Half of Ask DoorDash Orders Hit a New Restaurant. In Ten Years There Will Be More Dashers, Not Fewer.

Sarah Guo opens No Priors with Andy Fang and Stanley Tang, co-founders of a company most people still describe as food delivery. The one-liner undersells the brief they have been running. DoorDash completes more than three billion deliveries a year, connects some nine million Dashers, and reaches over 40 million consumers a month. The conversation is not about coupons. It is about asking the app for food in natural language, about a delivery robot named DOT, and about eight years of treating DoorDash as a robotics company.

DOT is the physical proof. It weighs 300 pounds, travels up to 20 miles per hour, and is about one-tenth the size of a car. It is, Tang says, the only delivery robot designed to move not just on sidewalks but on bike lanes and on the road. It has been live in Phoenix — Tempe feels, Guo says, like Waymo in San Francisco — for almost two years, and it went fully autonomous L4 last year. The rest of the episode is how a dorm-room PDF-menu experiment got to that vehicle, and what it implies for the nine million humans still carrying bags.

Fang starts with the software bet. A couple of years ago DoorDash tried to make a play in agentic commerce and first went long on voice. Voice was not the thing; maybe it will be later. What landed was a natural conversational experience. People could translate what was in their head into the interface instead of researching online and optimizing keywords. Nuanced restaurant discovery and grocery tasks both got easier. The traction, he says, has held up as the rollout expanded.

The behavior change is specific. On restaurants, 50% of Ask DoorDash trajectories are orders from places the person has never ordered from before. Fang calls that huge: new-restaurant trial has historically been one of the hardest metrics for DoorDash to move. On grocery, basket sizes run about 40% larger. People photograph the fridge and say stock me up, plan a week of meals around dietary constraints, cook pasta with the family this weekend, or just reorder the usuals — all faster than tapping through the old UI. Guo had never thought of DoorDash as hard to use. The numbers suggest latent demand that search was not serving.

People build eating habits and also want diversity. Conversational discovery is a way to express that, and it has a social-currency side: a friend who finds you a great new place looks like a hero, and DoorDash starts to sit in that role. They also invested in world knowledge the models do not have — what is trending on the internet, what forums are talking about — because training cutoffs are too early and people want to eat what is cool. Trust, Fang says, is built that way. Guo’s own Sunday-night extended-family dinner is the office-lunch problem in a house: poll who is coming, remember allergies, ask if anyone wants something special, then order. Put it on autopilot. Fang says the office manager staring at a cutoff time and a list of dietary preferences is the same job, and it is exactly what a conversational order is for.

Robotics at DoorDash did not start when autonomy became fashionable. Tang says they have been looking at it since 2018, when it was not obvious the technology would work. They still thought it would transform — and maybe disrupt — their space. Founder-led companies get to fund speculative bets on the horizon so they are not the ones being disrupted. Fang’s line, which Tang repeats, is that the next DoorDash will not be the same product with a better UI. That would be dumb. It will incorporate AI, agentic commerce, autonomy, robotics, drone deliveries. Eight years later Waymo is on the road and they are glad they wrote the check early, when the business case was less amazing than it is today.

Capital allocation followed the company’s original method. DoorDash itself was a Stanford dorm experiment: a site called PaloAltoDelivery.com, eight PDF menus, a Google Voice number. Only after the demand was real did it become a company. That philosophy has lasted thirteen years and was applied to autonomy. In 2018 the intention was not a giant robotics program. It was Tang plus half an engineer’s time, a skunkworks, go explore, form partnerships, learn. They did not think they needed to build hardware. Startups already had robots. DoorDash would be the platform: APIs, distribution. They spent several years working with everyone from sidewalk-robot players to robotaxi companies.

Three lessons came back. First, autonomy was a when, not an if; Waymo later confirmed it. Second, plopping a robot — or an LLM — into the network does not make things happen. You have to build infrastructure around it: which deliveries to take, operations, merchant integration, the consumer experience. That became the autonomous delivery platform. Third, and the reason they built in-house, the startups were not focused on a use case. They built technology first and retroactively hunted for a problem. In software, Y Combinator drills “build something people want.” In hardware and robotics, Tang says, people do the opposite, and the result is never quite the right fit.

The market split into two unhelpful buckets. Sidewalk robots are two-to-three-mile-per-hour water coolers on wheels, simple and effective in a narrow band. DoorDash’s average delivery is about three to five miles, and typical delivery time is 15 minutes once you exclude cooking. A two-mile-per-hour robot never closes that loop. On the other end, robotaxis are 4,000-pound vehicles designed to carry people, with chairs and air conditioning, going fast. A couple of burritos do not need that. Pickup and drop-off are different too. You can walk half a block to a Waymo. A package cannot. Tang calls it the first and last 100 feet: how food leaves the merchant, how the vehicle finds the driveway, how it reaches the porch people actually expect.

Working backwards from the customer, the right metaphor for dense-suburb 3-to-5-mile trips — where most deliveries happen — was an autonomous motorcycle, scooter, or bike-profile vehicle. Not 4,000 pounds: about 300. Faster than a sidewalk bot: 20 to 25 miles per hour. Nobody was building that. Rather than wait, they decided to control their own destiny, iterating against real DoorDash deliveries and the operational insight in 10 billion historical drops. That is what became DOT. Guo’s hypothesis for why so many teams still go technology-first is ChatGPT brain: people assume a general model will turn into applications for free, and treat the rest as secondary. Tang agrees that might work in software. DoorDash is a physical-world business. The physical world is messier.

No two of the three billion annual deliveries look the same. Downtown San Francisco is not Dallas, and neither is snowy Helsinki. Pizza is not ice cream; dinner is not grocery; retail, pharmacy, and parcels are in the mix now. McDonald’s and Starbucks integrate differently from a mom-and-pop sandwich shop; a drive-through is not a strip-mall counter. The advantage, Tang says, is simply called DoorDash: 10 billion deliveries of data, 40 million monthly consumers, and hard-won knowledge of what happens when things go wrong. Fang adds that they launched Tasks a couple of months ago so the Dasher fleet can collect data points for world models. Unlike ChatGPT, the V1 of physical autonomy is expensive, and there is real pressure to be better than today’s alternative on quality or cost. Otherwise, he asks, what are we doing?

Guo’s robotics point is that a cherry-picked demo is cheap; covering the real distribution of environments is the industry’s hard problem. Tang’s hiring pitch is the same idea in reverse: do you want prototypes in a PhD lab, or something that ships? Autonomy engineers spent a decade without seeing their product used. DoorDash Labs’ culture is pragmatic — get it into the world and let contact teach you — which is why they were doing autonomous deliveries in Phoenix for over two years before most people knew. Building a demo that works once is not the same as a scaled fleet, a scaled service, a scaled business. Run seven days a week, ten hours a day, and things start breaking.

Dirt covers a camera: how robust is the stack? DOT drives like a bike, hugging the right side or the bike lane, so fallen leaves can put the right two wheels on foliage and the left two on asphalt. Torque to each wheel has to change; middleware and low-level controls have to notice. Hard braking in an edge case can let regenerative braking overpower the battery and cause an electrical shock the safety system must absorb. Boot time was a joke until it was not: an engineer’s Jenkins script, hacked together in a couple of hours, was fine for a demo and then crashed half the time, taking 30 to 45 minutes, across hundreds of robots every morning. Multiply by 500 and it is a productivity disaster. Then come depots, maintenance, charging, manufacturing, merchant education, pickup and drop-off. Autonomy is only one layer of a physical-world business.

The homeliest problem is finding the customer. Punch an apartment into Google Maps and the pin is often tens of meters off. A human Dasher glances and knows which building, which door, which gate. A robot drives to the pin and faces a row of similar storefronts. That drop-off history — where humans actually left the bag — does not live in Google Maps. It lives at DoorDash, and it is the first-and-last-100-feet dataset. Guo, an investor in a home-robot company called Sunday, offers the dishwasher cat: you will not invent that in an office. The only way to learn the distribution is to put machines in the real world until the data shows you the boundary. Fang’s complement is operations. A lot of AI researchers treat ops as dirty, or as something to postpone. Physical-world AI, he says, is exactly what a world-class operations team is for: collect, annotate, deploy, run the fleet.

Phoenix is climbing from a handful of robots toward 100, then 1,000, then — Tang’s joke with Guo — toward three billion deliveries. He splits scale into three components. Autonomy was the science-fiction question five years ago. Waymo and Tesla have made their breakthroughs; DoorDash made L4 last year on a stack built entirely in-house and purpose-built for delivery. You cannot copy-paste Waymo onto a bike-profile vehicle that constantly moves between road and sidewalk. As far as Tang knows, nothing else in the world behaves like DOT. They still have to bring it to the Bay Area and more cities, and more edge cases will appear. Funny thing, he says: autonomy is becoming less of a blocker.

Operations is next. Phoenix restaurants do not look like San Francisco, London, or Helsinki. Integrations, pickup points, the interface layer, fleet management — every city is a new exam. The surprise bottleneck is hardware. Five years ago everyone called hardware a commodity. The first hundred robots were hand-built. The next thousand or ten thousand raise supply chain, component reliability, and how long the machine has to last — which they can actually measure from the field. They partnered with Also, a micromobility company spun out of Rivian, with RJ Scaringe as founder and chairman, because those people know how to scale vehicles. The problem five years ago was whether autonomy worked. Now it is operations, commercialization, hardware, and manufacturing. That, Tang thinks, is where DoorDash’s scale and ops advantages shine.

Inside a 10,000-plus-person company, Guo asks who owns the productivity shock. They acquired Metis last year to infuse AI-native ways of working into a workforce trained on old process. Coding is the obvious win, and they published Dashbench to see how models and harnesses perform on their own coding tasks — and to calculate ROI on the bill. Fang looked at spend a week before the taping: June was about 20 times January. It has since flattened, partly by cutting waste, partly by routing cheap tasks to open-weight models and keeping closed-weight flagships for hard ones, same intelligence, lower price. Seat growth is actually fastest in non-technical orgs: analysts, operators, account managers prepping QBRs with strategic merchants. With frontier labs, a familiar wall appears. Internal teams say the model is okay; scrub the data, wrap an RL environment, send it over, and the model crushes the simplified task; put it back on real enterprise data and it drops. The open question is harness versus missing capability in the training distribution, and that answer decides the next wave of spend beyond coding.

Guo asks the obvious long-term question: do the Dashers go away and the dots take over? Tang’s prediction is the opposite. In a world full of robotics, drones, and AI, ten years from now there will be more Dashers, not fewer. The arithmetic is scale. Over 9 million Dashers, the business growing 25% year over year: 5x or 10x from here, where does supply come from? Half of America is not going to run deliveries every month. New modalities and efficiency have to fill the gap, but demand will also surge as delivery gets more affordable. Guo looks forward to six packages a day. The fleet Tang wants is multimodal: DOT for suburban 3-to-5-mile strip-mall runs, drones for rural or poor-road lightweight orders, humans for the complicated multi-step grocery job that climbs stairs and pick-packs. The consumer app does not change. The merchant keeps one integration and suddenly has Dashers, robots, drones, and the AI tools on the same pipe. That local-commerce ecosystem, across 40 or 50 countries, is what he thinks is hard to copy.

Fang’s forecast is about how people buy, not only what delivers. Google search queries have gotten longer over the past couple of years. He reads that as people getting more comfortable talking to agents and apps the way they talk to a person. Lower friction there is more orders — food, grocery, retail. The other shift is agent-first design. They launched a DoorDash CLI the week before the episode. Agent traffic on the web, he likes to note, already exceeds human traffic. If college kids in a garage rebuilt DoorDash today, it would probably be agentic first; they would rather build that themselves than wait to be built around.

One concrete CLI story: a startup office manager realized DoorDash sold more than lunch — convenience and groceries too — pointed a camera at the pantry shelf, and fired the agent whenever the shelf ran low. That use case is awkward in a tap-and-scroll app and natural once an agent can place the order. Fang and Tang keep returning to the same method: write down the use case and the real constraints, distance, time, cost, then pick the technology. If a design needs the user to change habits in order to work, it probably has not found the need. Guo closes by admiring the ambition on both the interface and the physical network. The founders sound like they intend to keep running both bets in parallel.