IN2018, ON A TRIP to his ancestral homeland, Alexandr Wang listened as China’s brightest engineers gave impressive presentations on artificial intelligence. He found it odd that the researchers conspicuously avoided any mention of how AI might be used. Wang, whose immigrant parents were nuclear physicists at Los Alamos National Laboratory, where the first atomic bombs were designed, was unsettled.
“They were really dodgy on what the use cases were. You could tell it was for no good,” recalls Wang, the cofounder of Scale AI, who has no second “e” in his first name so that it has eight characters, a number associated with good fortune in Chinese culture. Scale was then an up-and-coming startup providing data services primarily to self-driving auto-makers. But Wang began to worry that AI might soon upend a world order that, excepting the fall of the Soviet Union, has remained mostly stable since World War II. “If you think about the history of humanity, it’s mostly been punctuated by war except the last 80 or so years, which have been unusually peaceful,” he says from Scale’s sixth-floor headquarters in downtown San Francisco, as the occasional (partly) self-driving car zips by below. “A lot of that has been because of American leadership in the world.”
At first glance, Wang, 26, exudes the skittish energy of a fresh college graduate. He listens to “sad girl” musicians like Gracie Abrams and Billie Eilish and dresses “gorpcore,” an in-vogue style of fashionable hiking clothes. He posts Instagram photos with actor Kiernan Shipka of Mad Men fame and spouts pithy nuggets on Twitter: “The best problems can only be solved by blood, sweat, tears, spirit and an overwhelming sense of purpose,” he wrote in one February tweet. At bars, he still gets carded regularly.
None of that matters in Silicon Valley and D.C., where he’s already a power player. His rise began with a bet he made in 2016 to “label” the mass of data required to power AI, primarily for self-driving cars. Someone needed to train the AI to know the difference between a paper bag and a pedestrian. He cornered that market and put Scale in a good position in another sector: generative AI. It was a prescient move that helped him garner a client list that includes the biggest names in AI—and the U.S. government.
“We’re the picks and shovels in the generative AI gold rush,” he says. It has quickly become a lucrative business for Scale, which says it pulled in $250 million in revenue last year, at a time when many AI startups aren’t yet making a cent. Its tech has been used by the government to analyze satellite imagery in Ukraine and by OpenAI to create ChatGPT, the bot that rocked the world with its ability to answer trivia and write poetry. Bret Taylor, former co-CEO of cloud software giant Salesforce, likens Scale’s rise to that of cloud computing darlings Snowflake and Datadog. Former Amazon consumer boss Jeff Wilke, one of Wang’s most trusted advisors, takes an even more enthusiastic view: Scale could become the Amazon Web Services of AI.
Investors awarded Scale a $7.3 billion valuation in 2021, making Wang the latest Silicon Valley insta-billionaire. But his fortune wasn’t built entirely on silicon. It was also built with a vast outsourced workforce that performs a rudimentary task crucial to AI: labeling the data used to train it. Those people—some 240,000 of them in countries including Kenya, the Philippines and Venezuela—work for Remotasks, a subsidiary Scale doesn’t mention in public marketing materials. In other words, if AI does someday liberate humans from mundane workplace tasks, it will have done so using a legion of workers in the Global South, many of whom are paid less than $1 an hour.
“They’re very, very important to the process of building powerful AI systems,” Wang says of his Remotasks workers.
Scale was conceived as a one-stop shop for supplying human labor to perform tasks that could not be done by algorithms—essentially, the antithesis of AI.
They’re also, increasingly, an ethical concern, with worries emerging about substandard working conditions and low pay. Meanwhile, competitors see Scale as a house of cards that has suffered layoffs and declining value on secondary markets in the past year that has stripped Wang of billionaire status. (Those markets now value his 15% stake at $630 million. Scale argues it’s worth closer to $890 million.) “Scale markets itself as a technology company,” says Manu Sharma, cofounder of rival startup Labelbox. “For us, they’re no different than any business-process outsourcing company.” Tech upstarts think they can do what Scale does better, while traditional outsourcers think they can do it cheaper.
“I would say that we’ve been working on this problem longer and have built more technology than anyone else,” Wang counters. He’s trying to follow Amazon’s playbook of managing the entire chain, from warehouses to shipping. For Scale, that means both the machines—which are increasingly automating the data work—and the human army, which is growing ever larger. “We’re always going to want a human in the loop,” he says.
BEFORE COLLEGE, Wang moved to the Bay Area to work for internet startup Quora, where CEO Adam D’Angelo gave him a crucial piece of advice: Four years of college is overrated, two is underrated. In the end, Wang spent just one year at MIT before heading to storied startup accelerator Y Combinator. There he teamed up with Quora alum Lucy Guo, another dropout, to start Scale in 2016. He remembers being “ridiculously young” at the time, just 19. “But I was just like, ‘Yeah, I know how to code. We’re going to go do this thing.’ ”
As it was first conceived, Scale was to be a one-stop shop for supplying human labor to perform tasks that could not be done by algorithms—essentially, the antithesis of AI. Accel partner Dan Levine was early to see its potential, offering the pair a seed investment of $4.5 million (and his basement as temporary headquarters) in July 2016. Within months, Wang and Guo realized Scale was a viable solution to a problem plaguing the self-driving car companies at AI’s then-frontier: They had millions of miles of on-the-road driving footage with which to train their autonomous vehicle AI, and not nearly enough people to review and label it. Scale could fill that need.
In 2018, Wang and Guo were named to Forbes’ 30 Under 30 list in enterprise technology. Guo subsequently left the company “due to differences in product vision and road map,” she says. “I think Alex has done a great job continuing to run the company.” Guo otherwise declined to comment for this story, and Wang declined to speak about their split.
Investor Mike Volpi first heard Scale’s name during a 2018 board meeting for autonomous vehicle (AV) startup Aurora. “Who?” he remembers asking. Scale’s data labeling service had become crucial for Aurora, he learned, just as it had for Uber and for General Motors’ self-driving subsidiary, Cruise. Volpi persuaded his firm, Index Ventures, to lead an $18 million investment in Scale that August, when its revenue was still shy of $3 million.
The AV wager was becoming a cash cow. Scale’s client list now included major international auto manufacturers such as Toyota and Honda, as well as Silicon Valley behemoths like Google AV subsidiary Waymo, according to a June 2019 fundraising pitch deck seen by Forbes. An account with Apple’s secretive self-driving unit alone was bringing in more than $10 million, the document said, putting annual revenue on track to surpass $40 million. (Scale declined to comment on the deck.)
When Peter Thiel’s Founders Fund made a $100 million investment that minted Scale as a Silicon Valley unicorn in August 2019, it kicked off a 20-month, $580 million fundraising spree, the final round of which valued the company north of $7 billion. It had taken Wang, then 24, just five years to become the youngest self-made billionaire in the world.
“There is pretty much zero accountability for those working conditions.”
BYTHE TIME Scale dominated the data labeling market for self-driving car companies, its name had become something of an irony. The more it scaled, the harder it became to keep up with the demand for human labor. Wang first turned to outsourcing agencies to fill gaps, but costs quickly spiraled. Gross margins, which hovered at about 65% in early 2018, approached a mere 30% by the fourth quarter. Wang needed to stanch the bleeding while still capturing both the human and machine sides of the AI data training supply chain.
Enter Remotasks, Scale’s in-house outsourcing agency. Created in 2017, Remotasks soon became a priority as the company’s AV business skyrocketed. In need of cheap labor, Scale set up a dozen-plus facilities in Southeast Asia and Africa to train thousands of data labelers. By mid-2019, Scale’s margins had recovered to 69%, according to the deck.
Scale has been careful to position Remotasks as a separate brand. Its website makes no mention of Remotasks; the reverse is also true. Early employees say this was done to make Scale’s strategy less obvious to competitors and shield the company from scrutiny. Scale told Forbes it separated the two brands for client confidentiality.
In a 2022 study into working conditions on 15 digital labor platforms, University of Oxford researchers concluded Remotasks met the “minimum standards of fair work” in just two of 10 criteria, flunking equitable pay—which early employees say is pennies per hour on average—and fair representation. They noted that the “obfuscation” of its association with Scale creates confusion that “can contribute to workers’ vulnerability to exploitation.” Lead researcher Kelle Howson compared data labelers on digital labor services like Remotasks to garment factory workers in many of the same countries. “There is pretty much zero accountability for those working conditions,” she added. Scale says it is committed to paying workers “a living wage.”
Beyond the ethical considerations, there are business questions, too. What Scale is doing with Remotasks isn’t hard to replicate. Kevin Guo, cofounder of Hive, a startup that once fielded its own Remotasks rival before shuttering it due to tough margins, contends that the sort of data labeling Scale does is a commodity business. “Anyone who puts up a team can compete with you, and it comes down to price really quickly,” he says.
WHILE REMOTASKS’ huge overseas workforce is critical to Scale’s private sector success, it’s a nonstarter for the company’s other focus: defense contracts with the U.S. government, which is unlikely to share classified data with foreign labelers. Wang is therefore building a much more expensive domestic AI army. Last year, Scale opened an office in St. Louis and announced plans to hire 200 people, many as data labelers.
“There’s two things I deeply believe,” Wang says. “One, AI is a huge force for good, and it needs to be applied as broadly as possible. Two, we need to make sure that America is in a leadership position.”
Train a custom AI model on live data from America’s 1.3 million active service personnel and you might just change the nature of war.
So far Scale has made $60.6 million from such contracts, according to a government database. The company touted a $249 million award in a press release last year—but the Defense Department told Forbes it is one of more than 70 companies eligible for the money. Scale has so far received one contract capped at $15 million and no payouts have materialized yet. The lion’s share of government spending on AI is still going to the likes of Northrop Grumman and Lockheed Martin, not Silicon Valley upstarts.
“Those companies, they’re really not that cutting-edge when it comes to understanding generative AI,” Wang says. For him, government partnership is a long game. The U.S. government has already used Scale’s expertise to make strategic sense of satellite imagery in Ukraine. And that’s just the beginning. Generative AI, he says, could someday be used more comprehensively. Train a custom AI model on live data from America’s 1.3 million active service personnel and you might just change the nature of war.
But it won’t be easy to get there. Generative AI models require far more complex training than their precursors. They too need additional human help, but instead of simply labeling data harvested from the internet, people need to create it. For AI to explain why puppies are cute in a way that sounds right to the human ear, you need people to train it using natural phrasing. “Human-annotated data turns out to be extraordinarily impactful to model performance,” says Aidan Gomez, cofounder of Cohere, a Toronto-based OpenAI competitor that counts Scale as its primary custom data provider.
Not all AI companies are sold on Scale. OpenAI, for example, relies on Scale’s human labelers but opts to use its own software to manage the data, says cofounder Wojciech Zaremba. Three engineering leaders who used Scale at prominent AI startups told Forbes confidentially that they have concerns about the quality of its human-made AI training data. One described a text-based generative AI model that was hampered by the labelers’ poor English. “Their data quality can be high, but also that’s not a given,” said another. Said a Scale spokesperson: “We stand behind our products and [their] results.”
“Wang didn’t get to where he is because he’s a boy genius—MIT pumps out a lot of teenage dropouts. He has an absolutely insane work ethic.”
Alternatives are emerging. San Francisco–based Surge AI, which debuted in 2020, offers data labeling tools and specifically targets AI companies. OpenAI, along with upcoming AI heavies Cohere and Adept, use both Scale and Surge. Then there are billion-dollar Bay Area labeling startups Labelbox and Snorkel AI, which focus on bringing AI to non-tech enterprises.
In January, Scale slashed 20% of its full-time staff. Wang cited “uncertainty” in market conditions. “We increased head count assuming the massive growth would continue,” he wrote in a blog post. Shares of the company are currently trading on private secondary markets at a 42% discount to the last funding round in July 2021.
Scale’s stakeholders remain confident Wang can keep the company ahead of its rivals. “He didn’t get to where he is because he’s a boy genius—MIT pumps out a lot of teenage dropouts,” says William Hockey, the centimillionaire cofounder of $8 billion fintech Plaid, who sits on Scale’s board. “He has an absolutely insane work ethic like nobody I’ve ever met.”
Scale recently signed a strategic partnership with consulting giant Accenture, which plans to use its services to help hundreds of companies build custom AI apps and models. And with nearly a quarter-million human labelers, Remotasks is still growing, Wang confirms. All this growth comes down to what he views as Scale’s ultimate purpose: playing a role in maintaining America’s AI supremacy.
“We’re in an era of great power competition,” he says. “American leadership—I don’t want to say it’s at risk, but it’s never been more important for us to retain that.”
Update: This story has been updated to reflect Plaid’s most recent valuation, to clarify the nature of Scale’s usage in Ukraine and with details of a government contract from the Defense Department.
