Ben Horowitz and Ali Ghodsi: How to Run a $100 Billion Business
Summary
- Databricks escaped the classic open-source trap by recognizing that Apache Spark’s popularity was not a business model. Downloads and Spark Summit proved demand, but customers could still ask, “Why can’t I just download the open source version?” After PLG stalled at roughly $3 million ARR, the company added proprietary differentiation, hired experienced commercial leadership and went all-in on enterprise sales.
- Ali Ghodsi’s operating system is aggressive self-education combined with direct access to ground truth. He advises founders to admit they are “zero,” interview the best practitioners, compare conflicting playbooks and hire people good enough to teach them. He and Ben Horowitz argue that CEOs must “fly low and fast,” because actual knowledge resides with customers and individual contributors—not neatly inside the executive staff or org chart.
- High intensity scales through leadership, organizational design and visible impact—not hours alone. Ali sets the tone by working nights and weekends and vets candidates through backchannel references, but explicitly rejects burnout as the objective. Ben’s sharper point: no motivational speech can overcome a “three-legged race” of dependencies where employees know extra effort will not change the outcome.
- The Microsoft partnership worked because a genuine product-for-distribution trade was reinforced by a painful commitment. Microsoft had a portfolio gap and roughly 60,000 sellers; Databricks had the product but would sacrifice “12 months of our roadmap” to integrate it. The team demanded a large pre-commit so someone inside Microsoft would care if it failed, then survived a deal that “died” around 10 times.
- Databricks evaluates acquisitions in the reverse order of conventional corporate development: people, product integration, then financials. Ali wants founders who will build for five years and code bases that can become one product; buying revenue first may create two years of growth but ultimately leaves “a bag of crap that doesn’t work together.” Ben argues the hidden casualty is sales efficiency, because every separate architecture creates more specialists, support systems and customer friction.
- A pivotal decision was rejecting an acquisition offer six times Databricks’ prior valuation. Ben acknowledged that selling would pay a16z handsomely, then framed the real cost as spending a lifetime wondering whether Ali had abandoned his “one shot.” The same ambition turned the seemingly absurd suggestion to add Databricks to FANG into a P95 engineering-compensation model—and preceded Ben’s 2019 prediction, at a $6 billion valuation, that the company would reach $100 billion.
- Even at Databricks’ scale, the AI talent market cannot be treated as a pure bidding contest. Ali believes many reported $100 million offers are exaggerated by CEOs with incentives to reset compensation expectations; his counterweight is mentorship, learning and real ownership. He contrasts smaller startups with Databricks’ scale, citing a $100 billion valuation and 10,000 employees. He also stresses luck: starting in 2012 might have been too early, 2014 too late, while the actual 2013 start barely survived a frozen Series C market—“there’s a lot of randomness.”
Deep dive
1. Spark’s success became Databricks’ commercial enemy
Ali took the CEO job in 2016 after two or three years inside the company, knowing that painful pivots would challenge its identity. Apache Spark had become a worldwide sensation, with surging downloads and a successful conference, but “your biggest enemy is your open source project” when Amazon and other cloud vendors can distribute it themselves.
The original strategy—make Spark enormous, then offer the best Spark—had not produced enough proprietary differentiation. Customers’ decisive objection remained: “Why can’t I just download the open source version?” Databricks therefore had to build software around Spark that enterprises would actually pay for, rather than treating community adoption as proof that monetization would follow.
Ben’s assessment is that Ali combines genuine technical depth with unusually fast commercial learning. He had run engineering, understood product strategy in detail, then caught up on go-to-market and business development with help from advisers such as John O’Farrell: “He learned everything so fast.”
What separates Ali further, in Ben’s view, is that “he doesn’t hesitate” and trusts his own inspection of a threat. Building a data warehouse looked like “a pretty big swing” from Databricks’ position, but Ali was paranoid enough to investigate the competitive risk and confident enough to act rather than avoid evidence that rivals were trying to kill the company.
2. Becoming commercial requires admitting that your instincts are wrong
Ali’s learning model begins with an uncomfortable admission: “You know nothing, you’re zero.” He sought 30-minute meetings, breakfasts and dinners with reputedly excellent engineering and product leaders, asked “lots and lots and lots of dumb questions,” then compared their contradictory playbooks until he understood what great execution looked like.
The managerial multiplier, drawn from Andy Grove’s High Output Management, is hiring people so capable that the leader learns from them and then requiring that pattern recursively. Databricks initially violated it: virtually every leader was a computer-science PhD, “including sales,” because founders mistook someone who could talk engineering for someone qualified to sell.
Ali and Ben describe question-based feedback rather than verdicts. Ali might ask which displayed numbers yield a claimed 5% conversion; Ben says his own version was blunter—“I’m just trying to understand basic math”—while Ali’s better version is, “How do you think it’s going?” Both approaches force the owner to inspect the problem before the boss supplies a verdict.
Ali reframes criticism as optional assistance toward the recipient’s own goal: ignore it if you wish, but this change might improve your odds of winning the next job or project. Ben adds that candor must be continuous; annual-review surprises are offensive, while frequent correction prevents the common firing conversation in which an employee says, “I only got thumbs up all along.”
3. Intensity survives scale only when effort still changes outcomes
Ali’s first lever for preserving intensity across roughly 10,000 employees is “setting the tone at the top.” If employees know the CEO works at 9 or 10 p.m., at 2 a.m. and on weekends, the example becomes cultural behavior rather than a demand imposed by someone observing a different standard.
Hiring can screen for stamina, but Ali warns that candidates who advertise themselves as the hardest workers are often “the opposite.” Backchannel references are more revealing: ask not whether someone was great, but how consistently they “grind the midnight oil,” and former colleagues will usually answer candidly.
His position is explicitly not that every problem yields to more hours. Thresholds differ, work must remain sustainable and Databricks intervenes when a team’s work-life-balance scores deteriorate—while Ali jokes that groups scoring near 100% may have the opposite problem. “I don’t think you want a culture where people are burning out.”
Ben and Ali’s deeper constraint is organizational design: capable people work hard when they possess autonomy and can see their impact. Put them into a dependency-heavy “three-legged race,” and they rationally conclude effort will not matter. Leaders must make teams feel they are winning—or, during a crisis, show a “rock solid” path to winning that makes the requested sacrifice intelligible.
4. CEOs need a broad view and one leg driven deep into reality
Ali follows product launches and even progress reports in detail, answering small launch emails with congratulations because the acknowledgment itself motivates builders. Databricks’ old principle was “be a co-founder”: employees should behave as owners, surface ideas regardless of tenure and feel they can influence what the company creates.
Direct access has rules. A CEO can talk to the person closest to the work and listen for understanding, but issuing instructions around that person’s manager creates chaos; Ben’s preferred pattern is to diagnose at the front line, then send direction back through the chain of command.
Ben argues that executive summaries are structurally unreliable: leaders spin, lack detail and are themselves searching for bottlenecks. “All the knowledge in a company is with the individual contributors that are doing the work and the customers.” Hence a CEO who cannot “fly low and fast” will never receive enough truth to debug the organization.
Ben describes the attention model as a T: remain broad, then drive one leg extremely deep into the current priority. Ali says that priority might be engineering, a competitive sale or an HR failure. The org chart is “just a communication architecture,” so meeting cadence should be asymmetric; one executive may require daily contact and another quarterly—and if a staff member cannot do the job, “they can’t be fixed.”
5. Microsoft worked because both sides had something irreplaceable
Databricks pursued Microsoft because its roughly 60,000 sellers represented transformative distribution. Multiple supposed introductions to Satya Nadella disappeared into scheduling limbo, but after Ben and Satya discussed the strategic trade at a16z, one email propagated through Microsoft and produced around 25 messages from people suddenly clearing their calendars.
Timing opened the door. Hortonworks was reportedly threatening to withdraw comparable functionality unless Microsoft paid more, while Ali adds that Hortonworks was on-premises and in the cloud, creating another mismatch. Microsoft wanted to close a portfolio gap against AWS; Databricks wanted distribution. Ali’s rule for major partnerships is therefore unforgiving: “There has to be a give and get that actually is kind of commensurate.”
John O’Farrell advised Databricks to demand a sufficiently large pre-commit that “somebody in there is going to get fired if it doesn’t go well.” The negotiating device was to emphasize that integration would consume “12 months of our roadmap,” ask Microsoft for its sales forecast, then request only a portion of the large number Microsoft supplied.
Microsoft strategist Takeshi Numoto worried the small company might take a large payment and “get drunk off of it”; Ali had to prove his appetite would remain. The deal still died about 10 times, including near launch, so Ali repeatedly took the “nerd bird” to Seattle and influenced internal opponents face to face. He credits Microsoft with ultimately delivering as an exceptional long-term partner. Ali also says Satya’s growth-mindset push made the deal possible when it would have been impossible five years earlier.
6. Revenue-first acquisitions borrow growth by destroying integration
With acquisitions including Tabular, Neon and Mosaic, Ali starts with founders and teams: can they work together culturally and keep building for five years? Next comes the product experience, customer reaction and even the programming language, because incompatible build systems and code bases may not compile together. Revenue multiples and three-to-five-year financial plans come last.
Conventional corporate development reverses that order. It models revenue acceleration, assesses the product second and treats founders as people to pay off. The CEO departs, key employees follow, previously blocked subordinates get promoted, the acquirer inserts its own managers and eventually “the company is dead,” with neither the original talent nor a coherent platform remaining.
Ben’s go-to-market warning is load-bearing: separate architectures require separate sales engineers, post-sales teams, access-control models and customer education, driving field efficiency “through the floor.” Product integration is therefore not engineering polish; it determines whether the acquisition can be sold and supported as a Databricks product.
Ali concedes that financial engineering can work temporarily: the acquired revenue lifts year one, growth may improve again in year two and a favorable stock multiple can make the transaction look attractive. Long term, however, it becomes “a bag of crap that doesn’t work together,” eroding a brand built from every customer’s experience—damage Ben says marketing cannot reverse.
7. Talent quality is an acquisition constraint, not an integration detail
Ali repeatedly vetoed one otherwise attractive, successful company because he believed its employee base would dilute Databricks. Exceptional and very weak teams are relatively obvious; the costly judgment lies in the middle, where leaders must determine whether mediocre results reflect weak talent or merely missing capital, distribution and support.
That assessment cannot be “an Excel sheet exercise.” Databricks’ people interview the target’s people, spend time with them and ask whether the founders will become genuine co-founders.
Acquisition conduct also compounds as reputation. Future founders inspect whether earlier acquired leaders received influential roles or quit amid conflict, while every concession during the 20 or 30 days of definitive-agreement negotiations becomes precedent for the next transaction. Treating founders well is therefore both an integration choice and a competitive advantage in later auctions.
Ben’s cultural caution is that the acquirer does not automatically absorb the target; the reverse can happen. “Merger of equals” is especially dangerous because neither cultures nor talent pools are actually equal. Ali’s desired outcome is acquired engineers remaining the people who make the combined software great.
8. Implausibly large goals changed Databricks’ operating math
When a candidate wanted accelerated vesting in case Databricks was sold, Ben’s response was that Ali was severely underselling the opportunity: “We are Oracle in the cloud,” with the potential to become worth 10 times Oracle. Ali’s first reaction was simpler—“Ben’s crazy”—but both Ben and Marc Andreessen repeatedly forced the founders to examine why the ambition might be feasible.
During a 2017 financing discussion, Ali identified Google and the other FANG companies as the recruiting bottleneck. The response was to add Databricks to the acronym—“FANG DB.” Rather than leave it as motivational theater, the team calculated market capitalization per employee, translated its own available dilution into compensation capacity and concluded it could pay engineers at the P95 level.
Sarah Wang recalled joining a16z in 2019 and working on Databricks’ Series F at around a $6 billion valuation, when Ben predicted it would become a $100 billion company. The discussion treated the difference between $6 billion and $7 billion as immaterial relative to that long-term target while Databricks kept doing the work.
Ben’s underlying thesis was the rarity of simultaneously finding an enormous market and an entrepreneur capable of filling it. Venture portfolios more often contain a strong founder constrained by market size or a huge opportunity constrained by the founder; Databricks presented both.
9. Refusing an offer at six times valuation kept the opportunity open
A real acquisition offer arrived at six times Databricks’ prior valuation. Ali regrets telling his co-founders before deciding: people mentally stopped work, calculated their payouts and began reading his mood through daily gossip. Their excitement pushed him toward, “Maybe they’re right. Maybe we should just sell.”
Ben made his conflict unusually explicit. Selling would be excellent for a16z and return its investors’ money many times over, and he would support either decision. But his own experience at Loudcloud and Opsware told him how rarely anyone receives a company this promising: “I guarantee you you’ll never have an idea this good again as long as you live.”
The decisive question was not whether Ali would become rich, but whether he could tolerate wondering forever how far he might have taken it: “That was the one thing. I should have taken it all the way.” Ali ended the conversation and immediately decided, “We’re never doing this. We’re done. This is not happening.”
10. AI compensation headlines obscure what ambitious talent actually needs
Ali views the AI labor market as genuinely intense but partly performative. He doubts that $100 million offers are broadly real beyond exceptional cases such as Character.AI, and notes that CEOs benefit from publicizing $100 million poaching attempts: every lesser offer then feels insulting. Ben observes that Sam used the same narrative against Meta, effectively raising the reference price.
Ali contrasts smaller companies with Databricks’ scale, citing a $100 billion valuation and 10,000 employees. Databricks can pay significantly for selected talent, while smaller companies must sell the future value people can create together. Employers can also offer early-career people learning and impact—motives Ali believes remain more important than the headlines imply.
Intern questions have shifted from succeeding at Databricks to whether a three-month internship wastes the fleeting chance to build AGI or superintelligence. Ali tries to reduce that FOMO: “You have a few decades.” A CEO’s two minutes of attention, concrete career help or promise of mentorship can be immensely valuable to someone just out of school.
His favorite acquisition profile combines big-company experience with a failed startup: the former teaches process and bureaucracy, while the latter supplies grit, humility and respect for how difficult Databricks was to build. Maintaining relationships with departing employees matters because founders often return as more appreciative, effective “boomerang” hires.
11. Databricks survived because its pivot and the market arrived together
Ali’s counterfactual is brutally narrow. Had Databricks started in 2012, its crisis would have arrived in 2014, before cloud and AI demand could sustain it; starting in 2014 would have delayed repairs until 2017, leaving hyperscalers and competitors too far ahead. The actual 2013 timing came from waiting for Matei to finish his PhD thesis: “There’s a lot of randomness.”
Survival was still precarious. Redpoint reportedly stopped returning calls after a Series C handshake, so a16z and NEA—already the Series A and B leads—co-led the round when nobody else would. Revenue was thin beyond the conference, cash burn was high and Ali seriously considered taking a Berkeley professorship: perhaps Spark was the success and the founders simply “weren’t business guys.”
By 2015, PLG had decisively failed at roughly $3 million ARR, free Spark adoption was not converting and even hitting the company’s low plan would, in Ben’s words, still lead to bankruptcy. With little left to lose, Databricks entered 2016 by adding proprietary software, embracing B2B enterprise sales and hiring executives who had “seen the movie before.”
Hiring Ron was both uncomfortable and extraordinarily lucky: a classic salesperson among PhDs, he forced customer focus and became indispensable. The continuing contributions of the original founders multiplied that luck—Reynold pushed warehousing, Arsalan helped integrate go-to-market, Matei kept innovating and Patrick led major engineering groups—an unusually durable founding team around the commercial reinvention.