The Drone Is Only 15 Percent of the Problem

Keller Rinaudo Cliffton tells Sarah Guo on *No Priors* that his path was a winding one. In college he built computers made of RNA and DNA that operate inside human cells, aiming for molecular doctors that could recognize cancer at the cellular level. After graduating he spent a year and a half as a professional rock climber, then got obsessed with robotics. Academic labs were producing cool technology, he says, but almost nobody was making it reliable enough that millions of people could actually depend on it.

The more his team learned about robotics, the more logistics looked like the right fit: boring, repetitive work. The more they learned about logistics, the clearer it became that the system served the "golden billion" — the richest billion people on Earth — and did a very bad job for everyone else. Five and a half million children die each year from lack of access to basic medical products, he says, and the world pretends the problem is unavoidable or excusable. If they were going to build a new logistics system that was automated, zero-emission, and ten times as fast, they wanted it to be the first that served all people equally. That vision became Zipline.

They did not start there. Keller's first company, Romotive, grew out of not having jobs and putting robots on Kickstarter, where they sold $150,000 worth of machines — a fortune to them at the time. They assembled them in Tony Hsieh's Las Vegas apartments; the Zappos founder had lived in the same college dorm ten years earlier and became a mentor. The robots were laser-cut acrylic platforms you clipped a phone onto, for teaching kids to program or doing telepresence. People bought them, then put them on a shelf after a month or two. Homes are not controlled environments, and folding laundry is not a repetitive task.

Around then Keller met the founder of Kiva Systems, whose orange warehouse robots later sold to Amazon. His reaction was immediate: someone is going to design Kiva for outside the warehouse, and that will be a world-changing company. A simple idea, brutal in execution — and a problem they would have to commit a decade-plus to, which he now calls an underestimate. Ten years in, he says they are maybe one percent of the way there.

Medical logistics was a calculated first use case, not a sentimental one. Launching a brand-new physical system meant unit economics, regulation, and technology were all hard; healthcare helped on every front. Demand was obvious to customers, willingness to pay was higher because a delivery could save a life, and they could walk into the aviation authority with the minister of health and argue that the lives on the line justified taking some technical and regulatory risk. Healthcare is still Zipline's bread and butter, the majority of deliveries globally.

They were not smart enough, Keller insists, to know what to deliver. In 2016 they told Rwanda's minister of health they would cover every hospital and every medical product. She looked at them and said, "shut up just do blood." She did not fully believe them, but she gave them a shot: a government contract to deliver blood to 21 hospitals. She walked them through why blood is a logistics nightmare. Fifty percent of transfusions go to mothers with postpartum hemorrhaging and 30 percent to children. The catalog includes platelets, plasma, cryoprecipitate, and packed red blood cells, then ABO types and Rh factor, then wildly different shelf lives — frozen plasma lasting a year, packed red cells lasting 30 days refrigerated, platelets lasting six days at room temperature while being constantly agitated.

They said yes and thought blood would be easy. It nearly killed the company. Serving those 21 hospitals in the first year was incredibly hard, and only later did the network expand into the system Guo is interviewing him about. The first lesson was that the customer, not the founder, picks the wedge.

Keller describes the product as teleportation: any doctor, nurse, or healthcare technician pushes a button on a phone and summons a product to their GPS coordinates. It should feel magical, nearly instantaneous, and reliable in any weather, 24/7. Behind that button, Zipline now serves 3,000 hospitals and health facilities worldwide, with more than 20 global distribution centers expected by the end of January.

An order hits a fulfillment area stocked with medical products — blood was first — then goes to flight operations for a pre-flight, loading, and launch. There are no runways: aircraft leave on a catapult and come home on something that looks like an aircraft-carrier arresting system. The plane flies itself to the facility, descends to about 30 feet, never lands and never stops, and drops the package on a simple paper parachute into a "mailbox," an imaginary rectangle on the ground the size of a couple of parking spaces.

The aircraft returns, batteries are swapped, and it can be back in the air about two minutes later. Distribution centers were designed for about 150 flights a day; for the first three years that number felt like a pipe dream. Some centers now do north of 300 deliveries a day. Six years on, Zipline delivers a majority of several critical medical products in the countries where it operates, and Keller calls it the largest commercial autonomous system on Earth of any kind.

Guo lists the stack in one breath: fixed-wing long-range drones, hardware, software, sensors, autonomy, weather, doctor-facing UI, hospital integrations, traffic control, and the operational muscle of the fulfillment centers. When Zipline launched in Rwanda in 2016 the team was about 20 people. They had fooled themselves into thinking an awesome vehicle would be enough. The vehicle, Keller says, is only about 15 percent of the solution — a realization he thinks most robotics and autonomy companies still have not had.

Customers do not care how cool or even how capable the robot is. The only question is whether the product goes from A to B ultra-fast, reliably, 24/7, fast enough to save a life. Behind the curtain sit air-traffic-control software for the regulator, computer-vision pre-flight checks, unsexy data logging — each aircraft generates a gigabyte that has to be uploaded and used to make vehicles more reliable — and detect-and-avoid so aircraft can stay clear of each other and of other planes.

There was no master plan. After testing in Half Moon Bay, California, they thought they were ready to operate at national scale in Rwanda in any weather. They expected to onboard the first hospital in two weeks; it took nine months, and for those nine months they served one hospital, not 21, pulling constant all-nighters. They were never the biggest or best-funded player — a large Seattle e-commerce company was investing billions in similar technology — but they were, by far, the most practical. They launched into the real world and asked customers to pay immediately, with no excuses. Putting the first version in a paying customer's hands, Keller says, is a deeply humbling experience.

What they thought would be fatally hard often did not matter. They spent heavily on a lockstep two-by-two flight-computer architecture; to this day he is not sure they have ever had to use it. What actually hurt them were unsexy details: data logging; 43 kinds of fasteners in the first aircraft, versus two today; getting GPS lock fast enough between packing and launch, a problem the entire vehicle has to be designed around; delamination inside electric motors after they airshipped the first 20 aircraft to Rwanda. The culture that survived that year is still taught to new hires among nearly a thousand people: assume we are idiots, because that assumption is proven correct more often than not.

For a technical audience, Guo asks the basic difference between AI and the autopilot already on commercial flights. Keller's answer is counterintuitive: there are not very many things to hit in the air, so the self-driving problem aloft is "thousand times easier" than on roads. Hard perception problems — a child running in front of a car, another airplane flying too close — are still not solved at human safety level in either domain. But the sky is a big place, and Zipline could operate at multinational scale without solving cutting-edge computer vision first.

The company that inspired them was Tesla, not Waymo. Tesla shipped what it thought was a dead-simple product — put a battery in a car and sell it — while Waymo bet on a fully autonomous moonshot and would not sell a thing until it worked. Zipline took the Tesla path: launch something simple, make money, learn, then layer in autonomy. Other drone-delivery companies took the moonshot, spent billions, and have not served a single customer. Zipline's data, problem knowledge, and sense of what customers will pay for are now far ahead, and the company is integrating new autonomy every quarter, including a full detect-and-avoid stack in FAA certification.

In U.S. airspace, drones have no right-of-way; they must avoid everyone else, and they often cannot see them. Transponders are technically required but compliance is low in general aviation, and even a human pilot's see-and-avoid is, Keller warns, extremely poor. The requirement is a "two mile hockey puck" around the vehicle — a cylinder of awareness. Lidar and radar are heavy, power-hungry, expensive, and narrow; radar might have a 30-degree field of view on a 50-pound electric aircraft where every gram of payload matters. Cameras are lighter but fail in clouds, so they cannot be the sole solution.

About four years ago an engineer floated a "really stupid idea": what if they just listened, and triangulated other aircraft the way ears roughly locate sound? Dolphins, they note, use active sonar; this would be passive. The team tried to disprove it quickly. Early experiments failed to kill the idea; a two-person prototype also failed to kill it; it became the plan of record. Mechanical engineers had to make a microphone array work at 80 miles an hour, while signal processing and machine learning pulled aircraft out of wind noise. They collected their own data — microphones on airplanes, plus an array on the office ground recording overhead traffic around the clock. Four years later the hardware is in production, has flown in shadow mode, and has regulatory permission in some countries to control the aircraft and deconflict. The nets can identify make and model. Microphones are cheap, weigh almost nothing, and give a natural 360-degree view. The dumb-sounding idea was the one that actually fit the constraints, and it is what will let Zipline fly beyond visual line of sight and in clouds.

Starting in Africa looked odd from the Bay Area, where healthcare deserts exist too. Zipline wanted a public system so they could work with a country and serve every hospital at once, rather than knocking on U.S. health systems one by one. They also needed a small, innovative, agile government willing to make regulatory exceptions the FAA cannot. Rwanda checked all three boxes in 2016, and put up with a first year in which the company served one hospital for nine months.

Weeks before the taping they announced a $61 million national partnership: Zipline now delivers all medical products to every hospital and health facility in Rwanda and is the country's largest logistics network. They have added home delivery to patients, agricultural products that raise farmer productivity and cut childhood malnutrition, and a national e-commerce layer. Rwanda made what Keller believes is the country's largest investment in a private company, and is now a shareholder. Ghana has about 1,800 facilities; Nigeria launched late the previous year; Côte d'Ivoire and Kenya were coming online that month; the United States already had three distribution centers.

The meta-challenge of getting Zipline everywhere is that hardware is hard, especially when you design, manufacture, and operate the whole service. The company has to be excellent at supply chain, supplier industrialization, manufacturing, logistics, electrical and firmware and software and mechanical engineering, aerodynamics, aeroacoustics, fulfillment-center operations, regulatory approval, flight ops, and maintenance. Unlike software, you cannot scale 100x in a year; like Tesla, you are constrained by how fast you can build factories. The flip side is defensibility. Hard hardware companies at scale are among the most defensible on earth, for the same reason it is hard to build a new smartphone and compete with Apple. Keller thinks many of the most important companies of the next 20 years will be hardware and infrastructure businesses — more capital-intensive, harder to build, with more inertia.

He is blunt about the last two decades of American tech: too much appetite for incremental SaaS, too little for building in the real world. It was very difficult to raise money for Zipline for the first six years. 2022 was a breakout year: new geographies, a Walmart deal, more than 200,000 deliveries — more than the previous five years combined — and hiring former Tesla CFO Deepak Ahuja. In 2023, on already-signed contracts, flight volume would rise 400 percent, about twice as many autonomous deliveries in one year as in the company's entire history. Inflation in labor and gasoline, the same forces behind the recession, were accelerating customer adoption. U.S. systems including Intermountain Healthcare, Novant Health, MultiCare, and Michigan Medicine were using Zipline for autonomous instant delivery to homes, which Keller calls the other half of telepresence: if a doctor can diagnose you on a phone, the medicine should be on your doorstep in five minutes. Four new distribution centers were coming online in January, plus a three-year product he would not name that he thinks will be transformational for logistics.

Guo asks what founders who want audacious problems should steal from Zipline. Keller has already named ruthless practicality. He also remembers a Christmas party when the whole company fit around one table at a hole-in-the-wall Chinese restaurant. Someone asked the odds of success. He said about one percent. People were upset, until he added the other half: one percent of a world-changing outcome — automating logistics, serving people equally, universal healthcare access, zero-emission last mile — is worth it. They hired people who could live with real, likely failure.

The scariest, most ambitious ideas also have less competition. Zipline looks around and sees no one operating at anywhere near its scale. Incremental problems put you in perfect price competition with everyone else. Setting sights five or ten years out, then working in unfancy, scrappy, practical ways for a decade, is how you build important companies — and there are investors who will fund that, because those ideas often become the biggest companies.

Technology has a bad rap, he says, because the world assumes it will serve the richest people on the U.S. coasts while childhood malnutrition, women's reproductive health, democracy, free speech, and energy get left to governments and nonprofits — which, he argues provocatively, are bad at solving problems and incapable of deploying technology. Problems that affect six billion-plus humans need engineering, technology, and sustainable business models. The smartest graduates going to a large search engine to help sell 0.1 percent more ads is not a great use of talent. Zipline's proof point is that you can build a multi-billion-dollar company on problems that matter.

Do not read the African operations as philanthropy. "That is exactly the wrong way to understand what we do." They make money in all of these countries and are gross-margin profitable at almost all distribution centers. Sustainable unit economics, he argues, is how you scale a solution to six billion people. Entrepreneurship, not charity, is the engine.