Summary

MacDonald calls autonomy existential for Uber — and its largest single standalone investment — but thinks the endgame favors the platform. “Autonomy is as bad as it’s ever gonna be today,” and he expects it eventually to become better in all use cases, though today it can be slower, have imperfect pickup points, and fail in some weather or geographies. With 300M trips/week across 75 countries and average fares of ~$2.50–4 in India and Brazil, “it’s gonna be decades” before AVs displace the majority of trips. He concedes Harry’s counter: dollar share could flip much sooner if AVs take San Francisco, LA, Washington, DC, Miami, New York, and Boston.
On Waymo vs. Tesla, his call is “more than two winners” — and “in the end, distribution wins.” China already has four or five AV players “which by the way will over time become eight or ten,” so he doesn’t see the rest of the world converging on one. His McDonald’s/Starbucks analogy is that owners of expensive fixed assets will have incentives to work with marketplaces to drive utilization, even with their own 1P apps.
The viral “blew through the AI budget in four months” story gets demystified — and the real ROI mechanism is headcount discipline, not line-item attribution. A pod of 30 top AI engineers paired with businesspeople has cut weekly pricing allocation from 15 hours to 2, forecasting from 8 to 2, and marketing QA from two weeks to two days — but freed hours refill with other work, so “the way companies ultimately have to extract AI efficiency… is just in your target setting, hold the constraints tighter.” His hedged five-year headcount call: “we could do everything we do today with less people,” but new businesses may need more.
He admits he was wrong — and “too short-termist” — on Uber One, which he now calls the most efficient long-term consumer lever Uber has. He used to put as much of each incremental dollar as possible into price or driver supply (“ridesharing at the end of the day is price, reliability, and safety”); membership’s compounding LTV, cross-sell into Eats, and churn resistance showed him he was wrong, though Uber is only within “spitting distance” of Amazon Prime/Costco-class programs.
The path from 200M to 500M monthly users is price — an answer he jokes IR will hate. A $35-per-direction UberX commute in NYC “is still a luxury product”; moving users from 6 to 25 transactions a month requires cheaper modes (trains, bikes, scooters) as you “deconstruct car ownership,” an asset that “sits idle 98% of the day.”
On agentic disaggregation, Uber will participate with frontier labs but refuses to be aggregated on price. He’s opposed API deals feeding real-time cars and prices into comparison apps — “I wanna be the front end” — and argues the “managed transaction” (lost items, pickup problems, driver interactions) is why the doomer scenario hasn’t played out; he thinks Brian Chesky “was right” that chat isn’t the interface for every booking.
The Uber China and delivery threads carry the competitive DNA: $52M/week burned on price subsidies in the final weeks before the DiDi deal, competing without WeChat “with one hand tied behind our back.” Today he’s personally running delivery — its leader departed, and he’s working “my day job and my night job.” The pending Delivery Hero deal would add geographic scale and local brands, while Uber remains not number one in US food delivery against DoorDash.

Deep dive

1. The 14-year operating formula: optimize visibly for Uber, and be willing to be wrong

Harry’s setup frames Uber at a $160B market cap, $52B in FY2025 revenue, more than a quarter-trillion dollars in annual gross bookings, and 200M monthly consumers. MacDonald is Uber’s longest-tenured active employee — “15 in May” — and gives his followership formula: filter every decision, big and small, through “what is the best thing for Uber,” so that even when you’re wrong, people trust the optimization target. Pair that with domain depth: rideshare has existed since 2009, he’s worked on it since 2012, and “I know that better than anyone in the world at this point.”
His signature epistemics, via Bezos: “If you wanna be right most of the time, you gotta change your mind a lot.” He volunteers he’s “wrong every single day.”
The confessed error — which Dara pre-seeded with Harry, insisting “say it’s from me” — was resisting membership. Running mobility, MacDonald constrained Uber One’s capital envelope because his gut said put as much of each dollar as possible into price or driver supply: “I probably was short-termist in my thinking there.”

2. Why membership is now the best dollar-in, dollar-out lever

The unit math as he runs it: the baseline metric is IGB (incremental gross bookings) per incentive dollar, and raw ROI is brutal because “we only make seven and a half percent of your dollar” — a 2:1 revenue return is still negative ROI, justified only by engagement and LTV growth.
Membership wins because it compounds: member cohorts ride more over time, consolidate mobility spend onto Uber, cross into Eats “instead of DoorDash or Deliveroo,” and churn less — “all these downstream long-term impacts that sort of multiply the value of that first dollar.” Price and promo dollars, by contrast, “dissipate faster.”
The gap to Prime/Costco: low consumer comprehension of the 5% mobility cash-back, and a structural handicap — Uber’s variable-cost model means “I don’t have a lot of free things to give away,” unlike a hotel with excess room nights. The sweet spot is “high perceived value, low cost” features, and Uber’s marketplace makes those scarce.

3. Innovator’s dilemma at ~$250B in gross bookings

Anything new must show “a path within a few years to multiple billions of dollars of GMV” to be significant — and MacDonald concedes flatly (“Yeah, totally”) that this constrains experimentation. The core “$225 billion blob… swallows up your organizational capacity to do anything else.”
Uber’s answer is Growth Bets: dedicated capacity — of a notional 2,000 mobility people, 100–150 on pre-product-market-fit ideas — because incubating “as 5% of your job” fails. Harry’s counter-model, Revolut’s Nik Storonsky (“the single best founder I’ve ever interviewed”): 26 parallel experiments, $2M each, weekly 20-minute check-ins, and decisions about whether to fund the next round. MacDonald loves the cadence — “operating on weeks, not months or quarters, is how a new business should run” — and the sing-for-your-supper discipline, since big-company incubations “just get fat on the resources.”
The offsetting advantage: 200 million monthly consumers of distribution — “which is the mother of invention,” Harry interjects; “100%,” MacDonald agrees — though internally every product fights over “how we spend our pixels.”

4. Getting to 500M users starts with a price problem

His IR-unfriendly answer: the constraint isn’t a price war, it’s that “the vast majority of transactions in transportation broadly happen at a price point that is way lower than our core products.” The NYC commuter UberX at $35 a direction “is still a luxury product.”
The target state: from 6 average monthly transactions to 25, via cheaper modes — trains on Uber in London, bikes, scooters — “because once you deconstruct car ownership, it’s not just about UberX.” The owned car is “the most inefficient asset that anyone owns… it sits idle 98% of the day,” and in “maybe not 5 years, but 15 or 20 years,” he thinks nobody owns a car or holds a license — like Harry, who runs the Uber-versus-car-ownership math daily.

5. Autonomy is existential — and the ATG divestiture was still right

Why existential: AVs are “a better product than our core product in many use cases,” those use cases grow, and he thinks they eventually become better in all use cases. Today he acknowledges they can be slower, pickup may not be at the door, and they may not work in all weather conditions, geographies, or pickup points. Over time, he expects them to become safer and provide a preferred in-car experience of privacy, work, and sleep. “Autonomy is as bad as it’s ever gonna be today… every single day it’s gonna get better.” It’s Uber’s largest single standalone investment: equity stakes, purchase commitments, autonomous infrastructure, and data-collection fleets.
On Harry’s I-told-you-so about stop-starting with Travis: MacDonald credits Travis’s foresight circa 2012–2014, when “the narrative was ahead of the reality by a lot,” but rejects rose-colored counterfactuals. At ATG’s divestiture, mobility “had lost 84% of our top line in three weeks” in COVID, the company was burning billions annually, and “we were trailing” — debatable whether the field or just Waymo. Focus turned the core into “cash-flowing machines,” and “almost any metric you pick from that point in time is up and to the right.” Still: could he snap his fingers and have ATG be a leading AV player? “Yeah, I think that would be a good thing for us.”

6. The AV forecast: tiny trip share, decades in emerging markets, but dollars concentrate in US cities

His refusal to predict the human/robotaxi mix in five years rests on three legs: the denominator is 300M trips/week against a few million AV trips a month (triple-digit monthly growth stays “a relatively tiny drop”); the human business is itself growing faster than the rest of the US in SF and LA; and Uber spans 75 countries where Brazil fares run “$3.50, 4 bucks USD” and India “$2.50, 3 bucks” — “it’s gonna be decades until the cost of autonomy compresses” to that labor cost, and those markets are the majority of trips.
Harry’s sharpening — which MacDonald accepts as “for sure the counter” — is that trip share understates dollar share: if AVs take majority share in SF, LA, Washington, DC, Miami, New York, and Boston, “that’s a big chunk of our bookings… that’s kinda the ultimate question.”
On the margin-squeeze bear case — few AV winners with leverage over Uber’s take rate — his delivery analogy: McDonald’s and Starbucks have 1P channels and billions in fixed assets, yet work with marketplaces because utilization rules. “Whether that’s a store or a car, I think that’s gonna be true… ultimately we have distribution, and ultimately they have expensive fixed assets that need utilization.” His categorical close: “In the end, distribution wins” — with the acknowledged tail risk that if only one player reaches the finish line, “that is a problem for us.”

7. China war stories: $52M a week, overlapping payrolls, and limited WeChat access

The endgame mechanics: investment was negotiating leverage, so in the final weeks before the DiDi deal “we were burning fifty-two million a week in China just on price subsidies.” Travis’s maxim — “we need to raise more money than all our competitors in the world combined” — broke down once SoftBank and other investors were also capitalizing competitors during the free-money era.
The story he calls nuts: when DiDi merged with Kuaidi and combined HR systems, they found roughly 200 of ~2,000 employees on both payrolls — an overlapping-payroll dynamic he “in a million years” couldn’t imagine in the US. Harry’s aside: “feels like a frontier AI lab employee.”
His sober verdict on the exit: Uber competed “with one hand tied behind our back” — at one point being unable to operate on WeChat was “like trying to compete in the US without email or a phone number” — and a US tech company winning China mobility was never geopolitically plausible. The silver medal was “a better outcome than the vast majority of Western companies”; the hardest part was the all-in Uber China team, many of whom Uber moved into global roles.

8. The AI budget headline, decoded — and how to actually budget for AI

Both viral headlines got distorted, he says: CTO Praveen’s budget line was about unpredictable, deliberately-driven usage growth (“you’re setting a budget number in November for a tool that’s growing vertical”), and his own hard-to-draw-a-direct-line-to-consumer-features comment “wasn’t insightful at all” — it just let AI skeptics and “fundamentalist AI evangelists” each confirm their priors. “There’s just nuance in the middle that is true.”
The tangible wins: a pod of 30 of Uber’s best AI engineers paired with business and G&A people, going process by process — weekly pricing allocation across thousands of markets from 15 hours to 2, finance reforecasting from 8 hours to 2, and marketing QA from two weeks to two days. Against Harry’s pushback, echoing his reference to Karp, that outside coding and support ROI “is still not material at best,” MacDonald’s concession-plus-mechanism is that freed hours refill with other presumably high-value work, so “you do have to be a bit top-down and belief-based about it” — and extract the gain by holding headcount targets tighter, not by tracing “I need two less operations analysts.”
Budgeting fix: pool headcount and compute — “your head count budget is X, our compute budget is Y. Just add X and Y together and then spend it as you see fit” — plus smart model routing, external providers (Harry mentions Fireworks), and cost dashboards. Harry’s sustained needle on usage/cost leaderboards — “do I want to be number one or bottom?… then it’s a bad leaderboard, man” — draws a partial concession: blunt metrics get gamed, but adoption leaderboards for domain-specific tools (every support agent on the AI assistant, not everyone on Cloud Code) are worth interrogating.

9. Frontier labs, agents, and why the managed transaction protects the front end

On Karp’s fox-in-the-henhouse warning about feeding data to labs that then compete: Uber’s lab experience “has been great” and it’s moving from experimentation to implementation, but the physical-world component makes direct competition difficult — “I don’t see OpenAI launching a ridesharing service… going around to tens of thousands of cities and getting locally licensed and putting boots on the ground.” Harry’s reference to Dario’s passion for last-mile delivery in Barcelona gets a watch-out response from MacDonald.
On agent-led disaggregation: he’d happily fulfill “get me my usual Uber” from ChatGPT or Claude, but “get me a car” price-comparison indifference is the threat. He’s opposed aggregation-app API deals — “I wanna be the front end… today we win that first look with two hundred million consumers” — and any lab partnership is a negotiation over how much data crosses the line, where the transaction starts, and who handles failure cases and responsibilities across the experience. “You shouldn’t picture the experience that goes right as the archetype.”
He defends Chesky’s roasted take that chat isn’t the right hotel-booking interface: “I think he was right. Some experiences are more visual, some experiences are more managed.” Harry’s rejoinder — the “Fortnitification of markets”: strip out transactional bookings and Airbnb shrinks to experiential travel — gets a measured “I think it’s a reasonable perspective,” not a rebuttal.

10. The Delivery Hero deal, the DoorDash fight, and running two jobs

The Delivery Hero logic: delivery is nearly as big as mobility, growing faster, and was more country-constrained; the pending deal would expand the footprint “in one fell swoop” and bring local brands with real mindshare — Argentina, Korea, and the Middle East — that “are not easily supplemented,” plus localized combined mobility-delivery offerings. He notes Uber gets “almost zero credit” for having a market-leading food business roughly the size of mobility. The transaction remained subject to regulatory and shareholder processes.
Since delivery’s leader left, he’s run it directly — “my day job and my night job,” with a scheduled evening shift — and admits Uber is not number one in the US: “operating from a position of strength gives you a nice tailwind, and so we’re having to play the challenger role, which we relish.” He also flags burnout risk: “you can only push above the red line for so long.”
On ex-Uberites saying “if Travis were here, we’d be number one in food”: he invokes the arena — sideline counterfactuals are cheap — and gives DoorDash its due: “Tony’s a tremendous entrepreneur… they move quickly, they’re aggressive, they take risk.” Whether Uber ever had a shot at buying DoorDash: “I honestly don’t know.” On Postmates, he thinks Uber “get[s] a harder rap on M&A than is deserved” — deals can advance capabilities, talent, and gap-awareness invisible from outside.

11. Travis, Dara, and the quickfire

From Travis: creative problem-solving as an organizational value — walking into any meeting, asking “a few pointed questions,” and in 15 minutes evolving the thinking of week-deep experts — and explaining the why behind answers, which “creates mini versions of yourself.” MacDonald operationalizes the latter through published principles.
From Dara, the quote he keeps: “Management comes from an org chart. Leadership comes from the heart” — low ego, first over the fence, pushing from a good place. On surviving both eras when most staffers picked a camp: in bad times leaving felt like abandoning teammates, in good times “we’re conquering the world” — plus “take what you can get from your leaders; you’re not gonna get everything from any one individual.”
Quickfire specifics: he’s become a longevity bull (“I actually do think we’re gonna solve all of human disease at some point”); he uses both labs but picks OpenAI, driven by voice — “probably 50 times more than anything else” as his most-used AI feature. Best advice ever: Rachel Whetstone’s commencement thesis, sent around 2015 or 2016, “always say yes” — bet on yourself, and even failure returns “a better version of yourself.”

From $18B to $300B: How Nikesh Arora Rebuilt Palo Alto Networks 179: The Distillation Storm: A Technology Race No One Is Discussing in Public