As Alexandr Wang walked down the staircase to the open atrium at Scale AI headquarters in San Francisco last June, employees didn’t know if he was still their boss. A day earlier, the data labeling company had announced a bombshell deal: Meta was acquiring 49% of Scale for $14 billion, and its founder CEO was leaving to lead Mark Zuckerberg’s newly-formed superintelligence lab. Amid a swirl of confusion, many Scale employees thought he had already left. So, surprised to see him, the workers applauded as he headed to the stage for the company all-hands. “I literally teared up,” Wang tells Forbes now. “In another life, I'd be so excited to continue at Scale.”

He doesn’t remember exactly what he told them at the meeting, but according to one person in attendance, Wang began by recounting how he’d started building Scale as a freshman at MIT, then started to cry. “This is so stupid. Why am I doing this?” the person recalled him saying.

Wang’s remark was in reference to his tears, but it seemed like an apt question to apply to the entire situation. In the wake of the surprise deal, which had leaked a few days earlier, all of Silicon Valley was similarly asking: Why was Wang giving up his own growing company, then worth $13.8 billion, to go work for Meta, which was playing catch-up in AI to Google, OpenAI and Anthropic?

Scale had been a powerhouse in the market of human data, where armies of clickworkers and experts — like PhDs, lawyers and engineers — generate data to train cutting edge AI models like Google’s Gemini and OpenAI’s ChatGPT. Scale had also become a go-to vendor for the Department of War since it first brought large language models to the Pentagon in 2020. And the operation had, for a moment, turned Wang into the world’s youngest self-made billionaire . Now the tie-up with Meta threatened to railroad what had become a stalwart business providing infrastructure to the most valuable companies in AI. After all, the thinking went, what frontier lab would want to entrust its data to a company almost half-owned by Meta?

Wang, who took roughly ten Scale employees with him to Meta, says now that it was an “incredible opportunity” for both companies. The other half of the all-hands meeting was Wang formally introducing his successor Jason Droege, previously Scale’s chief strategy officer, and before that an executive at Uber and Axon, the company that makes the Taser gun. Despite all the drama of the previous week, the meeting itself was a tight 30 minutes, with no Q&A from employees. Instead, Wang and Droege kept it quick. “We shake hands, give each other a hug, say we're both excited about our futures,” recalls Droege. “And then the next moment, we've got to go do it.” (Droege still holds the “interim” CEO title, but internally he’s seen as its long-term leader.)

“We definitely anticipated turbulence. There's no question. I mean, the facts of the situation are, our founder went to another lab to help them.” Jason Droege, CEO, Scale

Almost one year on from the deal announcement — and a decade after its founding in May 2016 — Scale unsurprisingly looks like a much different company. Droege’s first order of business as CEO was to shift the company’s investment away from the data labeling business, and toward a model where it helps large enterprise companies, like Ernst & Young, Paramount and Cisco, as well as public sector clients, like the U.S. milidevelop their own internal AI applications.

The strategy appears to have paid off. Scale tells Forbes it tallied just shy of $1 billion in revenue last year, up from $870 million the year before . (Throughout the entire history of the company, it has brought in $2.5 billion in revenue.) Part of the revenue growth could be a lot of help from its new shareholder Meta: As part of the arrangement, Meta agreed to pay Scale at least $450 million a year for five years for its services, or more than half of its annual AI spend, whichever is less, Forbes reported at the time. Wang and Droege declined to comment, but it would represent nearly half of Scale’s annual revenue.

Droege also declined to break down the split between the data and apps business. He says data labeling is still the “vast majority” of revenue, but he expects revenue from the app business to overtake the data business within the next 18 months. Meanwhile, Wang, now at Meta, is still rooting for the company he cofounded. “There's this perception that Scale is cooked or done or whatnot, and that's just absolutely not true,” says Wang, calling the company’s resilience a “huge narrative violation.”

W ang, whose parents were physicists at the Los Alamos lab where the atomic bomb was created, started Scale ten years ago with cofounder Lucy Guo, a Thiel Fellow who dropped out of Carnegie Mellon as part of billionaire Peter Thiel’s program to get promising young people to ditch college. The pair, who met while both working at the Q&A platform Quora, had dreamed up several iterations of the company before it landed on AI infrastructure.

Things began to go sideways between Guo and Wang over future company roadmaps, Guo said last year on entrepreneur Emma Grede’s podcast Aspire. It all came to a head after Guo told someone she thought Wang “should get fired.” That person subsequently told Wang, Guo theorized on the podcast. “I was bitter about who I thought told him that because I thought there was a certain level of trust,” she told Grede, adding, “I felt a little betrayed.”

In the end, Guo left the company. “We don't really talk about this that much,” Wang says when asked about the situation, with a nervous laugh. “What they don't tell you at YC is how common the founder breakups are. But I think that's something that we worked through. And I'm really proud of everyone on the team in the company, to be able to be very successful beyond that.”

Seven years later, it was Wang’s turn to leave. The call from Zuckerberg first came last spring, as he was looking for a new AI chief after the disappointing performance of Meta’s flagship Llama 4 model. The two had met years earlier when Wang went to him for his advice on running a startup. Asked about Zuckerberg’s pitch to work for Meta, Wang downplays it. “I don't even know how much more complicated it was than Meta has this incredible opportunity in AI, and it's very, very exciting, and let's figure out how to work together,” he says, adding that Zuckerberg’s “conviction was just very, very clear.”

Have a tip? Contact reporter Richard Nieva at rnieva@forbes.com or RNieva.26 on Signal.

Wang then turned to his board members to weigh the opportunity. “When you look at transactions like this, deals come about when the pros and cons are almost tied,” says Mike Volpi, an early Scale backer from when he was a partner at Index Ventures. For his deliberations with Wang, it went beyond just the deal itself. “A lot of the discussion wasn't so much about the exact mechanics of the transaction,” he says. “It was, what does Alex want to accomplish with his life?”

The $14 billion deal was announced last June, and investors and employees alike received a big windfall. Most employees received a dividend that was roughly equal to half of their equity package, a source familiar with the deal said.

Some employees were understandably upset by Wang’s decision to depart. “I definitely think there's a lot of emotion when stuff happens,” Wang says when asked if he had any feelings of guilt over leaving. “I think I always had confidence that the team was going to be able to do incredible work.” Layoffs followed as well. Last August, two months after the deal was announced, Scale said it was axing 200 full-time employees, roughly 14% of its workforce, and discontinuing work with 500 contractors (part of Scale’s corporate contract workforce, not to be confused with its data labelers who work per contract). Now the company has about 1,300 full-time employees, and plans to hire another 500 by the end of the year, a Scale spokesperson said.

E ven before the Meta deal had been signed, Scale knew the transaction would complicate its relationships with the frontier labs. “We definitely anticipated turbulence. There's no question,” says Droege. “I mean, the facts of the situation are, our founder went to another lab to help them.” He adds, “Most of them have come back [as customers], but not all of them.”

To assuage fears that Scale would be in the pocket of Meta, Droege had to undergo a charm offensive to put customers at ease. Some clients wanted to visit Scale’s office to meet face-to-face with leadership and researchers. Others wanted a point-by-point breakdown of the parameters of the deal. Meta would not be getting a board seat, Droege emphasized to clients, and while Wang would remain on Scale’s board, no board member would or ever had knowledge of Scale’s work with customers, he said. “I am a shareholder of Scale,” Droege says now. “My incentive is to make sure that the non-Meta customer is as happy as possible. I am not a Meta employee.”

The most high-profile loss was OpenAI, according to two people familiar with the business relationship. It was a blow not just because of OpenAI’s industry stature, but also because it was the frontier lab that first gave Scale its start in generative AI when Scale began training GPT-3, even before ChatGPT’s launch in 2022 ignited the global AI frenzy. Meanwhile, Google initially ditched Scale as a data labeling vendor after the Meta announcement, but resumed its work a few months later, according to two people familiar with the situation.

One customer, of course, Scale didn’t have to worry about losing: Meta, where Scale played a major role in training Muse Spark, the first model built under Wang’s stewardship. Unveiled last month, it’s so far gotten mixed to favorable reviews, but has been seen as a genuine comeback for Meta in the model wars. “It’s been a labor of love,” Wang says of the new model. “Scale was a very, very critical partner. The success of Muse Spark, I think, is a testament to great work across many of our partners, but Scale is certainly a big part of that.” Beyond that, Wang declined to discuss his work at Meta.

Meanwhile, Scale has attempted to upend its business since the Meta deal. Prior to the tie-up, Scale’s focus was 70% on data labeling and 30% building apps for enterprises and governments, Droege says. Now he’s flipped the focus, he says, with annualized revenue from the apps business tallying $200 million at the end of last year. The company has also added new enterprise customers, including the Mayo Clinic, BP and Allianz. The Mayo Clinic, for example, developed a system with Scale that reads and interprets fragmented medical records. Ernst & Young, meanwhile, uses Scale to build its internal AI agents, including one that helps it perform due diligence for clients who have hired the accounting firm to guide it through M&A deals, says Tony Qui, chief technology officer of the firm’s strategy consulting arm. Qui says the company chose Scale because of a recommendation from OpenAI. (It was a few weeks before the Meta deal, which at first caused some concern, but Qui decided to continue with Scale after meeting with Droege and other company leaders.)

Scale claims its perch as both an enterprise vendor and a data labeler is a unique advantage. Since it works with frontier labs to train their models, it knows what the cutting edge of the technology looks like — and it’s what enterprises usually want to implement soon after. “Scale is the only one that really can offer both of that,” says Coatue partner Lucas Swisher, who led the firm’s Series B investment into the company.

Scale has also doubled down on its business working with the U.S. government — continuing the effort that began under Wang. Since then, it’s led Project Thunderforge, the Pentagon’s effort to integrate AI agents into mission planning, with a $500 million contract the department awarded Scale last week. Scale was also chosen as a contractor last month, along with Palantir and Anduril, for Golden Dome, President Trump’s $185 billion effort to create a missile defense shield, but both Scale and the DoW declined to discuss specifics of Scale’s role in the project.

“We still can, in some sense, double dip on the investment.” Mike Volpi, early Scale investor

Scale’s work with the DoW has shined particularly in data labeling, says Cameron Stanley, chief digital and artificial intelligence officer for the DoW, going back to its work in computer vision for Project Maven, an effort to bring machine learning into military intelligence workflows. The company’s practices in organizing metadata are “top-notch,” he said. But Scale’s particular talent is in allowing the Pentagon to build off of disparate datasets, as well as understanding how to work within the department’s bureaucracy to push projects forward. “Their ability to pull in vast amounts of different data and make sense of that data, and then structure it in a way that we can actually train algorithms off of it, it's pretty unique,” Stanley told Forbes . He declined to specify whether or not Scale’s tools had been used in military combat. A Scale spokesperson declined to disclose specifics on its work with the DoW.

Droege insists the reason for the shift away from data labeling isn’t because of a new ceiling created by clients not wanting to work with a vendor part-owned by Meta. Instead, it’s because the data labeling business more broadly has slowed its growth, he argues. The market for building enterprise apps, by contrast, is wide open as every big company tries to make its transition to an AI world.

Either way, Scale’s data labeling rivals, like Surge, Mercor, Handshake and Invisible, have pounced, and they haven’t been shy about flamethrowing. Mercor, for instance, hired a former Scale employee last year who allegedly stole trade secrets concerning Scale’s customer strategies. The lawsuit was settled in January for an undisclosed amount, according to a legal filing, which hasn’t been previously reported.

Since the Meta deal, one rival CEO anecdotally says Scale appears to have fallen out of the shuffle when it comes to data labeling contracts. “I haven’t heard of them in the past couple months,” they said. “They’re just not a name that comes up.”

Meanwhile, Scale is pushing forward with a potential IPO in the cards. It was one of the considerations when Scale did the deal with Meta in the first place. “We still can, in some sense, double dip on the investment,” says Volpi, allowing for the possibility of a “second exit” like the leftover company going public. Droege downplays the timeline. “Scale will very likely become a public company at some point,” he says, but adds that the company is “super early in the planning process.”

For now, the company is in uncharted territory. In the past few years, there have been a handful of similar deals where startups have made splashy aqui-hire deals with big tech companies. Founders from buzzy AI companies including Inflection, Adept, Character and Windsurf have collectively landed at Microsoft, Amazon and Google. Droege forcefully rejects being lumped into that category.

Why? “Because we actually had a business. Those companies didn't have a business,” Droege says. “And so I think there's a big difference.”