What is happening inside Meta's AI unit?

How many AI teams does Meta actually have now?

Meta’s AI setup is actually not one clean lab anymore.

It is now a layered organization, with a frontier-model group at the top and several support, research, product, and infrastructure teams underneath.

The main umbrella is Meta Superintelligence Labs, or MSL, led by Alexandr Wang. Inside it, the structure is becoming clearer: TBD Lab focuses on frontier-model training, FAIR remains the long-term research group, Products & Applied Research works on turning models into actual Meta products, and MSL Infra handles the compute and systems layer.

There is also Applied AI, a newer group formed in March 2026 to support MSL’s researchers. WIRED reported that it has about 6,500 engineers and product managers.

That is a large number. Meta had 77,986 employees at the end of March 2026, which means this one support unit would represent roughly 8% of the company.

It looks more like Meta created a large internal service layer to help a smaller group of elite AI teams move faster.

Meta has a high-profile frontier lab at the top, but underneath it there is now a large human system producing tasks, feedback, tests, examples, and support work for models.

So what is Alexandr Wang actually doing at Meta?

Alexandr Wang is now one of the people reshaping Meta’s AI organization.

In June 2025, Scale said Meta made a major investment valuing Scale at more than $29 billion. Scale also said Wang would join Meta to work on AI, while staying on Scale’s board. Meta’s own leadership page now lists him as Chief AI Officer.

After that came the reorg. Wang’s memo split the superintelligence effort into research, training, product, and infrastructure. TBD Lab, the smaller team focused on training and scaling large models, sits closest to the frontier-model race. FAIR is now more directly connected to model development. Products and Applied Research is supposed to bring models into Meta’s apps. Infra exists because this entire strategy depends on massive GPU and data-center capacity.

Wang’s arrival did not simply add another senior executive. Instead, it changed the operating model. Meta has moved from a more open, research-heavy AI culture toward a tighter, faster, more centralized model-building culture.

Did Meta basically buy Scale to fix its data problem?

Meta clearly wanted Scale’s data expertise, but the deal did not solve the whole data problem by itself.

The official transaction already says a lot. Scale said Meta’s investment valued the company above $29 billion and would “substantially expand” the commercial relationship around data solutions. That matters because Scale’s business is built around data, evaluation, and human feedback for AI systems. Meta did not just hire an AI executive but also brought in the founder of one of the most important AI-data companies and deepened the commercial relationship at the same time.

But the follow-up signal is more revealing. TechCrunch later reported that Meta’s TBD Lab was also relying on Scale competitors, including Mercor and Surge, for next-generation model-training work. Some researchers reportedly preferred the quality of those vendors’ data. Meta disputed the idea that Scale quality was the issue, so that claim should not be overstated.

Still, the pattern is meaningful. If a company spends more than $14 billion on a strategic Scale relationship and still works with rival data vendors soon after, it suggests the bottleneck is getting the exact kind of high-quality data needed for a specific model run.

That is the deeper point. In frontier AI, “data” no longer just means a huge pile of internet text. It’s more about difficult coding tasks, reasoning problems, expert feedback, agent behavior traces, safety tests, and edge cases that expose model weaknesses.

Meta is now trying to secure that supply chain from several directions at once: Scale, rival vendors, internal employees, and product usage.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Are Meta engineers really being asked to do data-labeling work?

Yes, but “data labeling” is not quite the right phrase.

According to WIRED, it is closer to expert task-writing and model evaluation. Some Applied AI employees are reportedly asked to generate puzzles that test how reliably models solve problems.

Others are asked to produce complex coding problems that help AI scientists train and evaluate frontier models. Some employees are expected to complete two tasks per week.

That sounds modest until we look at who is doing it. These are engineers and product managers who may have been building Instagram, infrastructure, ranking systems, ads, messaging, or developer tools.

Now some of them are creating training and evaluation material for AI models. The work is more skilled than old-school click-labeling, but it can still feel like a demotion if your previous job was building products at global scale.

There is another signal pointing in the same direction. Meta’s Model Capability Initiative, reported by Reuters and WIRED, records employee screen activity, clicks, keystrokes, and computer behavior in certain work contexts to train AI agents. Employees pushed back hard, and WIRED reported that an internal protest post was seen by nearly 20,000 coworkers.

So the direction is clear: Meta is turning employee knowledge and employee behavior into AI training material.

Why would Meta put strong engineers on puzzle tasks?

Because the easy data is no longer enough.

For older AI systems, a lot of improvement came from more general text, labels, and human preference data. Today, the frontier fight is more specific. Models need better reasoning tests, harder coding tasks, better agent traces, richer safety evaluations, and examples of how humans actually complete work on a computer.

Meta’s own Muse Spark evaluation materials show that the company tests the model across reasoning, multimodal ability, coding, tool use, and health knowledge.

These are not areas where generic low-quality data helps much. If a company wants a model to improve at coding, it needs hard coding tasks. If it wants agents to use software, it needs examples of software use. If it wants reasoning models to improve, it needs tasks that reveal where reasoning breaks.

That is why the work can be strategically useful and frustrating at the same time. From the model team’s perspective, a good engineer writing two hard coding problems per week can be valuable. From the engineer’s perspective, it can feel strange: you joined Meta to build products at scale, and now your work becomes input for a model team.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Is the “AI gulag” line real or just drama?

The phrase is dramatic. The frustration behind it looks real.

WIRED quoted one Applied AI employee calling the unit “literally the gulag.” That should not be treated as a literal description of working conditions. It is an anonymous employee using extreme language.

But the surrounding facts help explain why people are angry: the unit was formed recently, has about 6,500 people, includes engineers and product managers, and some employees say they had little real choice except to join or leave.

There are several morale signals pointing in the same direction. WIRED reported a livestream incident where an employee interrupted an internal presentation. It also reported that some Applied AI workers call themselves “draftees.” Across Meta, more than 1,600 employees reportedly signed a petition against the employee-tracking AI initiative. Another WIRED story said a protest post about the tracking tool had been seen by nearly 20,000 employees.

Those numbers matter because they show this is not just one colorful quote. The “gulag” phrase is the part that travels online, but the real story is forced redeployment, surveillance anxiety, layoffs, and employees feeling that AI is being imposed on them rather than built with them.

So, all things considered, this looks more like a morale crisis than a simple communication issue.

Is Meta’s AI unit actually producing anything these days?

Meta is producing real AI output, but the output still looks early compared with the size of the bet.

The clearest product signal is Muse Spark, launched in April 2026. Meta described it as the first model from Meta Superintelligence Labs. Axios reported that it was code-named Avocado, built over nine months, and designed to narrow the gap with OpenAI, Anthropic, and other leading labs. It powers Meta AI in the app and on meta.ai, with expansion planned across Facebook, Instagram, WhatsApp, and other surfaces.

But the same reporting also made an important point: Muse Spark does not set a new state of the art, and Meta acknowledges gaps in areas like coding. That fits the broader picture. Meta is now shipping, but it is still trying to catch up in parts of the frontier-model race.

Meta’s own evaluation framing also shows where the company wants to be judged. It evaluates Muse Spark across reasoning, multimodal performance, coding, tool use, and health knowledge. That is basically the modern frontier-model scorecard. In Q1 2026, Zuckerberg also highlighted the “release of our first model from Meta Superintelligence Labs” in Meta’s earnings release.

So Meta AI is not just reorg charts and ambition. It is actually shipping, yes.

But today, the visible product is still the first serious output of the new organization, not proof that Meta has already overtaken the frontier labs.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Why is Meta spending so much if the product is still early?

Because Meta is trying to buy time, compute, talent, and data all at once.

The Q1 2026 numbers are huge. Meta reported $56.31 billion in quarterly revenue, up 33% year over year. It also reported $19.84 billion in capital expenditures in just one quarter. For the full year 2026, Meta now expects capex of $125 billion to $145 billion, up from the previous range of $115 billion to $135 billion.

That is far beyond a normal “let’s add some AI features” budget. Meta is using the cash machine from its ads business to build AI infrastructure at massive scale. The reason it can do that is also visible in the numbers: Meta had 3.56 billion daily active people across its family of apps in March 2026, and ad impressions were up 19% year over year. The old business is still strong enough to fund the AI race.

But that creates pressure inside the company. Once spending reaches this level, AI stops being a side initiative. It becomes the priority around which other teams are reorganized. That helps explain the elite hiring, the data deals, the task programs, the employee-tracking initiatives, and the push to bring Muse Spark into Meta’s apps.

So the spending is not just a financial detail. It explains why the internal transition feels so intense.

When a company raises capex guidance by $10 billion in one quarter, the rest of the organization has to adapt quickly.

What is proven, and what is still only reported?

The cleanest way to read the situation is to separate official facts, strong reporting, and contested claims.

The official facts are solid. Alexandr Wang joined Meta as Chief AI Officer. Scale confirmed Meta’s investment valued Scale above $29 billion and said the commercial relationship with Meta would expand. Meta confirmed Muse Spark as its first model from Meta Superintelligence Labs. Meta reported Q1 2026 revenue of $56.31 billion, capex of $19.84 billion, 77,986 employees, and a 2026 capex range of $125 billion to $145 billion.

Then there is strong reporting.

WIRED says Applied AI was formed in March 2026, has about 6,500 engineers and product managers, and includes employees doing puzzle and coding-problem tasks for model training and evaluation.

WIRED also reported the internal livestream incident, the “draftee” language, the 1,600-plus employee petition against tracking, and the 50-to-one manager ratio in some teams.

The more delicate part is the vendor-quality story. TechCrunch reported that some TBD Lab researchers preferred data from Scale competitors such as Surge and Mercor, while Meta pushed back on the idea that Scale data quality was the issue. So we can say there are signs of tension in Meta’s data supply chain. We should not say that Scale “failed” Meta.

Overall, the story is well supported, but not every detail has the same evidentiary status. The reorg and spending are official. The employee anger is strongly reported. The exact Scale-versus-rivals quality issue is still less settled.

Is Meta copying OpenAI, or doing something different?

Meta is borrowing the frontier-lab playbook, but it is applying it inside a very different company.

The OpenAI-style part is clear: a superintelligence lab, a small elite model team, huge compute, expensive hiring, reasoning models, multimodal models, and a race toward frontier capability. TBD Lab is the closest thing to Meta’s inner model room.

But Meta has something most AI labs do not have at the same scale: distribution across WhatsApp, Instagram, Facebook, Messenger, Threads, smart glasses, and billions of daily users. That changes the strategy. Meta does not only need a model that performs well on benchmarks but also a model that can work inside feeds, chats, creator tools, glasses, ads, and business messaging.

This is why Muse Spark matters even if it does not lead every benchmark. It is the first test of Meta’s real advantage: can it put AI directly into places where billions of people already spend time?

The tension is that Meta is trying to operate like a frontier research lab while still being a giant social and ads company. Those two cultures do not naturally fit. One rewards small elite teams, fast model runs, and technical intensity. The other has thousands of engineers, product roadmaps, privacy constraints, moderation issues, ads incentives, and consumer UX trade-offs.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Why are employees angry if Meta is doing so well financially?

Because Meta can be financially strong and culturally unstable at the same time.

That is exactly what the numbers suggest. Q1 revenue was up 33%. Operating margin stayed at 41%. Net income was up 61%, helped by a tax benefit. Daily active people across Meta’s apps reached 3.56 billion. From the outside, Meta still looks like an extremely strong business.

Inside the company, the experience appears different. Employees are dealing with layoffs, AI reorgs, redeployments, tracking software, new manager ratios, and teams being dissolved or merged. Business Insider reported that Meta had already dissolved two major AI teams in less than six months. WIRED reported that Zuckerberg admitted mistakes in recent organizational changes and promised more stability.

That gap is central to the story. Meta is not restructuring because the core business is collapsing but because the core business gives it enough money to chase AI aggressively. For employees, that can feel especially destabilizing. The company is healthy, but it is still asking people to accept a much harsher operating model.

In the end, Meta’s AI push is being funded by a very strong ads business, but the internal organization is being reshaped as if the company were in wartime.

So what is really going on at Meta AI right now?

Meta is trying to become an AI-first company faster than its organization can comfortably absorb.

The pieces now line up. Meta hired Wang from Scale. It created Meta Superintelligence Labs. It split the AI org into frontier training, research, product, and infrastructure. It launched Muse Spark. It raised 2026 capex guidance to $125 billion to $145 billion. It deepened its data relationship with Scale, while also reportedly working with other data vendors. It created a 6,500-person Applied AI unit to support model work. It also started collecting employee computer-behavior data for AI agents, then faced major internal backlash.

Taken together, this is not just “Meta doing AI.” It is a company-wide conversion. Meta is using money, people, compute, product distribution, internal work data, and vendor data to catch up in frontier AI.

The uncomfortable part is that the human layer is now visible. We usually talk about AI through chips, models, benchmarks, and apps.

Here, we can see the people in the middle: engineers writing coding tasks, employees being tracked to train agents, managers absorbing very large teams, researchers deciding which data is good enough, and workers wondering whether they are building the future or being turned into input for it.

So today, the clearest answer is this: Meta AI is not empty, and it is not quietly failing. It is moving fast, spending massively, shipping early models, and creating real internal strain along the way.

The immediate news is the backlash inside Applied AI, but the bigger story is that Meta is learning that superintelligence is not only a model problem. It is also an organizational problem.

If you want more recent data on this point, please see our latest AI infrastructure market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.