Pick One Bottleneck. Stay Uncomfortable With It for Years.

Long Strange Trip — Ali Ghodsi, Databricks, with Brian Halligan — Sequoia Capital, k7wPdCNfljQ

In 2015 Databricks had Apache Spark and almost no business. GAAP revenue came in around $1.5 million. Ali Ghodsi was hearing through the grapevine that the board was interviewing other CEOs. He applied for a Berkeley faculty job. He had never wanted the chair. Andreessen Horowitz, in his telling, was religious about founders — Ben Horowitz had been asked when he would bring in a “real CEO” and written about it — so they gave Ghodsi a trial without even a CEO salary. A year in, equity and pay went up. At an all-hands Horowitz said they hadn’t thought he would do it either. Years later, when Halligan asked Horowitz who the best CEO in the Valley was outside his own portfolio, he did not hesitate: Ali.

Ghodsi’s operating system is not a personality cult. It is a bottleneck. Companies fail, he said, because of one giant clog, maybe two. Identify it. Pour orders of magnitude of attention on it for a year, two, sometimes three. If you overdo it, good — you probably still will not unclog it. Everything else is the daily avalanche: hiring, drama, board, legal, missed revenue. That is not strategy. That is disaster landing on your knees.

Always the harder option

He grew up in Sweden programming, dreamed of Electronic Arts in LA, $80k floated in high school, parents said university. At university he learned he had only known a little programming. In 2000 an internship at Room 33 — Europe’s largest mobile startup, he said — offered him the number-two seat if he dropped out. He went back to school. The bubble burst. The same university offered a professorship without a PhD. He took a hard PhD at a top school instead, later became a professor there, then left that for a “measly” Berkeley postdoc because they would not take him as faculty. Pattern: take the thing that challenges you. CEO was the thing he knew nothing about. Academia he already knew how to play. Finite years on the planet.

First-time CEOs, he said, should expect a 100-hour-week learning curve, network with the best, read the books, then apply first principles instead of copying. Bruce Lee’s Jeet Kune Do: build your own playbook from ingredients so it is self-consistent. Then find the bottleneck. For Databricks it was obvious: open source check, tech check, no commercial engine. He handed engineering and product to a co-founder and ignored them for two years. He revamped the executive staff within a year to a year and a half. The culture had been anti-sales, product-led, AWS swipe-a-card. They had run a “zero-touch” experiment before he was CEO: never talk to a customer. Revenue had been rising; it flatlined for two quarters. At $1.5 million, a single AE crushing a $1–2 million quota would be the whole company. PLG was not working. Hire sellers.

What kind? He ran a thought experiment: look at companies where enterprise sales works. Are the best AEs PhDs? None at the top. Technical screening was the wrong test; Halligan added that a PhD is a disadvantage. They had been hiring for smarts. Ron Grisco knew what great looked like — Ghodsi gives him 99% of the credit on sales. Great sellers were professionally aggressive, high EQ, hard to ignore, and they mapped the power base: no one person buys a $100 million deal. Dave, an early AE, got a meeting a customer had refused; the exec told them to get Dave out of the building, he had emailed the boss. Ron had told him get in somehow. Keith took a meeting that went hostile, asked what her team was working on because everyone is comfortable talking about that, and left with a hug and an apology. Techniques for dealing with humans, Ghodsi said, have not changed in 30 years. Trust the pros.

Ron was a classic PTC/BMC meat-eating seller who had sold free FTP and made money on it, plus an engineering undergrad and a Stanford GSB master’s in engineering “for fun.” He had taken Cyclone 0 to ~$50 million ARR, then seen hundreds of millions after Axway, stayed 11 years. Builder of the car, not just a driver. Longevity meant Ghodsi could fight with him and he would not quit. He is still CRO. The engine Ghodsi cited: $1 million to $7 billion ARR. VPs of sales now churn every two or three years because companies scale faster than people have seen. That pairing is rare.

Hire before you need them. Study the enemy. Don’t copy.

By the time you know you need an exec, it is too late. Infinite time, best person; five or six years, phenomenal; a rush, junk. Hire ahead of the curve so you can increase false negatives and cut false positives. A bad exec takes a year to fire if you have guts, another search, gun-shy, compromise candidate, two and a half years gone. Be extremely picky. Skip phenomenal people who later succeed and hold a grudge. Most of his remaining staff have been there a decade. Two miss-hires come to mind; in postmortem the signs were in the interviews — a big fight with someone, he told himself it was none of his business — or he rushed. Back-door references: talk to all past managers. Maybe 10–20% of front-door references are truthful.

Snowflake had double the revenue and was growing faster. Consensus said they were the sales company, Databricks the product company. Ghodsi thought that was unfair once Ron and later Andy had built an army, but the math was real. Overtaking would take years. Study the enemy, apply strength to weakness, do not copy. Snowflake had disrupted hyperscaler warehouses with a great product. Weaknesses: proprietary format from Oracle roots (open source a small fraction), weak AI despite the claims, expensive TCO because winners extract margin. Databricks went at all three: open lakehouse, you own the data; AI roots since 2009, even though AI was not a big commercial thing in 2019–20; TCO typically a third. Not “rip out Snowflake.” Coexist: pull the ML-amenable workloads, move those to open format, account by account. The category that named the playbook was Lakehouse.

Internally it was controversial. Seasoned people said don’t: funny word, house on a lake, you are trying to talk to execs. Strategy firms surveyed and agreed, proposed bland names — Unify Data Science Platform. Sales agreed it was too technical. They marched anyway. Online ridicule, data-rapids jokes; any PR is good PR. Then the company went maniacal. A glowing article that omitted the word was a loss. A customer win that didn’t say lakehouse was a loss. Facebook ads converting on Spark — marketing showed the ROI, Spark sells — he killed them. “I don’t care. Get rid of the ads that work.” Comp plan multipliers and spiffs, not debates. If you are convicted, put it in the plan and the argument stops. It took the whole company, years.

They were already at multiple hundreds of millions of ARR in 2018–19. He was unhappy. He was painting public-market investors a vision and then truth-seeking: we are just selling Spark. He did not want the job enough to run a one-trick pony. Splunk was amazing and still one trick. Amazon, Google, Microsoft could not be defined that way. Take the big bets or get stuck. Public markets have heard too much earnings-call BS; they say show me the money. Databricks was cloud-only AI when on-prem was the business; five years later people asked how they did it so fast. Non-consensus and right, for four or five years. Conviction without staying power, Halligan later said, is a January diet.

Conflict aversion is disqualifying

People who have worked for him called him a killer, more aggressive than you think possible. Mild-mannered CS professor. Chip on the shoulder: immigrant in Sweden, switching schools, United States softened him. Number one terrible CEO trait, learned fast: conflict aversion. Wonderful humans. Worst possible in the chair. Everyone is pushing. If you cannot say no — hallway approvals, elevator jams — the company goes in every direction except the bottleneck. People do not understand nuance. Make it black and white. Cheat sheet. 12,000 employees otherwise cancel each other out like bouncing atoms.

Dunbar hit around 250, near the CEO transition. Before that the CEO knows everything; managers of managers are names. After, you cannot tap every shoulder. Structure, reviews, KPIs, so the ship points without you knowing every name. Directionally he agrees with Jack Dorsey’s “company as AI, two layers to the front line” and Brian Armstrong collapsing levels. Separate the org chart from how information actually moves. Twenty years ago they were the same tree. Now models are, in his definition, AGI: smarter than most people around you most of the time. He asks audiences; ~90% of hands. Moscone, 32,000 people, almost everyone. Humans hallucinate too. Enterprises use the same models as stupid chatbots and coding agents. Nobody raises a hand for hundreds of agentic coworkers proposing, updating, collaborating. The gap is enterprise context — meetings, email, decisions — captured as ontology. Databricks built Genie for that; internally it is magical. Customers buy the product and do not build the ontology the same way.

He is not bought in on 25 direct reports plus player-coach vibe-coding 80% of the day. Managers still do human work: unhappy people, career choices, unclear goals. You cannot do that for 25 and ship code. Unhappy employees who never meet the manager. He thinks it is BS.

Does he still code? Never stopped, weekends, third grade. November of last year he started committing to production to feel the gap. A connector in two days; the team took three quarters per person. They said his was a POC, not secure, not production. He wanted to know if they were right. They came back: compress three quarters to seven and a half months. He put in an FTE who would not take the BS. After three quarters: seven connectors in one quarter. Models got better, they said. Off the record: a quarter collecting customer feedback for a beautiful PRD stalled the pipeline; Salesforce/Workday/NetSuite setup they hated, now consultants; testing at the end, now up front with AI; seven people, seven connectors, so one person on parental leave does not freeze it. Process re-engineering. That is why he thinks humanity needs a decade at least to absorb AI — even a tech company of people who want it. He still codes, less than earlier this year.

You cannot Google-doc your marriage

A day is not one cadence. Bursts around board, strategy, the annual summit. Three buckets: the bottleneck only he will work because everyone else is running trains; the daily CEO sludge, compressed; product, 3,500 engineers, stay current end to end. Weekends and early mornings on the 10x unblock. A calendar packed 8 a.m. to 6 p.m. means he has been a slave to other people’s monkeys. He tells the CEO office: we did nothing for the company today.

Staff Mondays, 60–90 minutes, one or two annual goals — he wanted one, got two, used to be three. Wednesdays and Fridays, 8 a.m., 30 minutes, bottleneck or around the horn or the news, some people on the phone, some skip. Halligan asked if Monday is theater, why not a Google Doc. Ghodsi: we are talking about humans. Why meet your wife and kids? Put it in the goddamn Google Doc. Claude can reply. Natural gravity builds walls between sales, legal, marketing. If the e-staff never hang out, they prioritize their second teams. QBRs two to three days, product and GTM, regions and departments. Offsites at least quarterly, aimed at the top bottlenecks. A lot of trust: people there 10-plus years, he knows how each will screw up and how they will be great.

Why IPO? He thinks they would be worth more public right now; that answer varies. Last raise: they took five, interest around $20 billion, could have set price much higher, not optimizing for that. They already run a private NASDAQ: tenders across ~100 countries, tax, 12,000 employees plus alumni, maybe 15,000 people. At some point that is what public is for. Not during a crazy transition — companies going from perpetual licenses to SaaS go private, do it, come back. They are free-cash-flow break-even; they do not need Anthropic/OpenAI capital. Wait for calmer water. SAS apocalypse, then SAS is great, then all money in semis, then sell the semis: he does not want to live in that mood swing. Public markets are bad at revolutionary transitions. Give it a year.

If you are a normal conflict-averse Homo sapiens who became CEO: do not lie to yourself. Self-defense mechanisms invent excuses. Truth-seeking to the extreme cannot be conflict-averse. Going to the gym sucks for everyone. Eat vegetables. If you want to be an athlete who does not work out because your family doesn’t, you are not in the sport. Nobody likes conflict. Get over it or do not take the job.

Halligan’s close: don’t be a wuss. He over-indexed on being liked at HubSpot, Glassdoor prizes; he is not sure that correlates with long-term success. Jobs didn’t. Second acts need more conviction than founding ideas. Swing the pendulum all the way. Middle means people stick with the first act. One or two focus areas a year. Overindex on the bottleneck, because changing gets harder as you get bigger.