At 6:40 in the morning, a klaxon horn sounds three times. “Gas!” a man in a hard hat and fluorescent vest yells out. There’s a hissing noise, and the helium starts flowing. From the tanks stacked like cordwood on a nearby truck, the gas moves through a series of hoses until it’s 55 feet up, then through a copper pipe and into the top of a plastic tube that hangs down to the ground, like a shed snake skin held up for inspection.

It’s a Wednesday in late June in Winnemucca, a solitary mining town in northern Nevada that has avoided oblivion by straddling the I-80 freeway. Along with two Basque restaurants, the Buckaroo Hall of Fame, and a giant W carved into the side of a hill, Winnemucca is the test site for Project Loon, a grandiose scheme launched in 2011 to bring the internet to huge swaths of the planet where sparse population and challenging geography make the usual networks of cell towers a nonstarter. Instead of building and maintaining earthbound structures with a range of a just few miles, Loon plans to fly packs of antenna-outfitted balloons 60,000 feet above the ground, each one spreading the gospel of connectivity over nearly 2,000 square miles.

Loon is testing in Winnemucca because the skies are mostly empty and there’s an airport for when the higher-ups want to come in by private plane straight from Palo Alto, just a short flight away. Today, the team is testing a new iteration of its communications system, which could support 10 times as many users as its current setup.

Half an hour later, the balloon is ready to go, held in place by a red horizontal bar and protected from the wind by walls on three sides. At the command of an engineer wielding a blocky yellow remote control, this structure, known as Big Bird, rotates 90 degrees to the left. Like Rafiki holding up newborn Simba in the opening scene of The Lion King, the various arms of the crane complex push the balloon up and out. As it takes on the weight of its payload—a triangular assortment of solar panels, antennas, and varied electronics—it freezes for just a moment. Then it’s up and away with the wind, climbing 1,000 feet a minute.

As far as routines go, it’s spectacular. “Never gets old,” Nick Kohli says. “Ne-ver gets old.”

When Kohli joined the nascent effort that was Project Loon in 2012, his job was to run around the world finding and collecting downed balloons from the Mojave Desert, rural Brazil, the coast of New Zealand. Loon was part of Google X, the arm of the search company that fostered audacious projects applying emerging technologies to stubborn problems in novel ways. One such project was self-driving cars. (In 2015, when Google restructured, creating its parent company Alphabet, Google X was renamed X.)

Kohli—not your usual Googler—is oddly qualified to survive the apocalypse. He didn’t get the grades for med school, so he trained as an emergency room technician—a background which, combined with his pilot’s license and eight years of search-and-rescue operations in the Sierra Nevada, made him just what Loon was looking for. This practical skill set and eye for operations makes him one of the many new kinds of people X needs to fulfill its mission: expanding Alphabet’s reach beyond the computer in your lap and the phone in your pocket.

With Alphabet’s help and resources, Kohli (who now runs flight operations) has seen Loon evolve past watching balloons fly hundreds of miles off course, to the point where a launch like today’s is nothing special. It’s just another step toward delivering the complex system Loon envisions in the future.

Today, X is marking a major step forward in that mission by announcing that Loon is “graduating”—becoming a stand-alone company under the Alphabet umbrella. Along with Wing, another X effort that delivers goods with autonomous drones, Loon will start building out staff and putting together its own HR and public relations teams. Its leaders will get CEO titles, and its employees will get an unspecified stake in their company’s success. Generating revenue and profit will matter just as much as changing the world.

Loon and Wing are not the first projects to get their diplomas from X (and, yes, employees get actual diplomas). Verily, a life sciences outfit with plans to monitor glucose levels with contact lenses, made the leap in 2015. And lo, the self-driving effort made the leap in December 2016, taking on the name Waymo. Cybersecurity project Chronicle ascended to autonomy in January.

The dual graduation of Loon and Wing—both big, ambitious, projects—marks a watershed for X and perhaps the moment when the secretive research and design division starts to make good on its mission. For the technological giant that has made its billions in advertising, X isn’t a junk drawer for unusual projects that don’t fit elsewhere in the corporate structure. It’s a focused attempt to find a formula for turning out revolutionary products that don’t just sit on a screen but interact with the physical world. By launching Loon and Wing into the world, X will soon discover whether it can effectively hatch new Googles—and put Alphabet at the head of industries that don’t yet exist.

But Alphabet’s attempt to birth the next generation of moonshot companies raises two questions. Can this behemoth grow exponentially? And do we want it to?

The Loon lab in X’s Mountain View headquarters is piled with the results of generations of falls and spills. Loon is based on a simple idea—replace ground-based cell towers with high-flying balloons—which concealed a beguiling series of technical problems. In 2013, after a year of work, the balloons still had a nasty habit of popping or falling to earth after a few days. (They carried parachutes to soften the blow to their electronics payloads, and the team would warn air traffic control of their descent). Before setups like the one dubbed Big Bird, when the launch process resembled a gang of kids trying to will a kite to take off, a puff of wind could derail the whole thing.

Now, a custom-designed “mother of all crates” keeps the system safe during shipping. Key components ride in a silver box made of metalized styrofoam that reflects sunshine and holds in warmth. An 80-foot-long flatbed scanner examines swaths of polyethylene for the microscopic defects that can reduce a balloon’s survival at 60,000 feet from months to days. Mapping software tracks the floaters across continents and oceans, using machine learning to identify the wind currents they can ride to wherever they need to be. With all these tools, the Loon team is learning: The company can launch a balloon every half hour and keep them in the air for six months or more.

This is the sort of development X allows for. For six years, Loon’s engineers and designers and balloon recovery operatives haven’t had to worry about funding or revenue streams or hiring HR people or who’s running their PR strategy. They’ve had access to Google’s machine learning expertise and to X’s “design kitchen,” a 20,000-square-foot workshop for prototyping any mechanical device they could think of. They haven’t needed a detailed business plan, let alone revenue or profits. They’ve been allowed to fail over and over, each time learning a little bit more.

X chief Astro Teller pitches X as a place for making the world better, but he doesn’t hide the benefits for Alphabet, including new revenue streams, strategic advantages, and recruiting value. And while he won’t reveal the moonshot factory’s employee count or operating budget, he makes clear that no matter how much money you might think X spends, it’s piddling compared to the value of what it creates.

All across X, teams pursuing an extravagant array of moonshots are finding their own ways to fail, with similar protected status. Ideas are welcome as long as they involve new ways to solve thorny problems. They come from all over. Some surge from the brains of employees. Others come from Teller or Google cofounders Larry Page and Sergey Brin. X employees plow through academic papers and stack up frequent-flier miles attending conferences, looking for the seeds of projects they could grow into something real. One unnamed project came from a researcher’s NPR interview: Someone at X was listening and asked her to come in for a chat.

Wherever they come from, most ideas stop first at the Rapid Evaluation Team. This small group meets a couple of times a week, not to advocate ideas but to shoot them down. “The first thing we’re asking is: Is this idea achievable with technology that will be available in the near term, and is it addressing the right part of a real problem?” says Phil Watson, who leads the team. Breaking the laws of physics means no dice. “You’d be amazed at how many kinds of perpetual motion machines have been proposed,” he says.

These meetings combine the unfettered thinking of a smoke-filled dorm room with the brutalizing rigor of a dissertation defense. The team has considered generating energy from avalanches (unfeasible), putting a copper ring around the North Pole to make electricity from Earth’s magnetic field (too expensive), and building offshore ports to simplify shipping logistics (a regulatory nightmare). They once debated working on an invisibility device. The tech seemed doable. “We kept saying, we should do it because it’s awesome—no, we can’t do it, because it’s going to cause more trouble, and it doesn’t solve any real problems,” Watson says. “It certainly would make criminals much more effective.”

The ideas that make it through this first evaluation are whisked to the Foundry, where whoever’s leading the fledgling project works through questions about the operations of the business they might create, something engineers aren’t always eager to do. This stage is led by Obi Felten, who came to X in 2012 after years of launching Google products in Europe.

In her first meeting with Teller, Felten learned about all the secret stuff X was cooking up, including internet balloons and delivery drones. And she started asking the kinds of questions you get from someone who launches products. What’s the legality of flying balloons into a different nation's airspace? Are there privacy concerns? Will you work with the phone companies or compete with them? “Astro looked at me and said, ‘Oh, no one’s really thinking about any of these problems. It’s all engineers and scientists, and we’re just thinking about how to make the balloons fly.’”

The Foundry uses this intense interrogation to root out the things that could kill a project down the line, before X has poured in piles of money and time. Take Foghorn, X’s effort to create a carbon-neutral fuel from seawater. The tech was amazing and the problem was huge, but two years in, the team realized they had no viable way to compete with gasoline on cost—and were reliant on technology that was closer to research than development. X killed Foghorn, gave everyone on the team a bonus, and let them find new projects to push. Ideally, the Foundry makes sure that the right projects get killed, as quickly as possible.

This is based on a simple premise: The sooner you can kill one idea, the faster you can devote time and money to the next one. Trying to change the world and make enormous new companies means shunning the traditional signs of progress. Uncovering the things mostly likely to doom whatever you’re doing is the only way to achieve success. Because once it’s good and dead, you can go back to the well for the next thing—the thing that might be the moonshot that lands.

Any project hoping to qualify as X-worthy must fall in the middle of a three-circle Venn diagram. It must involve solving a huge problem. It must present a radical solution. And it must deploy breakthrough technology.

That definition, which X uses to separate the delivery drones from the invisibility cloaks, didn’t exist in 2010, when X first took shape. The effort started with an experiment: Larry Page asked a Stanford computer science professor, Sebastian Thrun, to build him a self-driving car. At the time, Thrun knew as much about the technology as anyone: He had led Stanford’s winning bid in the 2005 Darpa Grand Challenge, a 132-mile race for fully autonomous vehicles across the Mojave Desert outside Primm, Nevada. When Darpa held another race in 2007, the Urban Challenge, the agency thickened the plot by making the vehicles navigate a mock city, where they had to follow traffic laws, navigate intersections, and park. Stanford came in second (Carnegie Mellon won), and Thrun, who was already doing work with Google, came to the company full-time, helping develop Street View.

The Darpa Challenges had proven that cars could drive themselves, but the feds weren’t holding any more races. American automakers were focused on surviving an economic collapse, not developing tech that could devastate their businesses. Google was a software company, but it had mountains of cash, and it was clear that bringing this idea to market had the potential to save lives, generate fresh revenue streams, and extend Google’s reach into one of the few places where looking at your phone is not cool.

So Thrun quietly hired a team, passing over the established academics who led the field in favor of a younger crew, many of them Darpa Challenge veterans, with less ingrained ideas about what was impossible. (They included Anthony Levandowski, who eventually found himself at the center of a bruising lawsuit with Uber, which the companies settled in February.) Page set his own challenge for the team, selecting 1,000 miles of California roads he wanted the cars to navigate on their own. Thrun’s squad called it the Larry 1,000, and they pulled it off in a conventional-wisdom-busting 18 months.

This move into the physical world was fresh ground for Google, whose taste for projects outside its core business had yielded Gmail, Google Maps, and Google Books—cool stuff, but still software. And the sight of Toyota Priuses chauffeuring themselves around the streets of Mountain View inspired possibilities, including more projects that didn’t consist solely of 0s and 1s.

But self-driving cars had fallen in Google’s lap. Finding other similarly hard, complex, worthwhile problems would require some infrastructure. Page made Thrun the company’s first “director of other,” in charge of doing all the stuff that didn’t line up with what investors expected from Google. Because Thrun was focused on the self-driving team (and after 2012, on his online education startup, Udacity), his codirector, Astro Teller, took the helm of a ship whose purpose and direction remained nebulous.

In an early conversation with Page, Teller tried to hash it out. “I was asking, ‘Are we an incubator?’” Teller says, sitting back in a chair with his trademark rollerskate-clad feet kicked out in front of him. Not exactly. They weren’t a research center, either. They were creating new businesses, but that didn’t convey the right scope.

Finally, Teller reached for an unexpected word. “Are we taking moonshots?” he asked Page. “That’s what you’re doing,” Page replied.

Creating a research division to build groundbreaking products is a mainstay of companies whose worth is tied to their ability to innovate. The tradition goes back at least to Bell Labs, founded in 1925 by AT&T and Western Electric. Made up of many of the smartest scientists in the country, Bell Labs is known for creating the transistor, the building block of modern electronics. It also helped develop the first lasers and, courtesy of mathematician Claude Shannon, launched the field of information theory, which created a mathematical framework for understanding how information is transmitted and processed. Along with eight Nobel prizes and three Turing Awards, the lab produced the Unix operating system and the coding language C++.

This breadth was key to Bell Labs’ success. There was no way to know what the next breakthrough would look like, so there was no point in demanding a detailed plan of action. Its leaders were fine with “an indistinctness about goals,” Jon Gertner writes in The Idea Factory: Bell Labs and the Great Age of American Innovation. “The Bell Labs employees would be investigating anything remotely related to human communications, whether it be conducted through wires or radio or recorded sound or visual images.”

Yet Bell labs functioned within some parameters. Its most valuable tool was basic research: Bell’s scientists spent years probing the fundamentals of chemistry, physics, metallurgy, magnetism, and more in their search for discoveries that could be monetized. And while “human communications” is a broad mandate, their work didn’t venture far outside what could conceivably improve AT&T’s business, which was telephones.

Silicon Valley got its first great innovation lab with Xerox’s Palo Alto Research Center, whose researchers stood out not for their scientific breakthroughs but their ability to take existing technology and adapt it for new aims that had never been considered. PARC created the laser printer and Ethernet in the 1970s and early ’80s and laid the foundation for modern computing by leading the transition from time-shared monsters that fed on punch cards to distributed, interactive machines—aka personal computers.

But in Silicon Valley, it’s best remembered for Xerox’s failure to capitalize on that work. The lab pioneered graphical user interfaces—think icons on a screen manipulated by a mouse—but it took Steve Jobs to bring them to the masses. Xerox’s bosses didn’t pooh-pooh the tech, they just didn’t see how it concerned them, says Henry Chesbrough, who studies corporate innovation at the Haas School of Business at UC Berkeley: “Xerox was looking for things that fit the copier and printer business model.”

By giving its denizens a near-limitless mandate and maybe not quite so limitless funding, X thinks it can create products and services that previous labs might never have discovered—or might have cast aside. It doesn’t do basic research, relying instead on other institutions (mostly governmental and academic) to create tools whose uses it can imagine. It doesn’t rely on having the smartest people in the world within its walls and is happy to scout for promising ideas and lure them inside. And, most important, it’s charged with expanding the scope of Alphabet’s business, not improving what’s already there. For all those Nobel prizes, Bell Labs was valuable to its owners because it made phone calls better and cheaper. Xerox’s shareholders appreciated PARC because it earned them billions of dollars with the laser printer.

X isn’t making these mistakes, because its job isn’t to make search better. It’s to ensure that the mother ship, Alphabet, never has to stop expanding.

In that way, X’s project hasn’t been to pioneer self-driving cars or launch internet-slinging balloons or envision autonomous drones; the real purpose has been to build a division capable of engineering such businesses. Its fetishization of failure and its love for ideas that make everyone look up, even if only to shoot them down, are all in service of this single goal: If you’re not failing constantly and even foolishly, you’re not pushing hard enough.

That’s great for Alphabet and for people who like the idea of self-driving cars (especially those who can’t drive) or tracking their health with non-invasive wearables or basking in the light of the internet in the dark corners of the world or getting their cheeseburgers and toothpaste without contributing to traffic and planet-choking emissions.

But Alphabet, through Google, already has tremendous influence over our lives: how we talk to each other, where we get our news, when we leave the house to beat traffic. For most people, it’s a worthy tradeoff for free email, detailed maps, and free access to nearly unlimited information. X seeks to multiply that influence by moving it beyond the virtual realm. Critics already call Google a monopoly. Now imagine its dominion extending into our cars, into the food we eat and the goods we order, into our physical well-being—into how we connect to the internet at all. Google today wields heavy influence over the parts of our lives embedded in our phones. Are we ready to let it in everywhere else?

André Prager walks into the room pushing a cart piled with what looks like garbage. It’s mostly cut-up pieces of cardboard, with a few bags of plastic odds and ends mixed in. Wearing a T-shirt that reads “I Void Warranties,” Prager used to work on engines for Porsche. In his spare time, he has made a jet-powered chain saw and a turbocharged Vespa. Now he’s a mechanical engineer on Wing, X’s drone delivery project. This is his cart of failures.

Not long after Wing started up in 2012, the team realized that landing drones on the ground meant wasting energy on power-hungry vertical flight. Instead, they decided the aircraft would hover and lower its package to the ground—somehow.

The team’s first attempt was a bobbin-based system, where the package would be attached to a cord that would unspool from the drone. “It sounded like a great idea, because it was so simple,” Prager says. It quickly proved a complicated mess: Winding the things properly was a pain. Every package needed its own system, since the cord came off with the package, hardly an elegant customer experience.

They tried less complex mechanical systems modeled on clicky pens and cabinet doors. (Prager shows me one prototype off his cart, a square of cardboard with a broken pen, a thumbtack, and a straw taped to it.) Nothing quite worked—packages wouldn’t always unhook, or the hook would release then reattach, or something would break. “Then we said, What if we could do it without any moving parts?” says Trevor Shannon, another mechanical engineer, video-conferencing in from Australia, where Wing tests.

That thought led them to their current design, which is about the size and shape of a fingerling potato with an indentation that hooks onto the package. It’s easy to attach by hand, and when the payload hits the ground, the weight of the hook naturally pulls it off. An “underbite” stops it reattaching itself. Prager doesn’t mind its humble style. “We measure our success by how unimpressed people are when they see it,” he says.

The goal of Wing is to make it easier for people to get stuff, without all the wasted time and carbon emissions that come with moving things around in cars and vans. Since 2014, Wing has been running pilot programs around Australia, first in Queensland, then in Canberra, the capital. It started offering drone deliveries to ranchers in remote areas (lots of those Down Under) and is now preparing to start flights in the suburbs closer to the city. It’s delivering small packages that customers can order from Chemist Warehouse (Australia’s Walgreens) and Guzman Y Gomez (Australia’s Chipotle).

The real hurdle to doing that at scale, though, isn’t the delivery system, nor is it the technology: Batteries and aeronautic controls have made enough progress in recent years to float an armada of drone delivery companies. The problem is how to do this safely, especially in crowded, tightly controlled airspace over the US and Europe.

So in 2015, the team started building an unmanned air traffic management system that would connect all its aircraft and give each drone its own defined corridor to take it from origin to destination. “We’re trying to build the delivery trucks and the roads to drive on,” says Adam Woodworth, who will take on the title of CTO when Wing moves out of X. The hard part of this isn’t just developing a system that tracks aircraft, it’s getting everyone in the sky to run the same sort of system. Wing is working with the FAA and has made parts of its system open source, so others can make interoperable systems.

The funny thing about this problem is that it’s not the sort of thing X is built to solve. It doesn’t take engineering or prototyping or off-the-wall brainstorming. It takes careful relationship building and close conversations with regulators and competitors—entities for which success means getting something right the first time. And if Wing can’t make that work, its long-term survival is in doubt.

That marks a change the new company will have to embrace, as will Loon: Graduation from X means a different relationship with failure. These are becoming companies that are supposed to succeed in the conventional sense, by offering real services and bringing in real customers for real money.

Loon has flown more than 18 million miles. It has provided internet to Puerto Rico after Hurricane Maria and Peru after devastating floods. Now it has to do something harder. “It’s time to leave the nest,” says Alastair Westgarth, a telecom industry veteran who came to X a year and a half ago and who will become Loon’s CEO. It’ll be his job to nail down agreements with service providers around the world, working his balloons into their networks and keeping their customers connected. It’s important to stay audacious, Westgarth says, to keep pushing on innovation. “But by the same token, you don’t want to take existential risks.”

Out in the real world, failure is just failure. And slowly, the balance shifts from a death wish to a survival instinct.

X will keep an eye on the fledgling Loon and Wing as they try to make it for real, but it will soon turn its attention to finding new moonshots to take their place. It will be years before internet balloons and delivery drones either dominate the skies or crash to earth. Years the Rapid Evaluation Team and the Foundry may well spend spitting out untold numbers of failures and biting into a potential success or two. It will be far longer before we have answers about what X’s failing and tinkering and refining and launching means for the rest of us.

But back in Winnemucca, the launched balloon is climbing steadily. It’s headed into the desert and will spend a night in the area before moving on toward Denver, then Nebraska. Nick Kohli tells me that three balloons that launched from its site in Puerto Rico a few months ago are in the area.¹ My eyes flit back and forth in vain until Kohli directs my gaze and I spot the tiniest and whitest of tiny white dots bobbing silently along, 62,500 feet above my head. That’s about .005 percent of the way to the Moon, which, all things considered, isn’t that far at all.

More Great WIRED Stories

• We have no idea how bad the US tick problem is
• The Pentagon's dream team of tech-savvy soldiers
• PHOTO ESSAY: The annual super-celebration in Superman's real-world home
• It’s time you learned about quantum computing
• Boeing’s proposed hypersonic plane is really, really fast
• Get even more of our inside scoops with our weekly Backchannel newsletter

¹Story updated at 11:50 ET on Wednesday July 11 to correctly note when the overhead balloons were launched from Puerto Rico.