose the second door. Even then it opened only partway. The board gave him the seat as a trial run, and he did not get a CEO's salary.
Eleven years later, in August 2026, Databricks closed a $5 billion funding round at a $190 billion post-money valuation, led by Coatue with Blackstone, MGX, T. Rowe Price and Sixth Street Growth, and announced a revenue run rate above $7 billion. The comparison with 2015 is unfair in the usual way, since it sets audited annual revenue against an annualized run rate. Even allowing for that, the multiple is more than four thousand. A Swedish founder who met Ghodsi in San Francisco pointed out that if the company were listed in Stockholm, it would be the largest in Sweden.
Ghodsi has explained the decision several ways, depending on the interviewer and the year. The most revealing version describes a pattern he noticed in himself: "Every time I faced a choice between the comfortable known path and the genuinely challenging unknown, I had picked the challenge and it had opened my horizons each time." That is the founder-mythology version, and it is probably true. A plainer version runs underneath it, the one about the grapevine and the glass half empty and the faculty application kept ready in case. This man expects the house to come down. He takes the job anyway.
By the Numbers
From Pub Revenue to the Trillion-Dollar Conversation
$1.5MDatabricks US GAAP revenue in 2015, the year the board went looking for a new CEO
$7B+Revenue run rate announced in August 2026
$190BPost-money valuation after the $5 billion round closed in August 2026
$13.9MSeries A from Andreessen Horowitz, September 2013
20,000+Organizations on the platform, including 70% of the Fortune 500
80%+Share of new databases on Databricks launched by AI agents rather than humans
7Co-founders, all from UC Berkeley's AMPLab orbit
The Lights Over Tehran
Before the grapevine there was the blackout.
Ghodsi was born in Iran in 1978, the year before the revolution. His early childhood ran alongside the Iran-Iraq war, which began in 1980 and lasted eight years. He has described Tehran going dark as the lights were cut and Iraqi aircraft came over the city, and himself as a small boy crawling under tables in the blackout. One night the windows shattered and he was sure his own house had been hit. It had not. The houses around it had been. He reports that detail flatly, the way people do with things they have no reason to dramatize.
He is careful about what he takes from it, more careful than the people who repeat his story. "I wouldn't say that sort of experience makes you more resilient," he told Computer Weekly in 2021, "but it does prepare you for anything happening." Most founder biographies would turn the war into a forge that tempered the boy into a CEO. Ghodsi declines that, and the refusal says something about how he thinks. He claims a calibration, not a virtue: experience adjusted his sense of what counts as normal.
— Ali Ghodsi, to Computer Weekly, 2021I wouldn't say that sort of experience makes you more resilient, but it does prepare you for anything happening. One day, everything can be great; the next, everything can have collapsed. It helps you stay calm in a crisis.
In 1984 the family left. By one account Sweden was the first country that would take them. The move was a fall as well as an escape. Ghodsi has described the gap between a privileged childhood in Tehran and the hardship of growing up in what he called a Swedish ghetto, the immigrant estates at the edge of the welfare state, where a family's previous standing counted for nothing. Some displacement narratives move from nothing to something. His moved from something to nothing and then to something else, which teaches a different lesson. Fortune can be revoked, so it should never be confused with permanence.
He was about seven when he started programming. The machine was an old Commodore 64 with a broken tape recorder. In 1980s home computing, that tape drive was how games loaded, so with it broken the machine could not play games. Most children would have given up on it. He opened the manuals and read them, and since the computer would not entertain him he learned to make it do things. By high school he was skipping school and staying up through the night writing code.
The story is told often enough to sound like a parable, and it holds a structure that will come back. He had a machine that refused to do what it was sold for. He did not fix the tape drive. He found out what else the hardware could do. Twenty-five years later he ran a company whose free software nobody would adopt and whose customers would not use it unless they could pay for it. He did not fight that refusal either. He studied what the refusal was telling him.
A Network With No Center
Ghodsi spent his whole education in Sweden, from kindergarten to postdoc, as he puts it. It was a thorough Swedish education, and it left him with ideas about Sweden that do not match the postcard.
In 2002 he finished a master's in computer engineering at Mid-Sweden University. Alongside it he took an MBA in logistics and strategic marketing, an odd choice for a young systems programmer and the first sign of a split between engineer and businessman that would later work in his favor. Then came a doctorate at KTH, the Royal Institute of Technology in Stockholm, under the computer scientist Seif Haridi. The 2006 dissertation was titled Distributed k-ary System: Algorithms for Distributed Hash Tables. It dealt with peer-to-peer networks that have no central server, where any node can disappear and the system still has to find what it needs. A friend from the infrastructure world later described the research as work on completely decentralized systems, "precursors to bitcoin."
The subject fits the life. A distributed hash table assumes that any single node can fail at any moment and designs so that the whole survives. Someone who learned early that a house could be standing one night and its neighbors' houses gone the next might well find that problem natural.
He did not stay purely academic. He co-founded Peerialism AB, a Stockholm company building a peer-to-peer data transfer system. He did research at the Swedish Institute of Computer Science. From 2008 to 2009 he was an assistant professor at KTH. On paper this was a Swedish academic career with a side business, the kind that ends in a named chair and a summer house in the archipelago.
He came to dislike the country's version of how such careers are supposed to go. "I always felt Sweden was built for the benefit of the big companies, like Ericsson, Volvo and ABB," he told Computer Weekly. "Not so much for the little guy wanting to launch his own company." He is impatient with the Sweden that liberals in San Francisco and London admire. That country, he says, is the Sweden of the 1960s and 1970s: "The Sweden of the 1990s wasn't like that. And the Sweden of the 1930s was different again." The refugee who was educated free at every level is also the critic who saw the system's limits, and he finds no contradiction there. Sweden gave him everything except permission.
He found the permission elsewhere and explained it in terms that are really autobiography:
— Ali Ghodsi, to Computer Weekly, 2021The United States was created by refugees over the past 300 years or so. In Silicon Valley, you do find a lot of people who had to leave their home countries. And as a result, the country as a whole is like a startup, compared to where I grew up in Sweden.
He likes to name the precedents. He mentions the Hungarians who came in the mid-1950s, a wave that produced Andy Grove of Intel, and then the waves from India and Russia. He mentions the Iranians in particular: Dara Khosrowshahi at Uber, Omid Kordestani at Google, and the electrical engineers whose training in statistics and probability, he says, turned out to be good preparation for machine learning. In October 2021 Berkeley's Iranian student groups chose him as the first speaker in an entrepreneurship series called "The Tale of a Success." It was a public role he could easily have declined, and he accepted it.
One More Year
In 2009 Ghodsi got an offer to spend a year at Berkeley, working with a group of researchers who would later be treated as something close to legend. He told himself it would be a single year before returning to his professorship in Sweden. He has told what happened next with comic repetition:
"I was here for one year, and said, hey, this is really, really interesting... I'm not going to get another opportunity like this in my life again. So let me give it 12 more months. So I stayed one more year. Then two years had passed, and I said, this opportunity is so great I should give it one more year. I stayed three years. So you see where this story is going."
So the plan was always a year, and the year kept getting renewed. He never made a single decision to emigrate. What he made was a run of small extensions, each reasonable and each reversible, until they added up to something that could not be undone.
What held him was an idea. In his telling, the Berkeley group had realized that ordinary computers were no longer getting faster. "We've hit something called Moore's Wall," he said in 2017. "Moore's Law no longer applies." If single machines had stopped improving every eighteen months, the supercomputer could not keep up with the world's data, and the new computer was the datacenter itself, hundreds of thousands of commodity machines in the cloud that had to be made to act as one. That was a distributed systems problem, which meant it was exactly his kind of problem.
He worked with Scott Shenker, Ion Stoica, Michael Franklin and Matei Zaharia on distributed systems, databases and networking. He helped start Apache Mesos, a platform for sharing a datacenter's resources among frameworks. He co-invented dominant resource fairness, a method for dividing several types of computing resources fairly among users, which went on to influence scheduling design in Hadoop. And he was in the room for Spark.
Spark belonged to Zaharia more than to anyone. A Romanian-Canadian PhD student at Berkeley, he had studied Hadoop MapReduce, the dominant big-data framework of the time, and found it slow at iterative algorithms and interactive data mining, which were the workloads that mattered most for machine learning. His answer was an engine that kept data in memory between steps. Spark was conceived in 2009. Zaharia later became Databricks' chief technologist and a professor at Stanford. He was the immigrant engineer whose system was the asset, while Ghodsi was the immigrant engineer whose job became turning that asset into a company.
Andy Konwinski, another member of the group, described himself in a 2026 commencement address at Berkeley's Greek Theatre as "a kid from rural Wisconsin." He finished his Berkeley PhD in 2012, co-founded Databricks, and later co-founded Perplexity and Laude Ventures. On that stage he told 1,799 graduating computer scientists, data scientists and statisticians something an ambitious graduate rarely hears: "I think you need to stop succeeding." He meant chasing success. "Run at failure," he said. The problems that matter, he argued, are so hard that "you will likely fail," and "you have to work on them anyway."
It reads like a group biography of people who came in from the edges, from Tehran by way of a Swedish housing estate, from Romania by way of Canada, from rural Wisconsin, and ended up building the plumbing of the American economy. None of them seems to have planned it. The research produced papers. Ghodsi co-authored the Mesos paper and the Spark SQL paper, and the AMPLab's general paper on Spark lists him as one of fourteen authors. That last fact is worth stressing against the heroic-founder version of events. Spark was a collective achievement, and Ghodsi's later accomplishment depends on his never having pretended otherwise.
The Gift Nobody Would Take
The academics wanted to give their work away.
This is the least strategic part of the Databricks story and possibly the most important. The aim, as Ghodsi told it to Foundation Capital, was democratization: letting ordinary companies do the kind of data science that had made Silicon Valley's giants so powerful. The method was to publish papers, release code, and pitch large companies on adopting it. "They tried their best to pitch big companies on using their technology," the Foundation Capital summary says, "but no one bit."
That is the first refusal. The second is stranger. When the group began debating whether to commercialize, Ghodsi argued against charging too early: "We need to get adoption first. If we try to sell it now, nobody will use it." Adoption came, and with it came enterprise buyers who would not take the software for free. They said it was not enterprise-ready. In effect they demanded to pay. "We would tell them, 'Just take the software for free,'" Ghodsi has said, "and they would say 'No, we have to give you $1 million.'"
He described the founding to Goldman Sachs' David Solomon as an act of exasperation more than ambition: "So out of frustration, 2013, we started Databricks. We said, you know, if we don't start it and we're going to try to do it ourselves, no one else is going to do this." Seven founders, Ghodsi, Konwinski, Stoica, Zaharia, Patrick Wendell, Reynold Xin and Arsalan Tavakoli-Shiraji, incorporated in the summer of 2013. In September, Andreessen Horowitz put in $13.9 million, and the company told the trade press it meant to offer an alternative to Google's MapReduce.
Ghodsi joined as a part-time engineer. That is easy to overlook. The man who would be named to Time's list of the most influential people in AI started at his own company with one foot still in the university, eventually becoming VP of engineering and product while the professor's office stayed available.
Then the market began to notice. Reynold Xin, the co-founder who would become chief architect, entered the team in a public sorting benchmark: sort 100 terabytes as fast as possible. They spent cycle after cycle tuning and finished first. The results were published, the press picked them up, and in Ghodsi's memory attention arrived almost overnight. Spark Summit filled up and Spark meetups grew around the world. By 2017 the project had more than 300,000 meetup members.
That did not make it a business. "Your biggest enemy is your open source project," Ghodsi said later, in a line that every open-source founder should probably have on the wall. Amazon and the other cloud vendors could distribute Spark themselves. Each prospective customer could ask the obvious question: why can't I just download the open source version? The original plan, to make Spark enormous and then sell the best Spark, had not produced enough that was proprietary. Product-led growth stalled at roughly $3 million in annual recurring revenue.
Some of the damage was self-inflicted, and he says so. Asked by Unusual Ventures about his worst mistake as an entrepreneur, he named this period: "Not building a complete SaaS offering from day one and dabbling around with more 'secure' semi-hybrid SaaS architectures. Most shortcuts don't pay off." Asked what he wished he had known before starting, he said: "The importance of marketing, which I dismissed as nonsense." That is how a computer scientist thinks before a market corrects him.
By 2015 the company had won the open-source contest decisively and was losing the commercial one just as decisively. Ghodsi's later summary of the moment is two short clauses: "Open source success, check. Commercial success, zero."
Professor for a Professor
Ben Horowitz tells the story of the CEO change with the comic timing of a man who has told it many times, and his version differs from others in useful ways.
Horowitz's own path had two notable turns. He studied artificial intelligence at the University of Illinois in the 1980s, when, as he put it, "We all thought it was never going to work." He then founded Loudcloud in the middle of the dot-com bust, a company built on the edge of collapse, sold it for $1.6 billion, and went on to lead Andreessen Horowitz from its founding to $46 billion in committed capital. The firm's whole thesis was that technical founders could learn to be ruthless operators and should not be replaced with professional managers after the Series B. Databricks would test that thesis about as hard as anything could.
In Horowitz's telling, the call came from Scott Shenker, the Berkeley professor Horowitz described as "kind of the professor in the background kind of founder type character, very smart guy, very nice person." Shenker said the co-founders believed the company needed a new CEO and that Ali should be it. Horowitz did not know Ghodsi well, and his first reaction was dry: "You want to swap out one professor for another professor?" He suggested they at least see what an outside executive looked like. The founders said no. They all wanted Ali. Meanwhile, Horowitz said, "the company was in real trouble. Like, it was not going well."
Another version, in a 2026 newsletter retelling, has Horowitz at a tense board meeting in 2015, with revenue "lower than a local pub," pointing across the table and saying, "Let's go with Ali." The two accounts can both be true, since the decision may have formed on phone calls and been announced in a room. Ghodsi's own account adds one more layer: he was the last person to assume he would be chosen. On January 11, 2016, Databricks announced the leadership change. Ion Stoica, the professor who had been running the company, became executive chairman. The professor from Stockholm got the job.
Ghodsi's first and lasting decision as CEO was a judgment about what was blocking the company:
— Ali Ghodsi, CEO Insider, 2026Find the one major bottleneck and put an orders-of-magnitude disproportionate amount of attention on it. Not just a little more attention. A lot more. Almost to the point of going overboard. Because even with that level of focus, you probably will not fully unclog it.
The bottleneck was the commercial engine, and he did not know how to build one. Investors who later worked with him describe his method as plain admission. Founders should accept that they are "zero" in the functions they do not understand, interview the best practitioners they can reach, compare their playbooks where they contradict each other, and hire people good enough to teach them. He had no sales experience. He built one of the industry's most respected enterprise sales organizations largely by questioning salespeople until he understood them. He eventually brought in Ron Grisco to run it. He hired methodically, from a book: of Geoff Smart and Randy Street's Who, he said, "All Databricks hiring is built on this book." He checked backdoor references because, as he put it, "If what people advertised (and believed) about themselves were true, we'd all have the best dream team."
The harshest move was the one most first-time founder-CEOs cannot bring themselves to make. Within about a year he replaced essentially the whole executive team with experienced operators. Each displaced executive, many of them colleagues and some of them co-founders, was offered the chance to stay and report to the person replacing them. Of seven, only two left. "If people were smart enough," Ghodsi said, "they put their egos aside." The company's real advantage had not gone anywhere: "The others barely understood Spark... We're always a year or two ahead of everyone else." The founders still had the technical depth. What they gave up was the org chart.
The proof came in the spring of 2017, in a conversation with his CFO. Unusual Ventures later asked him for the most exciting minute of his entrepreneurial life, and this was his answer. The company was on track for $30 million in revenue that year, up from $10 million the year before. "We'd thought we were failing up until we heard that," he said. For the glass-half-empty founder, the company had gone from failing to working in the time it took to read one forecast.
Sixty Thousand Salespeople
Microsoft in 2016 was not obviously a startup's friend. It had spent a generation as the company young software firms feared most. Its sales force numbered roughly 60,000, and its Azure cloud had a gap in its portfolio exactly where Spark sat.
Ghodsi's deal with Microsoft is the most instructive transaction in the company's history, partly because it nearly failed so many times. The structure was a straightforward trade: Databricks had the product and Microsoft had the distribution. The cost to Databricks was severe. Integrating deeply enough to become a first-party Azure service meant giving up, in Ghodsi's words, "12 months of our roadmap," a full year of engineering for a company that had only just started to work.
So he demanded something in return that would hurt Microsoft too. The Databricks team insisted on a large pre-commitment, enough money that someone inside Microsoft would personally care if the partnership failed. In a company of that size, a deal without financial pain on both sides tends to get lost in the bureaucracy. The negotiations, by his count, "died" around ten times. Each time it was revived. In November 2017 Microsoft announced Azure Databricks as a first-party service. In February 2019 Microsoft joined the Series E, the $250 million round that valued Databricks at $2.75 billion, and Horowitz told CNBC that the old monster of Redmond was now "an outstandingly good partner."
The Microsoft relationship lasted. In July 2026 the two companies extended their integration into the 2030s. Databricks would use Azure and Microsoft's Cobalt processors, and Microsoft would build Databricks technology, including the Genie assistant, into Microsoft 365, Teams and Copilot. The year of roadmap given up in 2017 was still paying back nine years later.
Around this time came a decision that does not appear in the press releases. Databricks received an acquisition offer at six times its previous valuation. Plenty of founders would have taken it gladly, and Ghodsi admits he was tempted: "I was like, maybe they're right, maybe we should just sell." He called Horowitz, whose firm would have done very well from the sale. What Horowitz told him, as Ghodsi remembers it, is the closest thing to a moral in the founding story:
— Ali Ghodsi, recalling Ben Horowitz's counsel, on a16z's Boss Talk, 2025He said, hey, you can do whatever you want. You can sell, you're going to make a lot of money, and you'll be super successful in life. But, you know, if you're like me, you're going to look back the rest of your life thinking, you know, I missed that one shot.
Horowitz then turned the question around: So do you want to live with that or do you want to just have the money? Ghodsi did not sell.
The ambition kept growing past what seemed sensible. At one point a suggestion was made that Databricks belonged with the FANG companies, the era's most elite employers. It sounded absurd for a company that had recently been earning pub-level revenue. Ghodsi turned it into a compensation model, paying engineers at the 95th percentile so the company could hire from the same pool as the giants. In 2019, after the $400 million Series F valued Databricks at $6.2 billion, Horowitz predicted it would reach $100 billion. In August 2025 it raised $1 billion at a valuation above $100 billion.
Ghodsi is not sentimental about how the company got there. When he talks about the founding years he talks about luck: 2012 might have been too early, 2014 too late, and the 2013 start nearly died in a frozen market for Series C rounds. "There's a lot of randomness," he says. He regards survivors' certainty with suspicion, including his own.
The Lake and the Warehouse
For most of the 2010s the corporate data world had a kind of class system. At the top was the data warehouse, structured, governed, fast, expensive and proprietary, where business intelligence ran and executives got their dashboards. Below it was the data lake, cheap object storage full of raw files, logs, images and text, where data scientists worked and governance hardly existed. Companies paid for both, copied data between them, maintained pipelines to keep them in sync, and drew security boundaries across the gap.
The warehouse's newest and most celebrated form was Snowflake. Snowflake went public in 2020 and for several years earned roughly twice what Databricks did. Ghodsi has described a four-year effort to get the better of it. The weapon was a category he created, and the idea behind it fits an old line of his. Asked by Unusual Ventures how he would fight a giant, he joked about a slingshot and then answered seriously: "I'd analyze my competitors' weaknesses, figure out my strength, then apply my strength to that weakness. I live by this every day."
The weakness he found was not in Snowflake's product, which was excellent. It was in the business model: the warehouse held the customer's data in a proprietary format and charged rent for keeping it there. Databricks' strength was openness, the same openness that had nearly sunk it. The answer was the "lakehouse," which kept data in the customer's own cloud storage in open formats and added on top the transactions, governance and query performance that had been reserved for warehouses. Databricks built Delta Lake, an open-source layer adding ACID transactions to data lakes, and then a SQL product meant to run business intelligence directly on the lake. In January 2021, Michael Armbrust, Ghodsi, Xin and Zaharia set out the architecture in a paper at the Conference on Innovative Data Systems Research. It was fitting that a professor's company announced its strategic attack in an academic paper.
Category creation was hard inside the company as well as outside, he has said. Introducing a new word into an industry's vocabulary takes, in his phrase, "a strong point of view on what's going on in the market," and he admits it required pushing through "a lot of self doubt" against opinions that "did have validity at the time." He had done this before. In the early years, he recalled, most people "called us crazy" for staking everything on the cloud when on-premises vendors were expected to crush the company. Being right about the long trend has repeatedly meant being alone with it for a while.
The fight moved into open table formats. Apache Iceberg was started in 2017 by Ryan Blue and Dan Weeks, engineers who had dealt with the data lake problem at scale. With Jason Reid they founded Tabular in 2021 and raised $37 million through a Series B. By early 2024 Iceberg had become dangerous enough that Snowflake's CFO warned investors of "revenue headwinds" arising, among other things, from "the expectation that some of our customers will leverage Iceberg tables for their storage." Snowflake had a new CEO, Sridhar Ramaswamy, and a new open catalog called Polaris. In June 2024, after what a source described to Fortune as a bidding war with Snowflake, Databricks agreed to buy Tabular for more than $1 billion.
Ghodsi did not pretend to neutrality. "I think that this lakehouse approach, where you own your own data is obviously super disruptive," he told Fortune. "To be honest with you, I think they all see it as threatening to them, because they're going to lose revenues." That same June, Databricks open-sourced Unity Catalog, its governance layer. The pattern was by now established: give away the layer your competitor charges for, and make your money on the layer above it. In June 2025 Databricks disclosed that its SQL warehousing product was heading to a $1 billion revenue run rate. Its open lake had grown into a warehouse business of its own.
Wilderness Money
The AI wave did not take Ghodsi by surprise, at least not in his own account. He says the company was "focused heavily on AI at a time when that market barely existed," back when "a lot of folks were convinced the puck was headed in a different direction." In 2017 he described the mission as bringing artificial intelligence to "the rest of the Fortune 2000." Still, ChatGPT arrived in late 2022 like weather, and his response in the year that followed was quick, expensive and occasionally embarrassing in ways he has since acknowledged with unusual frankness.
In March 2023 Databricks released Dolly, an open-source language model with six billion parameters, named after the cloned sheep. The company said it showed "ChatGPT-like instruction following ability" and did not publish formal benchmarks backing the claim. It was the work of a company that wanted to show it belonged in the race. Three months later Databricks paid $1.3 billion for MosaicML, a two-year-old generative AI startup. A San Francisco paper calculated the price at about $21 million per employee. In Ghodsi's way of acquiring, the people came before everything else.
In March 2024 came DBRX, a mixture-of-experts model with 136 billion parameters, about 36 billion active for any given output. Databricks said it was competitive with the best open models of the time and that training it cost $10 million. That figure did not survive scrutiny. In June 2026 Ghodsi told Eric Newcomer that the company had actually spent about $20 million on the training cluster, and only $4 million on the run that produced the model. The rest, roughly $16 million, went to idle time, false starts and the mistakes of inexperience. "First of all the cluster is not running all the time — it's idle," he said. "You don't just yolo and say hey let's run the thing and then it comes out and it's like after a month of training it's like oh it was shit. So you do smaller training runs — so you work your way up to it."
The confession tells you more than the press release did. The habit of correcting your own official record years later, in plain language, belongs to someone who still carries some of the academic's obligation to publish errata.
What Ghodsi took from the model-building episode was a conclusion that put him at odds with the industry he was spending in. The models themselves, he decided, are a commodity. "Like, you can get gas from this gas station," he said on the BG2 podcast. "You can get gas from that gas station. It doesn't matter. Just compare price." He signed partnerships with all of them: $100 million with Anthropic in March 2025, a four-year agreement with Google for Gemini, $100 million with OpenAI that September. Databricks became a neutral marketplace for intelligence, and the value it held onto was the thing no model vendor had, which was the customer's data. "The dirty secret of AI," he told Fivetran's George Fraser, a fellow East Bay data founder, "is that the hardest part is the data."
He has also said, cheerfully, that artificial general intelligence is already here. Under the definition the AMPLab used around 2009, he argues, the systems now available would qualify, and the industry moved the goalposts once they were reached. In a Stanford lecture in the spring of 2026 he placed the real obstacle elsewhere, in what he called the context gap: the institutional knowledge locked in the heads of a few long-serving employees. That gap explains why so many enterprise AI pilots fail. To illustrate the scale of change, he cited Paul David's work on the electric dynamo, which took about forty years to show up as productivity gains, because the factories first had to be rebuilt around the new power source.
He also tells, with evident pleasure at the absurdity, about a meeting he attended with four humans and six AI note-takers. It looked, he said, like the opening scene of a movie in which the machines begin taking over the rooms where people meet.
The number he repeats most is the one that makes his commodity argument concrete. "Over 80% of the databases that are being launched on Databricks are not being launched by humans, but by AI agents," he said at Fortune's Brainstorm AI in December 2025. Developers "vibe coding" software in natural language generate applications, the applications need databases, and the databases appear, mostly on Databricks, mostly with no human ever choosing a vendor. In May 2025 the company bought Neon, a serverless Postgres startup, for about $1 billion, and that June launched Lakebase, a transactional database built for agents. Oracle's market, which Ghodsi said had stayed largely "the same for 40 years," now had a challenger whose customers were mostly software.
Gravity Upside Down
The money got large enough that its size was the story. In December 2024 Databricks announced a $10 billion round at a $62 billion valuation, one of the largest venture financings ever. The demand overwhelmed the company's back office. "I probably underestimated the sheer number of investors and the amount of interest," Ghodsi told Fortune. "We had close to $18, $19 billion of interest in this round." Thrive Capital led. One of its partners, Vince Hankes, wrote that it had become "a bit of a joke at Thrive" that so many of the firm's people had "a founder-crush on Ali." Insight's John Wolff said the company was approaching "escape velocity."
Then came $1 billion at over $100 billion in August 2025, more than $4 billion at $134 billion that December, and $5 billion at $190 billion in August 2026. Revenue kept pace closely enough to make the valuations look less like hype: a $4.8 billion run rate growing 55% in late 2025, $5.4 billion growing more than 65% by February 2026, and $6.9 billion growing 80% by June, with positive free cash flow over the prior twelve months.
What makes Ghodsi unusual in this market is not that he raised so much. It is what he declined to raise and why. The company could have gone public at the height of the AI frenzy, and he says so directly: "If I wanted to have a crazy, crazy valuation, we would have gone public in the last 12 months. I'm not saying it's guaranteed, but there's a decent chance we would have gotten retail investors excited." He then gives the reason: "But what goes up comes down." His stated goal was "a fair valuation that we can continue growing into over the next few years, every time we fundraise." A child who once watched a whole city go dark wants a price he can defend after the next collapse.
He has seen collapse in markets too. "In 2021, most of my peers, CEOs, they were like we got to IPO," he said. By 2022 those peers were cutting costs, and Databricks, still private, hired thousands of people. He has run what he calls "sky is falling" exercises, formal downside-scenario planning, since well before anyone required it, and he credits them with getting the company through the pandemic. Even at $134 billion he was openly calling much of the market a bubble: "Companies that are worth, you know, billions of dollars with zero revenue, that's clearly a bubble, right, and it's, like, insane." He expected the circular financing among AI companies to get "much, much, much worse" before it corrected. He compared the herd behavior of the industry to children's soccer: "all of them run towards the same ball."
He also says, with no apparent irony, that Databricks has "a shot to be a trillion-dollar company." He sets out the route as a "trifecta": transactional databases for agents, Agent Bricks for building agents on proprietary data, and applications vibe-coded on top of both. "I think if we just did that," he said of the database business alone, "we could maybe get all the way to a trillion." The pessimist who builds the downside model is the same man who names the trillion-dollar target, and he sees no conflict. Expecting the house to come down is exactly what lets him keep building.
On the IPO, a decade of questioning has produced a carefully rehearsed noncommittal answer. "I would not rule it out, but I also wouldn't take it to the bank," he told Fortune in December 2025. "We're ready. It's not an 'if,' it's just a 'when.'" At Brainstorm AI: "We will go public at some point. But to us, it's not a really big deal." In 2016 the next round was a reason to get through one more year. In 2026 the public offering is the next thing put off by one more year. The motif keeps repeating.
His working week is easy to locate. At the June 2026 Data + AI Summit, the conference that started as Spark Summit, he gave keynotes alongside OpenAI's Greg Brockman and his co-founder Arsalan Tavakoli-Shiraji. The company announced a real-time analytics engine, an architecture to combine transactional and analytical processing on one copy of data, and an agentic "co-worker" called Genie One. Databricks had about 10,000 employees by then. He had committed $1 billion to downtown San Francisco when the city's commercial core was emptying out.
One record has stayed almost unchanged through all of it. On the website of Berkeley's Department of Electrical Engineering and Computer Sciences there is a faculty page in the department's sparse academic template. Its title reads Adjunct Associate Professor. Its research areas are listed as database management systems and operating systems and networking. Under education it gives the 2006 PhD from KTH and the two degrees from Mid-Sweden University. The biography is two sentences long and says nothing about a company. At the top, under contact information, is a single line: aligh@berkeley.edu.
Part IIThe Playbook
The principles below are drawn from Ghodsi's decisions rather than his slogans, although he has been generous with slogans too. Most of them grow out of three facts in his biography: a childhood that taught him collapse can come overnight, an academic training that made him distrust intuition he could not test, and a first CEO job he took while not knowing how to do most of it. None of these is the standard advice about grit or vision. Several contradict popular founder wisdom outright, and Ghodsi would want them to.
The Ghodsi Playbook
- 1.Starve everything but the bottleneck.
- 2.Your open-source project is your competitor.
- 3.Declare yourself zero, then hire your teachers.
- 4.Make restructuring an ego test, and take it first.
- 5.Make the partner bleed.
- 6.Add time to every trade-off.
- 7.Rehearse the collapse before it comes.
- 8.Price the company you can grow into.
- 9.Buy people, then architecture, then revenue.
- 10.Attack the business model, not the product.
- 11.Commoditize the glamorous layer; own the context.
Principle 1
Starve everything but the bottleneck
The most quoted Ghodsi principle is also the most often misread. People hear "focus" and think of the standard advice to prioritize. What he described is closer to deliberate imbalance: "an orders-of-magnitude disproportionate amount of attention" on a single constraint, "almost to the point of going overboard." He builds in the expectation of partial failure. Even with that level of focus, he says, "you probably will not fully unclog it," because everything else keeps landing on the CEO's desk. Prioritizing makes a ranked list. His method turns the list into a single line.
In 2015 and 2016 the diagnosis was obvious once stated: open-source success, check; commercial success, zero. What was unusual was how fully he acted on it. He spent the next stretch replacing the executive team, learning sales from scratch, rebuilding hiring, and trading a full year of engineering for Microsoft's distribution. Each of those was a commercial-engine decision. The technical lead the founders held, "a year or two ahead of everyone else," was left to look after itself for a while, because it could.
The key is the time horizon he attached: two or three years on one problem. That is far longer than most leaders can keep up an obsession, and long enough for it to compound.
Tactic: Name the one constraint that would make the most other problems irrelevant if solved, then commit a share of your calendar to it that feels embarrassing, for at least two years.
Principle 2
Your open-source project is your competitor
Ghodsi's early instinct, to win adoption first and charge later, was right. It produced the conditions for the company's survival: customers who refused free software and asked to pay. The less comfortable lesson came next. Adoption is not a business model. Once Spark was everywhere, every cloud vendor could distribute it and every buyer could ask why they shouldn't just download it. "Your biggest enemy is your open source project."
His fix had three parts, and none of them involved hiding the open-source work. He added proprietary value above the open core: managed cloud operation, performance, governance, and eventually an entire platform. He moved to full SaaS, after admitting that the hybrid architectures had been a mistake and that "most shortcuts don't pay off." And he kept open-sourcing, often aggressively, as with Delta Lake and Unity Catalog, but chose which layers to open according to where the money was not.
The broader lesson is about the gap between popularity and capturing value. A project that people love and can copy in full gives you reach. It does not give you a business.
Tactic: For every component you open-source, write down specifically what a customer would pay you for that they cannot download, and if the list is short, treat the project's popularity as a threat rather than a moat.
Principle 3
Declare yourself zero, then hire your teachers
Technical founders usually fail at go-to-market in one of two ways. They hand it entirely to hired executives they cannot evaluate, or they apply engineering logic to sales and get it wrong. Ghodsi took a third path. He admitted total ignorance, "zero," and then went about learning, interviewing the best practitioners he could find, comparing their contradictory playbooks, and hiring people good enough to teach him.
This came with explicit rejections of conventional advice. He disagrees with "Hire your marketing leader before your sales leader." In his sequence, a creative sales leader sits with the founders and works out price points and buyers before anyone tells the story at scale. He dismisses "Trust your gut" and calls anything said about total addressable market "all fiction. Always." His operating texts are books: Who for hiring, Monetizing Innovation for pricing ("Helped me design pricing at Databricks"), Decisive for avoiding cognitive traps, and Good Strategy Bad Strategy, which he named as his favorite.
The method extends to staying in touch with ground truth. With Horowitz, he argues that CEOs must "fly low and fast," because the real knowledge sits with customers and individual contributors, not with the executive staff. Asked whether the Monday meeting was theater that a Google doc could replace, he answered with a sarcastic outburst about families ("If there's something you want to teach your kids, put it in the goddamn Google doc"). Trust, in his view, is built in person.
Tactic: For any function you do not understand, interview at least five of its best practitioners before hiring a leader, and write down where their advice conflicts, because the contradictions show you what you need to decide.
Principle 4
Make restructuring an ego test, and take it first
Within roughly a year of becoming CEO, Ghodsi replaced nearly the entire executive team with experienced operators and offered each displaced leader the chance to stay, reporting to their replacement. Five of the seven accepted. "If people were smart enough, they put their egos aside."
The offer works only because of the person making it. Ghodsi himself had accepted a CEO trial run without a CEO's salary, under a board that had been interviewing his potential replacements. The founders had picked him over an outside executive, which was a choice to put their hierarchy behind the company's needs. When he asked his executives to swallow their pride, he was asking them to do what he had already done.
The less obvious benefit is keeping knowledge. The displaced founders and early executives were the people who "barely understood Spark" in reverse: they understood it better than anyone. Firing them would have cut away the company's technical lead. Demoting them kept it.
Tactic: When you bring in senior talent above existing leaders, give the incumbents a real option to stay in a reporting role, and make sure your own recent history shows you would accept the same terms.
Principle 5
Make the partner bleed
Big-company partnerships fail quietly. Nobody cancels them. Nobody gets around to caring. Ghodsi's insight in the Microsoft negotiation was that a partnership needs pain on both sides to survive the bureaucracy of a 60,000-person sales force. Databricks gave up "12 months of our roadmap." In return, it required a large pre-commitment from Microsoft, so that "someone inside Microsoft would care if it failed."
The deal died about ten times and came back each time, which says something about what the pre-commit did. Once money was committed, a Microsoft employee had a career reason to revive it. The arrangement became a first-party Azure service in November 2017, an equity investment in 2019, and a partnership extended into the 2030s in 2026.
The general point is about incentives. A partner's enthusiasm in the first meeting tells you very little. The useful question is who on their side loses something if it fails.
Tactic: In any strategic partnership, find the specific person on the other side whose budget, quota or reputation is tied to your success, and if no such person exists, renegotiate until one does.
Principle 6
Add time to every trade-off
Ghodsi's decision rule is short and practical. Most hard decisions, he argues, trade a short-term gain against a long-term loss or the reverse, and the overlooked variable is time. "One way I do that is to think to myself: one year from now, will one of the options become 'more right?' If so, I pick that and I have time on my side."
The rule is consistent with the rest of his life to an almost uncanny degree. He stayed at Berkeley one year at a time. He chose the cloud when on-premises looked safer. He turned down an offer at six times the company's valuation after Horowitz recast the choice as a lifetime of wondering. He has put off the IPO one more year, more than once. In each case the near-term option paid better in money, comfort or safety, and the longer-term option became more right as time passed.
The rule is not reckless. It is a deliberate preference for choices that improve as time passes, made by someone who distrusts gut feeling and wants a test he can repeat.
Tactic: Before any major decision, write down which option you expect to look better in twelve months and why, and when the answer is clear, let it override the option that feels better today.
Principle 7
Rehearse the collapse before it comes
"One day, everything can be great; the next, everything can have collapsed." Ghodsi will not call his childhood a source of resilience, but he has made it a management practice. His "sky is falling" exercises, formal downside-scenario planning, require teams to model the bad case before it arrives. He credits them with steadying the company when the pandemic hit.
The practice explains how he can sound like two different people. He describes himself as paranoid about runway and also says Databricks has a shot at a trillion-dollar valuation. These are compatible. Planning for the downside is what makes the large bet survivable, because if you already know how much runway you have when revenue falls, you can afford to keep investing when it doesn't.
The 2022 contrast showed the payoff. Peers who rushed to IPO in 2021 spent the following year cutting staff. Databricks, still private and still planning for disaster, "was able to hire thousands of people."
Tactic: Every year, have each leader present a funded plan for a scenario in which revenue misses badly, and approve aggressive growth spending only after that plan is credible.
Principle 8
Price the company you can grow into
Most founders treat valuation as a score. Ghodsi treats it as a liability that will come due. "If I wanted to have a crazy, crazy valuation, we would have gone public in the last 12 months," he said in December 2025. "What goes up comes down." His stated goal is "a fair valuation that we can continue growing into over the next few years, every time we fundraise."
The timeline shows a company that raised a great deal but rarely got ahead of its revenue. Valuations rose with run rates, and the largest step-ups came alongside growth that accelerated rather than slowed.
$
The Climb, Round by Round
Valuations and disclosed run rates from public reporting. Each step was matched by revenue the company could point to.
2013
$13.9M Series A from Andreessen Horowitz. The company aims to replace MapReduce.
2015
US GAAP revenue of about $1.5 million. The board looks for a new CEO.
2017
On track for $30M in revenue, up from $10M. $140M Series D. Azure Databricks launches.
2019
$250M at $2.75B (Feb), then $400M at $6.2B (Oct). Horowitz predicts $100B.
2021
$1B at $28B with $425M ARR; $1.6B at $38B with more than $600M ARR.
2024
$10B at $62B, with roughly $18–19B of investor interest.
2025
More than $100B (Aug); more than $4B at $134B with a $4.8B run rate (Dec).
2026
$5B at $190B with a run rate above $7B. Still private.
Staying private is the other half of the principle. Ghodsi says plainly that the IPO is "not an 'if,' it's just a 'when,'" but he will not give up the freedom to invest through a correction: "If there's a major correction, and you suddenly have to start tightening everything and produce massive EBITDA, you can't spend as much time on massive opportunities for the future."
Tactic: Before accepting a valuation, model the revenue you would need in three years to justify it at a sober multiple, and if you can't defend that number, take less.
Principle 9
Buy people, then architecture, then revenue
Corporate development usually begins with the spreadsheet. Ghodsi starts at the other end. Databricks evaluates acquisitions on people first, whether the founders will build for five years. Then it looks at product integration, whether the code can become part of one product. Financials come last. Buying revenue first, he warns, may produce two years of growth but leaves "a bag of crap that doesn't work together." Horowitz adds the hidden cost: each separate architecture needs more specialists and support systems, and sales efficiency suffers.
M
The Acquisition Ledger
Major Databricks deals sorted by what they mostly bought.
| Target | Year | Price | What was really bought |
|---|---|---|---|
| MosaicML | 2023 | $1.3B | People Generative AI team, about $21M per employee |
| Arcion | 2023 | $100M | Architecture Data replication into the lakehouse |
| Tabular | 2024 | >$1B | Competitive Iceberg's creators, won in a bidding war with Snowflake |
| Neon | 2025 | ~$1B | Architecture Serverless Postgres that became Lakebase |
| Panther Labs | 2026 | Undisclosed | Expansion AI security operations for Lakewatch |
The rule also explains the hiring strategy. When Ghodsi called the AI talent market "a war" and "insane," he was describing his main acquisition logic applied one engineer at a time. He thinks many reported $100 million offers are exaggerated, and he responds with mentorship, learning and real ownership rather than trying to win an auction.
Tactic: Rank every acquisition target on founder commitment and code compatibility before anyone opens the revenue model, and walk away from deals that pass only on financials.
Principle 10
Attack the business model, not the product
Snowflake had about twice Databricks' revenue and a product customers loved. A direct product-against-product fight would have meant attacking where Snowflake was strongest. Ghodsi's stated method was to "analyze my competitors' weaknesses, figure out my strength, then apply my strength to that weakness." The weakness was proprietary storage, the warehouse's rent on data customers believed they owned. The strength was openness.
The lakehouse turned that into a category, the CIDR paper gave it academic standing, and open table formats made it something customers could act on. When Snowflake's own CFO warned investors of "revenue headwinds" from customers moving storage into Iceberg tables, the attack had reached the competitor's earnings call. Buying Tabular and open-sourcing Unity Catalog closed the loop: Databricks gave away what its rival charged for and sold what came above it.
⇄
Conventional Wisdom vs. the Ghodsi Approach
| Conventional wisdom | Ghodsi's approach |
|---|---|
| Hire marketing before sales. | Hire a creative sales leader first to find price and buyer. |
| Trust your gut. | Distrust intuition; use decision frameworks and the one-year test. |
| Size the TAM. | "It's all fiction. Always." |
| Out-feature the incumbent. | Open-source the layer the incumbent charges rent on. |
| IPO when the window is open. | Stay private while it preserves the ability to invest through a correction. |
| Build or buy the best model. | Treat models as gasoline; own the data and context. |
Tactic: Write down how your largest competitor makes its margin, then ask what you could give away for free that would erode that margin without hurting yours.
Principle 11
Commoditize the glamorous layer; own the context
The DBRX episode cost about $20 million for a model whose real training run cost $4 million, and Ghodsi admitted as much years later. The lesson he drew is the most important strategic position Databricks holds in AI: the model is a gas station. Enterprises switch providers "based on price," so the durable value is in proprietary data and the business context that only an organization's own people have. "It's a commodity now to have AI that has general knowledge," he said. "It's very elusive to get AI that really works and understands that proprietary data that's inside enterprise."
So Databricks signed $100 million partnerships with Anthropic and OpenAI and a multiyear agreement with Google's Gemini. It became the neutral place where everyone's models meet a customer's governed data. The strategic advantage is that Databricks wins whichever model wins, and every improvement in the models raises the value of the context layer they all depend on. The figure of 80% of new databases launched by agents is the result: once software chooses its own infrastructure, the platform that holds the context is the default.
The lesson applies well outside AI. In any technology wave, the layer that gets the attention and the capital tends to become the commodity, and the less exciting layer underneath it is where profits end up.
Tactic: Identify the hottest layer in your market, plan as though it will be cheap and interchangeable within three years, and invest in the asset that becomes more valuable as that layer gets cheaper.
Free playbook
Get The Business Model Playbook
58 business models, one visual page each: how the money flows, the metrics that matter, and who runs it. Free when you join the Faster Than Normal email.
Free. No spam. Unsubscribe anytime.
Part IIIQuotes and Maxims
In their words
— Ali Ghodsi, CEO Insider, 2026Every time I faced a choice between the comfortable known path and the genuinely challenging unknown, I had picked the challenge and it had opened my horizons each time. So I took the CEO role as the hardest thing I could possibly do.
— Ali Ghodsi, Powered by Battery interview, 2017I was here for one year, and said, hey, this is really, really interesting. We've hit Moore's Wall, there's this great opportunity, and I'm not going to get another opportunity like this in my life again. So let me give it 12 more months.
— Ali Ghodsi, to Fortune, December 2025If I wanted to have a crazy, crazy valuation, we would have gone public in the last 12 months. I'm not saying it's guaranteed, but there's a decent chance we would have gotten retail investors excited. As an AI company, valuation could have gone ahead of itself very easily. But what goes up comes down.
— Ali Ghodsi, to Fortune, December 2025There's definitely too much of everybody doing the same thing in general…It's kind of like watching kids play soccer—all of them run towards the same ball. Then, there's a big collision in the middle, somebody kicks the ball somewhere else—and then all of them run that way.
— Ali Ghodsi, ZOOM Marketing Positioning Leader SpotlightOne way I do that is to think to myself: one year from now, will one of the options become "more right?" If so, I pick that and I have time on my side.
Maxims
- Calibrate, don't romanticize. Hardship does not make you stronger, but it can teach you that collapse is normal, and that knowledge keeps you calm.
- Overallocate to one constraint. Focus that feels excessive is usually only just enough to move the real bottleneck.
- Popularity is not a business. An open-source project people love and can copy for free is your biggest competitor.
- Start at zero. In any discipline you have never practiced, admit total ignorance and hire people who can teach you.
- Restructure with your own ego first. You can only ask leaders to report to their replacements if you have already taken a demotion of your own.
- Make partners bleed. A strategic deal lasts only when someone on the other side loses something if it fails.
- Let time break the tie. When two options seem equal, pick the one that will look more right in a year.
- Price for the correction. Take the valuation you can grow into, not the one a frenzy will pay.
- Buy builders before revenue. Acquired revenue without integrated architecture turns into a bag of parts that don't work together.
- Commoditize the shiny layer. Assume the hottest technology will become gasoline, and own what it has to run on.
