Uber CEO on AI, Autonomous Vehicles, and the Future of Transportation
Dara Khosrowshahi (Uber CEO) with Patrick O'Shaughnessy. The throughline: Uber is the one scaled aggregator sitting where digital AI meets the physical world, and its bet is to own AV *demand*, not build the driver. The load-bearing market signal is the ROI reckoning landing inside a real P&L: **'We blew through our AI budget in a quarter, you know, for the whole year, essentially. And it is forcing us to adjust'** — so Uber is now **metering headcount** as engineers get *'superhuman in terms of their output'*, and explicitly tiering models — *'we're using the more expensive models to explore... once we scale some of these experiences... we're look to bring in more efficient models that are more efficient on a token basis or open source.'* This is the same enterprise-token-discipline thesis the [20VC roundtable](/issues/2026-06-07) and last issue's [All-In ROI thread](/issues/2026-05-31) flagged — Dara is the live proof, and the model-commoditisation hedge spoken from the buyer's seat. On AVs: **over 30 partnerships** (Waymo, Nuro, Lucid, Nvidia, Wabi, Wayve, Pony) and the conviction there *'isn't going to be a single winner'* — the same many-models logic running through the whole cohort, now applied to digital drivers. The single hardest number: **AVs on Uber's network are '30% or more busy than... 1P AVs who aren't using our network'** in trips and revenue per vehicle per day — the entire 'why partner with us' pitch, and the ROI lever for anyone funding **$60-70k** robotaxis. He sizes AV as **'another trillion dollar marketplace'**, with AV-hardware cost falling *'usually 30 to 40% per generation'*. Marketplace economics: **$10B+ free cash flow** on *'well over 10 billion trips a year'* (*'we are not a high margin business'*), **Uber One at 50M members growing 50% YoY**, **Reserve at a $5B+ run rate** from nothing five-six years ago. Capital allocation: **'I prioritize growth, I prioritize innovation over buybacks. If you're building the company right, you'll do both'** — Amazon, not Apple. The pre-mortem is social licence, not technology: AI and AVs are *'unpopular... with the general public'*, and Uber will *'go at the pace that society is prepared for us to move.'* And the China tell for the AV supply chain: Chinese manufacturing *'in terms of quality and cost at this point is unrivaled.'*
Key points
- **The ROI reckoning, now inside a real P&L.** *'We blew through our AI budget in a quarter, you know, for the whole year, essentially. And it is forcing us to adjust. You know, we are going to meter headcount increases because to the extent that my engineers are getting much more efficient, their throughput is increasing.'* This is the precise mechanism [last issue's All-In ROI thread](/issues/2026-05-31) and [the Mercor 'token spend exceeds salaries' episode](/issues/2026-06-07) describe, spoken from the buyer's chair — token cost is now binding on hiring, not just on margin.
- **Model commoditisation, hedged from the demand side.** *'We're using the more expensive models to explore... And obviously, you know, these frontier models, whether it's a OpenAI model or a cloud model, they really are terrific and they're great to experiment against. Once we scale some of these experiences and interactions, we're look to bring in more efficient models that are more efficient on a token basis or open source and again, lower cost.'* Frontier for exploration, cheap/open-source for scale — the live enterprise playbook that erodes premium-model pricing, the bear case under [Anthropic's IPO/ARR run](/issues/2026-06-07).
- **The single hardest AV number: a 30% utilisation premium for being on the network.** *'AVs that are on our network are 30% or more busy than, let's say, 1P AVs who aren't using our network. That 30% in terms of trips per vehicle per day, in terms of revenue per vehicle per day can make a huge difference in terms of your ROI of investing in these expensive cars.'* This is the whole demand-aggregator thesis in one stat — the ROI argument that pulls AV builders onto Uber rather than going 1P (first-party, direct).
- **'There isn't going to be a single winner' in AVs — 30+ partnerships, mirroring the foundation-model market.** *'We've got over 30 partnerships now with incredible partners like a Waymo to a Neuro and Lucid to an Nvidia... Just like you're seeing in the foundation model space, there isn't going to be a single winner. There are going to be many players on the foundation models. There'll be open source, smaller models as well.'* The same fragmentation logic the whole cohort applies to LLMs, now mapped onto digital drivers — and the structural argument that the aggregator, not any single driver, captures the spread.
- **AV sized as 'another trillion dollar marketplace', with a hardware-cost curve.** *'We think it's another trillion dollar marketplace... as we see the cost of AV software come down, and usually we see the cost of hardware come down... usually 30 to 40% per generation. You know, we think the lucid mid size that they are building for us and Neuro together, it would be 60, $70,000 car.'* A dated, quantified TAM call plus a per-generation deflation rate — the demand-elasticity bet (cheaper rides -> more rides) that justifies the capex.
- **Marketplace economics: high cash flow, low margin.** *'We've got over $10 billion in free cash flow, which is great, but it's on well over 10 billion trips a year. So we are not a high margin business.'* The corrective to any 'Uber is now a cash machine' read — the FCF is real but thin per trip, which is exactly why AI efficiency and the membership flywheel matter to the model.
- **The membership flywheel: Uber One at 50M members, +50% YoY, framed as Netflix.** *'We have our Uber 1 membership program. 50 million members now growing 50% year on year... the way I look at it is we're kind of like Netflix. For the same price, you get more content than anyone else.'* Cross-platform (about **13% of Eats bookings now come from the mobility app**) is the structural edge over single-line rivals — and the first-year-loss / multi-year-profit shape is the Amazon-Prime model he explicitly copies.
- **Capital allocation: Amazon, not Apple.** *'I prioritize growth, I prioritize innovation over buybacks. If you're building the company right, you'll do both.'* Asked directly *'are you Amazon or are you Apple?'* he answers *'somewhere in between'* but ranks organic growth and AV commitments (*'making commits to tens of thousands of AV vehicles, those commitments will be financialized'*) ahead of buybacks — relevant given the prior insistence that Uber would defend its FCF and buyback.
- **The pre-mortem is social licence, not technology.** *'You've seen how powerful AI is, but at the same time how unpopular it is with the general public... if it's going to cost my cousin's job, that's... that doesn't feel that good. And I think the same will be true of avs.'* Uber will *'go at the pace that society is prepared for us to move. Otherwise there will be a backlash.'* This is the same ~30%-approval, legitimacy-deficit risk [Joe Lonsdale and the Pope/Frankenstein thread](/issues/2026-05-31) priced as the cohort's most underweighted tail — here it's the named failure mode for the AV rollout itself.
- **The China tell for the AV supply chain.** *'The Chinese capabilities in terms of manufacturing, both in terms of quality and cost at this point is unrivaled'* — Uber needs *'that kind of a low cost player in the western hemisphere'* (a 'Foxconn for AVs') and it *'It's being worked on, but we're not there yet.'* The hardware bottleneck under the trillion-dollar-marketplace call, and a read-through to the China-capability vs Western-onshoring tension in [last issue's software-defined-warfare thread](/issues/2026-05-31).
- **Reserve proves Uber can stretch a brand from on-demand to planned — the option value behind hotels.** *'Reserve now is over $5 billion run rate. It didn't exist, you know, five, six years ago. And clearly we were able to, if we gave a tangible benefit, in this case, higher reliability for higher price, clearly we were able to move both demand and supply.'* The $5B Reserve run-rate is the evidence base for the hotels/trains pre-commit expansion — content depth driving retention, the same compounding logic Dara applies everywhere.
Notable quotes
We blew through our AI budget in a quarter, you know, for the whole year, essentially. And it is forcing us to adjust. You know, we are going to meter headcount increases because to the extent that my engineers are getting much more efficient, their throughput is increasing.
And the way that I put it is we're using the more expensive models to explore, hey, let's try a new interaction here. And obviously, you know, these frontier models, whether it's a OpenAI model or a cloud model, they really are terrific and they're great to experiment against. Once we scale some of these experiences and interactions, we're look to bring in more efficient models that are more efficient on a token basis or open source and again, lower cost.
We've got over 30 partnerships now with incredible partners like a Waymo to a Neuro and Lucid to an Nvidia that is building not just compute and, and sensors, but also a software driver.
Just like you're seeing in the foundation model space, there isn't going to be a single winner. There are going to be many players on the foundation models. There'll be open source, smaller models as well.
And what we're seeing is consumers love the product and AVs that are on our network are 30% or more busy than, let's say, 1P AVs who aren't using our network. That 30% in terms of trips per vehicle per day, in terms of revenue per vehicle per day can make a huge difference in terms of your ROI of investing in these expensive cars.
we think it's another trillion dollar marketplace. We think that over a long period of time, as we see the cost of AV software come down, and usually we see the cost of hardware come down, I mean this is universal, but usually 30 to 40% per generation.
We've got over $10 billion in free cash flow, which is great, but it's on well over 10 billion trips a year. So we are not a high margin business.
we have our Uber 1 membership program. 50 million members now growing 50% year on year. Every company has a membership program. But the way I look at it is we're kind of like Netflix. For the same price, you get more content than anyone else.
I prioritize growth, I prioritize innovation over buybacks. If you're building the company right, you'll do both.
you've seen how powerful AI is, but at the same time how unpopular it is with the general public.
The Chinese capabilities in terms of manufacturing, both in terms of quality and cost at this point is unrivaled.
reserve now is over $5 billion run rate. It didn't exist, you know, five, six years ago.
Themes
- Enterprise AI ROI reckoning
- Model commoditisation & token economics
- Autonomous-vehicle demand aggregation
- Marketplace economics & capital allocation
- AI social licence & regulatory pacing
Mentioned
People
Companies
Ideas
- AI budget blown / metering headcount
- frontier-for-explore, efficient/open-source-for-scale
- AV demand aggregation vs building the driver
- 30% network-utilisation premium
- no single AV winner
- AV as a trillion-dollar marketplace
- 30-40% hardware cost decline per generation
- $10B FCF / low-margin marketplace
- Uber One membership flywheel (Netflix/Prime model)
- growth and innovation over buybacks (Amazon not Apple)
- social-licence pre-mortem / AI unpopularity
- Chinese manufacturing unrivalled / Foxconn-for-AVs gap
- Reserve $5B run-rate / on-demand to planned
- ground truth & troublemakers-as-mutations