Kyle Vogt Pitched 120 Investors for Cruise. He Will Not Sell the Home-Robot Company

Sarah Guo introduces Kyle Vogt as a serial founder: Twitch, then Cruise, the autonomous-vehicle company GM acquired for a billion dollars, and now The Bot Company, a consumer-robot startup. Elad Gil asks why start Cruise when he did. The year was 2013. There was not really self-driving technology as it exists today. There was Google's project, and rumor had it they had spent about $100 million and hired the world's best engineers. Going after that looked a little crazy.

Even as a repeat founder from Twitch, credibility did not travel. Twitch had not been acquired yet, and the idea was still wild enough that money was hard. Vogt says he scraped the bottom of the barrel. He pitched about 120 investors over the course of probably a couple of years to raise what they needed. The thesis was not to copy Google's ultimate self-driving car. It was the lean-startup version of a moonshot: the minimum quantum of utility, cheaper or easier to execute, so you could get to market and iterate.

The first product was a retrofit. Take a regular car, add sensors, put a computer in the back, and get it to drive, an early cousin of Tesla Full Self-Driving, on one model, a BMW. Without a carmaker's blessing you reverse-engineer protocols and attach motors to steering wheels, so the business was not sustainable, but the technical proof was enough. It peaked at Y Combinator Demo Day, when Sam Altman sat in the car and they rode it to the venue. After about a year and a half they decided they had enough technically to go after the big fish, robotaxis. Uber and Lyft had become household names, with talk of going public, and a hole in their unit economics: paying drivers. Self-driving stopped looking like sci-fi and started looking like demand. They raised from Spark Capital, went straight at robotaxis, and within a year were acquired by GM. By 2015 they had prototypes in San Francisco obeying lights, changing lanes, and going from A to B from an iPhone app.

Gil sets up Tesla's camera-centric stack against Waymo's heavier hardware, robotaxis against selling cars. Vogt's first cut is commercial, not technical. To be fair, Elon nailed it from a business-model perspective. Tesla has been making billions of dollars of profit while developing self-driving cars, whereas everyone else has been burning billions to get to basically the same point. Start with custom vehicles, lots of expensive sensors, and a constrained environment, or start with an unconstrained environment and cheap sensors that do not quite work without a driver. Both paths are aiming at low cost, works everywhere, commodity sensors. Tesla won that hands down.

The popular claim that you cannot get to full self-driving from mostly self-driving is, given a long enough time span, almost certainly wrong. Musk does not have to finish by a date or run out of money; he makes money along the way. The risk is customers rage-quitting the program, and they are getting something they like in Tesla FSD. In 2013, 2015, even 2018, a fully driverless car on cameras and low-cost sensors was not viable. Look again in 2025. Large language models and generative models applied to perception and even motion planning changed the compute you need. From a single camera image, not even stereo, you can extract accurate depth, and those models get better every day. A 2025 technical bet, he says, does not involve a bunch of expensive lidars or exotic sensors. It involves the most commodity, high-volume sensors you can get, probably just several more of them than a typical driver-assistance system.

Hardware is still a limit. High-performance compute in an automotive temperature range and safety-critical environment is hard; Cruise made custom chips, Waymo almost certainly does too, and there is room for more automotive silicon. Most robotaxi deployments also rely on remote assistance. Multi-carrier cellular works where there is coverage. The missing piece may be Starlink or something similar as always-on fallback. Otherwise you have an open question on Highway 1 in California with no cell service: should an AV even be on that road if you cannot reach the rider? The right year to start a self-driving company, in retrospect, was around 2020, when hardware and software would be mature now. Start earlier with a nimble team, and the validation, testing, training, and public-road infrastructure still let you rip out parts of the stack and move fast.

On China, Gil offers two stories: mostly teleoperation and sizzle, or rapid progress with fewer safety constraints and parity with the West. From public videos, Vogt still sees a lot of teleoperation. Tesla may even start near one remote operator per vehicle; Cruise and Waymo were probably close to that. Brute-forcing deployment to collect data is fine by him. At 25 percent remote assistance you have already cut labor cost 75 percent, almost the diminishing-returns point. A thousand AVs with 250 people monitoring them sounds crazy and, on unit economics, is not. A one-to-four ratio is trivial with today's technology. Twenty-to-one or fifty-to-one later is single-digit points of margin. Anywhere less than one full human in the car already makes the economics work, on a path to better safety and lower cost.

Starting another company after Cruise is, Gil says, a brave endeavor, because anyone who has done multiple startups knows how painful they are. Vogt did some soul-searching and decided he is a builder. Sitting on the sidelines or helping other founders would be fun and would not scratch the same itch. He is 39 and feels he has at least one more startup in the tank. The first company, Justin.tv and Twitch, was straight out of college: survive, and it became games and entertainment. Cruise was for impact, technology that saves lives and gives time back. The third time he wants impact and fun: people he likes, hard technical problems, amazing products. That product is home robots.

The impact is hiding in plain sight. There are 24 hours in a day. Sleep eight, work eight to ten, and the remaining hours are scarce. People spend a surprising amount of that remainder on unskilled labor, acting like robots: making the bed, doing the dishes, folding laundry, picking up toys. Those tasks are not what make us human. They detract from our humanity, which is exactly why they are perfect to automate. Describe a machine that takes them away and it clicks. In five years, maybe ten, a house without multiple home robots will seem as insane as a house without a sink, a laundry machine, or a toilet, if the machines are affordable, which he wants them to be.

Gil likes the plumbing analogy: most of human history had no running water, just a bucket down a hill, and we stopped calling pipes technology. Vogt dates plumbing and electricity to the turn of the century, then a 1950s and 1960s surge of appliances, the housewife in the ads standing next to a microwave and a dishwasher. The last wave of excitement inside the home is 50 or 70 years old. It is time to revisit it. Getting a beer from the fridge to the couch has been the nerd dream since the dawn of robotics, and it is hard because a home is the opposite of a factory line: sloppy, different in every house, changing every day. Classical computer vision that reconstructs 3D objects or fits a map makes that computationally brutal. End-to-end learning, imitation learning, and reinforcement learning let you teleoperate a demonstration, collect human data, or learn from internet video, and give the robot some common sense. On top of that you talk to it in natural language instead of typing into an app. For home robots, unlike the 2020 AV timing, now is the time, even if it still feels a little early.

Gil asks what transfers from Cruise and Waymo, given the classic kid-chasing-a-ball problem. The big difference, Vogt says, is that a driverless robotaxi is not a product until it reaches superhuman safety, whether that is slightly better than humans or ten times better, which is where most driverless cars on the road today sit. There is no MVP, no launch with something partially useful. You cannot constrain public roads enough. Night or sparse areas help, and still a child can dart out anywhere, so you need a high number of nines before you have any product. In the home, and in most consumer and industrial settings, safety still matters. The functionality bar and the constraints you can put on the system let you launch much more quickly.

Is there a Tesla-like path of making money along the way? One approach he has seen is to sell a high-price robot now, a humanoid or something like it, fully teleoperated, at a crazy number such as $50,000 plus $1,000 a month, and call it the first robot that will do stuff in your house. The market is small; the approach is viable. The other side is robots that do not fulfill the whole household promise but do little useful bits. At CES this year he saw Roomba-like devices with a tiny hand that can pick up a sock in the way. Those incremental products sell, collect data, and train models on the way up the ladder toward the Holy Grail: a robot that takes the place of a butler and a housekeeper.

Gil notes that the AV market raised billions across dozens of startups, and the things that actually worked look like incumbents, Tesla and Waymo, plus Cruise under GM, and maybe Applied Intuition on software. Will robotics shake out the same way? Vogt expects the bubble in software AI and in the next robotics wave. A few big rounds get everyone excited, other investors chase markups, people drop out of PhDs, and most of those companies are low quality: half-in founders hedging a return to work, get-rich-quick hopes, or brilliant academics who leave to commercialize their research. That last phrase is a huge red flag. It means they will not be flexible on the solution; they will force a square peg into a round hole, which is deadly when a startup is constantly wrong and has to adapt. When the bubble pops, the industry will look wiped out. A handful of good companies started in that window will be fine. In hardware it is hard, and incumbents have an edge, but Aurora went public, Zoox was acquired by Amazon, and Cruise itself was a from-scratch company from the beginning of the cycle.

People project personhood onto machines, especially in the home. Cruise named every car. A car does not look like a creature, but once it moves on its own the brain starts treating it as an entity. You can pretend that does not exist and live with cognitive dissonance, or you can lean into it and integrate it the right way. Too much anthropomorphism implies human-like behaviors the product does not have. Rodney Brooks wrote an essay arguing that, for robots, appearance sets expectation. The more humanlike the machine, the higher the average person's bar: it looks like me, walks like me, has a face and talks like me, so it must do all the things I can do. In 2025, selling a humanoid and implying it can do all of that would be a leap of faith, many years short of those expectations.

Design, then, should match what the robot can actually do, or go the other way: instead of over-promising with a humanoid, surprise people with how much a less human form can do. Vogt would rather surprise and delight than set customers up for disappointment. Consumer skepticism follows a pattern he saw in self-driving, intuitive and still striking. About 75 to 80 percent of people said they would never trust one of those things. After one ride, that dropped to 20 or 30 percent. The more transformative and science-fiction a technology feels, the higher the skepticism. A home robot sounds like sci-fi. The most powerful way past that is not classical marketing and spec sheets. It is people using the product and telling someone they trust: I have this in my home, one of the most intimate spaces, and it works, and I love it.

Guo asks what moves the timeline between five and twenty years until robots are as expected as appliances. Vogt is sitting in his office, so the joke is that the clock starts when the podcast ends. The serious answer is the same feeling he had in 2013. The basic building blocks are there. For a low-cost home robot, cost matters because he wants a lot of people to have one, he can point at every technical challenge he can see today and at a technology built in the last year, a paper, a product, or a chip in the pipeline. Execution is how fast those pieces go together. The rest is unknown unknowns you cannot see until enough robots are in homes: a feature people insist on, or a behavior that makes them never want one again. No one else is really doing this, the cloud of uncertainty is large, and the timeline could be one year or twenty. The way to find out is to charge forward.

On manufacturing against Chinese robotics firms, the sentiment Vogt hears is almost do not even try. A US widget uses $100,000-plus engineers, domestic machine shops and tooling, slower iteration, higher cost, mixed quality. He thinks a global company can still use contract manufacturers and existing supply bases if you travel, pound the ground, and make the connections. Pure mechanical engineering services will lose on margin, so you have to build a brand. Complicated software and taste are harder to copy than measuring a part and redrawing it in CAD. Even with lots of teleoperation, he still thinks China is years behind the best US self-driving companies. Home robots are the wild west: nothing established to copy, and if the devices are connected and keep getting new models, there are angles for a durable US or other company, provided the cost gap is not so large that product and brand cannot win.

If he could wave a wand on policy, he wants more regulation of AVs, not less. Cruise got wiped out; Waymo is expanding much more slowly than the merits and safety of the technology would suggest. Early US airlines had no FAA. A crash meant getting sued out of existence, and many of the first airlines disappeared. The FAA was created so the country could have airlines: monitor safety, and in exchange give reasonable liability limits. That has not happened for self-driving. The only companies that stand a chance are giant tech firms that make money elsewhere and can afford the liability. Otherwise it is bleak. Home robots sit in a similar vacuum. There is no cybersecurity regulation. A Chinese-manufactured robot with cameras and a microphone can run in a house and send data who knows where, uninspected. Industrial-robot safety practices are a poor fit. Regulation, even in an anti-regulation administration, is a necessary enabler: it forces best-practice conversations and oversight that help companies and consumers.

Gil raises the drone counterexample: the FAA over-regulated, the US fell behind, and DJI took low-cost consumer and military-adjacent markets. Vogt agrees transportation innovation has suffered. eVTOL startups have fizzled into certification walls, and there is still no pathway to a pilotless plane. He wants two tracks: a mature track for passenger airlines as they exist, and an innovation track with grants and a phased release from prototype to regulated scale. Spell out what maturity looks like at each level, then open operating areas as you demonstrate it, the way self-driving did and the way Boom's supersonic jet went from a low, slow flight to the sound barrier as the data checked out. Regulation should stop irresponsible people from going zero to supersonic. Responsible people will take the steps if a pathway exists.

Gil asks the number-one thing Vogt would do differently after Cruise, besides have more fun. Vogt's own punch line is: don't sell it to GM. A big company whose core business is in another domain, selling pickup trucks and SUVs to Midwest customers, is not compatible with urban robotaxis. When push comes to shove they pick the core. Cruise, he says, got completely decimated by GM's lack of priority and then complete abandonment. Lesson one: he is never going to sell another company again, ever. Maybe it will IPO. There will never be an acquisition in his life again. If he is working on something today it is because he thinks it is important and he cares, and the day you sell is the day you have to let that go.

Lesson two is team size. He fell, with many peers, into Silicon Valley dogma: hire VPs first, they hire directors, they hire senior managers, keep fan-out under eight to one, run performance-review cycles. Bureaucracy and structure open communication gaps between the people doing the work and the people making decisions. The fix is to keep the company very small, make every seat count, put the absolute best person in every role, and never grow so large that it crumbles under structure, bureaucracy, and politics. Traditionally that meant small ambitions. He does not think that is true anymore. Coding assistants keep getting better, tools like Deep Research show up, and nearly every job function can be partially automated. As someone who programs a lot, he can now take on work that used to mean hiring an iOS specialist or a low-level Rust person for motor drivers. A couple of good engineers sitting next to an LLM and decent coding tools can adapt. Grand ambitions, tiny team: that is the plan for The Bot Company.

Gil asks where the stamina comes from: three high-risk companies back to back, and seven marathons on seven continents in three and a half days, and whether there is a Bryan Johnson-style supplement. Vogt has not tried young blood. He has a problem. Once an idea is in his head it burns a hole in his brain. He cannot sleep or focus until he sees it through. It is not twenty ideas a day. The stars align, this should happen now, and then he cannot let go. Delayed gratification plus the daily satisfaction of a little progress on something complicated is the fuel. The companies are hard, and he enjoys every day, and he cannot imagine doing anything else. When the taping ends, the joke from earlier still stands: he is going back to the office.