Okara
Start Free
Back to Blog
By · Published October 28, 2026 · 23 min read

Scale AI Valuation and Revenue: How It Grew to $29B

Scale AI was valued at $29B in Meta's June 2025 deal, with no new round as of October 2026. See every round, revenue by year and the 10 moves behind it.

Part of our Startup Growth Case Studies guide.

Scale AI is valued at $29 billion.

That figure comes from Meta's investment in June 2025, which Scale announced as valuing the company at "over $29 billion." Meta paid $14.3 billion for a 49% stake, Reuters reported. The stake carries no voting power, a Scale spokesperson told CNBC. As of October 2026, Scale has not announced a new funding round, so $29 billion is still its latest confirmed valuation.

Revenue is smaller than the valuation suggests. Scale told Forbes in May 2026 that it brought in just under $1 billion in 2025, up from $870 million in 2024. For 2026, a Scale spokesperson told The Verge that the company projects more than $1 billion.

Before Meta, Scale raised about $1.6 billion in venture funding. That covers roughly $600 million before 2024, according to TechCrunch, plus a $1 billion Series F that mixed new money with share sales.

The short version of why it grew: Scale turned messy human labeling work into an API, became the default data supplier for self-driving cars and then for large language models, and sold to the Pentagon early. The Meta deal then took its founder and drove away some of its largest customers. Scale has spent the 16 months since rebuilding around government and enterprise AI applications.

Here is how it happened.

Scale AI's growth in one table

DateMilestone
Summer 2016Alexandr Wang and Lucy Guo start Scale at Y Combinator as an "API for human labor"
July 2017$4.5M Series A led by Accel
2018$18M Series B led by Index Ventures; self-driving car companies are the main customers; Guo leaves
August 2019$100M Series C led by Founders Fund at a valuation above $1B
September 2019OpenAI uses Scale's labelers to fine-tune GPT-2 with human feedback
December 2020$155M Series D led by Tiger Global at more than $3.5B; Wang says Scale is breaking even
April 2021$325M Series E at a $7.3B valuation
January 2022Pentagon AI testing agreement with a $249M ceiling
January 2023Scale lays off 20% of staff
May 2023Launches Donovan, its AI platform for the military, on a classified Army network
Late 2023Annualized revenue run rate reportedly reaches $750M, up from about $250M in early 2022
May 2024$1B Series F led by Accel at a $13.8B valuation
2024Revenue of about $870M; $1.5B run rate reported at year-end
March 2025Wins the Pentagon's Thunderforge AI planning prototype
June 2025Meta invests $14.3B for 49% at a $29B+ valuation; Wang joins Meta; Google, OpenAI and xAI pull back
July 2025Scale cuts 200 employees (14%) and 500 contractors
August–September 2025$99M Army contract and $100M Pentagon agreement
2025Revenue just under $1B, according to Scale
April 2026Acquires defense analytics firm ICG Solutions
May 2026Pentagon raises its Scale agreement ceiling to $500M
August 2026Former Google Cloud COO Francis deSouza becomes CEO
October 2026Air Force agentic AI contract grows to $44.3M; Scale launches a post-training service with Nvidia

1. Scale AI started as an API for human labor

Alexandr Wang was 19 when he started Scale, and his co-founder Lucy Guo was 21, Forbes reported in 2023. Wang had spent one year at MIT, and both had worked at Quora.

The company joined Y Combinator's Summer 2016 batch. The first pitch was broad. In a July 2016 interview, Y Combinator described Scale as "an API that lets companies outsource manual tasks," where one line of code could hire a human for jobs like data extraction or categorization.

Guo explained the problem from her time at Quora and Snapchat. Both relied on people to review flagged content by hand, and she wished there was an API to do it.

Scale was a mid-batch pivot. Accel partner Daniel Levine wrote that he heard about it after Justin Kan "accidentally launched the company" on Product Hunt, and that Scale "really was just two founders, an idea, and a tiny bit of traction" when Accel led the Series A.

That $4.5 million round closed in July 2017. In his announcement, Wang said Scale had released six APIs, including image annotation and transcription, and was processing millions of API requests a month.

2. Self-driving cars made Scale the default labeler

The first big market was autonomous vehicles.

Self-driving cars collect huge amounts of camera, radar and lidar data. Every pedestrian, cyclist and truck in that data has to be labeled before a model can learn from it. Scale combined software with a large contract workforce to do that labeling quickly.

When Scale raised an $18 million Series B led by Index Ventures in 2018, Wired reported that its customers included Cruise, Zoox, Lyft, Nutonomy, Nuro, Pony.ai, Voyage, Embark and Starsky Robotics. The company had just 35 employees but drew on more than 10,000 contract workers.

Customers described Scale as core infrastructure. "Scale is basically providing the ground truth for our perception systems," Lyft's head of self-driving software told Wired.

The business grew fast enough to attract growth investors. In August 2019, Scale raised $100 million at a valuation of more than $1 billion, led by Founders Fund.

Wang, then 22, told TechCrunch that "ideally this is the last fundraise for us." Scale would go on to raise three more rounds.

3. Scale took OpenAI's unusual labeling job in 2019

Self-driving cars paid the bills, but a small project with OpenAI turned out to matter more.

In 2019, OpenAI fine-tuned GPT-2 using human feedback. Labelers picked the best of four model-written continuations, and OpenAI trained the model on their choices. The research paper states plainly: "We use Scale AI to collect labels."

The work was harder than labeling images. Summarization alone took 60,000 human comparisons, and the labels had to be collected while the model trained, OpenAI wrote. The answers were also subjective, so even the paper's authors agreed with each other only about 60% of the time.

The technique is now known as reinforcement learning from human feedback, or RLHF.

Fortune later noted that Scale helped OpenAI train GPT-2 with the team that went on to found Anthropic. When demand for RLHF exploded in 2023, Scale had already built the systems for it.

4. Scale went to Washington early

In 2020, Scale expanded into government and military work, according to Fortune, starting with systems for government geospatial data. Wang has said a 2019 trip to China pushed him toward national security.

Investors kept paying more. In December 2020, Tiger Global led a $155 million Series D at a valuation above $3.5 billion. Wang told CNBC that Scale was breaking even and had more than doubled its annualized revenue run rate year over year in the third quarter.

Four months later, Scale raised $325 million at a $7.3 billion valuation, co-led by Dragoneer, Greenoaks and Tiger Global, Fortune reported. The Defense Department and Air Force were already customers, according to Reuters.

Contracts followed. In January 2022, the Pentagon's Joint Artificial Intelligence Center picked Scale as a government-wide AI testing partner under a $249 million-ceiling agreement. In May 2023, Scale launched Donovan and deployed it on a classified network for the Army's XVIII Airborne Corps.

The government business started small. Forbes found that Scale earned $90 million from public government contracts between 2020 and the third quarter of 2023 and spent $1.95 million on lobbying over the same period. But it gave Scale a second business that would matter enormously after 2025.

5. Scale rode the RLHF boom when self-driving work cooled

The 2021 valuation came just before a downturn.

In January 2023, Wang laid off 20% of staff. "Many of the industries we serve, such as e-commerce and consumer technology, have been buoyed by the pandemic and are now experiencing a painful market correction," he wrote, as quoted by TechCrunch. He said he took "full responsibility" for overhiring.

ChatGPT had launched two months earlier. Every company building a large language model suddenly needed people to write examples, rank answers and correct mistakes.

Scale's revenue reflected that shift. Its annualized revenue run rate reached $750 million in late 2023, up from about $250 million in early 2022, Forbes reported, citing two people familiar with the matter. Forbes attributed the growth to "skyrocketing demand" for RLHF.

By May 2024, 90% of Scale's business came from generative AI spending, Fortune reported. Wang said annual recurring revenue had tripled in 2023 and was expected to reach $1.4 billion by the end of 2024.

That same month, Scale raised a $1 billion Series F at a $13.8 billion valuation, led by Accel. New investors included Amazon and Meta, and Nvidia returned. The round mixed new capital with secondary sales.

6. Scale replaced crowds with experts

The RLHF boom changed what good data meant.

Early labeling work could be done by large, low-cost crowds. Much of Scale's crowd worked through its subsidiary Remotasks, which the Washington Post in 2023 described as low-paying work with delayed pay, Fortune noted. Scale said payment delays were "exceedingly rare."

Language models needed harder examples. Scale started recruiting PhDs, lawyers, accountants, poets and native speakers through another subsidiary, Outlier, at roughly $30 to $60 an hour, according to job listings cited by Fortune.

In 2024, Scale told Remotasks workers that all generative AI work was moving to Outlier, "the human expert platform," while Remotasks would focus on image, video and self-driving data. Outlier's own site now claims more than 900,000 experts and more than $500 million paid out.

The labor model drew scrutiny. Former workers sued, alleging they were underpaid and misclassified as contractors, and the Labor Department opened an investigation. The department dropped that probe in May 2025, TechCrunch reported.

7. Scale turned model testing into a product

Scale did not only sell training data. It also sold ways to measure models.

In 2023, Scale provided the evaluation platform for an AI red-teaming challenge at the DEFCON hacker convention, supported by the White House science office, Fortune reported. Wang said then that evaluation would become a key part of the business.

That work became SEAL, Scale's Safety, Evaluation and Alignment Lab. In 2025, SEAL introduced 15 new benchmarks and published more than 450 evaluations across more than 50 models, according to interim CEO Jason Droege. Scale says benchmarks such as Humanity's Last Exam and SWE-Bench Pro are now widely used to judge new models at launch.

Benchmarks give Scale public credibility every time a lab releases a model. They also feed a newer business: in September 2026, CEO Francis deSouza wrote that Scale has partnered with government AI evaluation bodies in the United States, the United Kingdom, Korea and Singapore.

8. Meta paid $14.3 billion for 49% and took the founder

By early 2025, Scale looked set for its biggest year. It generated about $870 million in 2024 revenue, exited the year at a $1.5 billion run rate, and expected $2 billion in 2025, Bloomberg reported in April 2025.

A tender offer at a $25 billion valuation was also in the works. In April 2025, Forbes reported that Scale was wrapping up a tender that would let early employees and investors sell shares at that price.

Then Meta made a bigger offer. On June 12, 2025, Scale announced a "significant new investment" from Meta that valued it at more than $29 billion, along with a "substantially" expanded commercial relationship. Wang would join Meta, stay on Scale's board, and hand the CEO job to chief strategy officer Jason Droege, a founder of Uber Eats.

Much of the money went to existing shareholders. Scale said it would distribute proceeds from Meta's investment to shareholders and vested equity holders, giving them "substantial liquidity" while letting them keep their stakes.

The deal also came with guaranteed revenue. Meta intends to pay Scale at least $450 million a year for five years, or more than half of its annual AI spend, whichever is less, Forbes reported, citing a person who worked on the deal.

Wang became Meta's first chief AI officer and now leads Meta Superintelligence Labs, according to Meta. In April 2026, his lab released Muse Spark, Meta's first major model since the deal, CNBC reported.

9. The labs walked away, and Scale cut back

The Meta deal created a neutrality problem. Scale's biggest customers were Meta's direct competitors in AI, and many did not want a Meta-linked vendor handling their training data.

Google, Scale's largest customer, planned to cut ties, five sources told Reuters the day after the deal. Google had planned to pay Scale about $200 million in 2025, one source said. Reuters reported that Microsoft, OpenAI and xAI also planned to step back.

OpenAI confirmed it was winding down its work with Scale. A spokesperson told CNBC the pullback had been underway for six to 12 months as OpenAI sought more specialized data providers, and was not caused by the Meta deal. Contractors told Business Insider that xAI had paused some projects too.

Even Meta spread its work around. In August 2025, TechCrunch reported that Meta's TBD Labs was also using rivals Mercor and Surge, and that some researchers there saw Scale's data as low quality. Meta disputed that.

Scale cut costs. In July 2025, it laid off 200 full-time employees, about 14% of its 1,400 staff, and ended work with 500 contractors, CNBC reported. Droege wrote that Scale had "ramped up our GenAI capacity too quickly over the past year."

His memo also confirmed Scale was not yet profitable and was on a "drive towards profitability," Business Insider reported.

The gap shows in the numbers. Scale expected $2 billion in 2025 revenue before the deal, and finished just under $1 billion.

10. Scale rebuilt around applications and government

Droege's main move was to shift investment away from data labeling and toward helping enterprises and governments build their own AI applications, Forbes reported. Named customers include Ernst & Young, Paramount and Cisco.

Defense led the way. In March 2025, the Defense Innovation Unit awarded Scale a prototype contract for Thunderforge, which uses AI agents for military planning, with Anduril and Microsoft as partners.

After the Meta deal, the awards got larger:

Several of these figures are ceilings, not guaranteed revenue. The $500 million agreement lets any Defense Department unit start projects with Scale without a new competition, but spending depends on individual orders.

International governments became a second engine. In February 2025, Qatar signed a five-year deal with Scale to deploy AI across government services, Reuters reported. Droege wrote that the international public-sector business doubled revenue in 2025, with the Middle East the fastest-growing region.

The pivot came with a new leader. On July 30, 2026, Scale named Francis deSouza CEO, effective August 10. DeSouza was chief operating officer of Google Cloud and previously CEO of Illumina.

Scale says its applications business is on track to overtake its data business within 18 months, Axios reported. In October 2026, it launched a post-training service with Nvidia that customizes open-weight models for enterprises such as EY and Skydance.

How fast did Scale AI grow?

Scale is private and does not publish audited financial statements. These are the clearest revenue figures available, with their status:

  • Q3 2020: Scale's annualized revenue run rate more than doubled year over year, and the company was breaking even, Wang told CNBC (company statement, no dollar figure).
  • Early 2022: about $250 million annualized run rate, according to Forbes (reported).
  • Late 2023: about $750 million annualized run rate, according to the same Forbes report (reported). Wang separately told Fortune recurring revenue tripled that year.
  • 2024: about $870 million in revenue, with a $1.5 billion run rate by year-end, Bloomberg reported (reported). Scale later repeated the $870 million figure to Forbes.
  • 2025: just under $1 billion, according to Scale (company statement). The company says it signed more than $1 billion in new business and its data unit turned profitable, Droege wrote.
  • 2026: more than $1 billion projected, according to a Scale spokesperson (company projection).

Across its whole history, Scale has brought in $2.5 billion in revenue, it told Forbes.

The run-rate figures and full-year figures measure different things. A run rate annualizes a recent month or quarter, so the $1.5 billion exit rate for 2024 never turned into $1.5 billion of 2025 revenue. Scale has not said how much revenue it lost after the Meta deal.

Valuation moved much faster than revenue. Scale went from about $1 billion in 2019 to $3.5 billion in 2020, $7.3 billion in 2021, $13.8 billion in 2024 and $29 billion in 2025.

Headcount went up and down with the business. Scale had about 450 employees in February 2022, according to TechCrunch, and 1,400 before the July 2025 layoffs. Droege's memo to "1,300+ employees" in May 2026 was reported by Axios.

Scale AI's growth loop

Scale's growth through 2024 can be reduced to one loop:

  1. Find the AI field with the most urgent data problem, first self-driving cars and then language models.
  2. Build software and a contractor network that can deliver labeled data faster than in-house teams.
  3. Win the leading labs in that field as customers.
  4. Use their work to build better tools, quality checks and benchmarks.
  5. Point to those customers and benchmarks to win the next field and the next government contract.
  6. Raise money at a higher valuation and repeat.

The loop's strength was also its weakness. Scale's value depended on being trusted by competing AI labs at the same time.

The Meta deal broke that neutrality. Since then, Scale has been building a different loop: use what it learned training frontier models to build AI applications for governments and large companies, where a Meta connection matters less.

What could slow Scale AI down?

Dependence on Meta. If Meta pays at least $450 million a year, as Forbes reported, that would be close to half of Scale's annual revenue. Scale declined to comment on that figure, and the commitment runs for five years.

Lost frontier labs. As of November 2025, OpenAI, Google and xAI were not listed as customers on Scale's website, CNBC noted. The CFO said Scale still works with "all the major AI labs and tech companies" but declined to name them.

Expert-data competition. Rivals like Mercor and Surge were built around highly paid experts from the start. Mercor's CEO said on a podcast that "Scale lost the focus on product, on scaling quality," Business Insider reported. Scale responded that its data quality metrics were "at record highs."

Contract ceilings are not revenue. The Pentagon's $500 million figure is a maximum. Actual spending depends on how many Defense Department units start projects.

Leadership turnover. DeSouza is Scale's third CEO since June 2025, after Wang and interim CEO Droege. He arrives as the company tries to change what kind of company it is.

An unclear exit. One pre-IPO broker told CNBC that investors do not know what Scale's future exit looks like, asking: "Or did it kind of already have its exit?" Scale says it has about $1 billion on its balance sheet and does not need to raise money soon.

Borrowing Scale AI's public benchmarks

Scale earns attention with every model launch by publishing evaluations that labs and developers cite. Okara's Writer Agent drafts strategy-led articles like that and can publish them. Its GEO Agent tracks up to 15 AI search prompts in ChatGPT, Google AI Overviews and Google AI Mode where you want that work to appear.

More startup growth stories

Frequently asked questions

What is Scale AI's valuation?

Scale AI's latest confirmed valuation is $29 billion, set by Meta's $14.3 billion investment for a 49% stake in June 2025. Scale described the deal as valuing it at "over $29 billion." That was up from $13.8 billion in its May 2024 Series F. As of October 2026, Scale has not announced a new funding round or tender offer.

How much revenue does Scale AI make?

Scale told Forbes it made just under $1 billion in revenue in 2025, up from $870 million in 2024. The company projects more than $1 billion in 2026. Before the Meta deal, Scale had expected about $2 billion in 2025 revenue, Bloomberg reported. Scale is private and does not publish audited financials.

How much funding has Scale AI raised?

Scale raised about $1.6 billion in venture funding before 2025, including a $100 million Series C in 2019, $155 million Series D in 2020, $325 million Series E in 2021 and $1 billion Series F in 2024. Meta then invested $14.3 billion in June 2025, and Scale said it would distribute proceeds from that deal to existing shareholders. Investors include Accel, Founders Fund, Index Ventures, Tiger Global, Nvidia, Amazon and Meta.

Who founded Scale AI?

Alexandr Wang and Lucy Guo founded Scale AI in 2016 at Y Combinator. Wang was 19 and had left MIT after one year; Guo left the company in 2018. Wang ran Scale until June 2025, when he joined Meta as chief AI officer.

Who owns Scale AI and who runs it now?

Meta owns 49% of Scale AI but has no voting power, according to Scale. Wang remains on Scale's board. Francis deSouza, formerly chief operating officer of Google Cloud, became CEO on August 10, 2026, succeeding interim CEO Jason Droege.

Did Scale AI lose customers after the Meta deal?

Yes. Reuters reported that Google, Scale's largest customer, planned to cut ties, and that Microsoft, OpenAI and xAI planned to step back. OpenAI confirmed it was winding down its work with Scale but said the decision predated the Meta deal. Scale laid off 14% of staff in July 2025 and has since shifted toward government and enterprise customers.

Who are Scale AI's customers?

Scale's customers include Meta, the U.S. Department of Defense and its Army and Air Force, and the government of Qatar. Enterprise customers named by Scale or in reporting include Mayo Clinic, BP, Allianz, Ernst & Young, Paramount and Cisco. Scale says it still works with major AI labs but has not named most of them since the Meta deal.

Is Scale AI going public?

Scale has not announced plans to go public or filed publicly for an IPO as of October 2026. In 2024, Wang told Fortune the company was "very thoughtful about IPO timelines and preparing." After the Meta deal, a pre-IPO broker told CNBC that investors are unsure what Scale's exit will look like.

Sources

Tags

Growth stories