What Does Scale AI Do? Business Model, Meta Deal, Fallout

Ask ten people in tech “what does Scale AI do” and you’ll get ten different answers: a data company, a labeling shop, “the RLHF people,” a defense contractor, or “the company Meta bought a chunk of.” All of them are partially right, which is exactly why Scale AI is worth understanding properly. It sits in a layer of the AI industry that almost nobody outside it thinks about, and 2025 turned it into a case study in vendor risk that every company buying AI tools — not just frontier labs — should read closely. If you’re evaluating any AI vendor, including the ones powering your own AI chat agent, the Scale AI story is a useful lens.

This piece answers the plain question first — what Scale AI actually does and how it makes money — then walks through the Meta deal, the customer exodus that followed, and what the whole episode teaches about picking AI vendors you can trust.

What does Scale AI do?

In plain English: Scale AI turns raw, messy data into the labeled, ranked, and reviewed data that AI models are trained and tested on. It doesn’t build foundation models. It builds the unglamorous pipeline that makes those models usable, and every major lab depends on some version of it.

Four things sit at the core of what Scale AI does:

• Data labeling and annotation. Humans tag images, transcribe audio, classify text, and draw bounding boxes so supervised models have ground truth to learn from.
• RLHF and preference data. Reinforcement Learning from Human Feedback is the process that turns a raw language model into something that follows instructions and doesn’t ramble. Scale’s contractors write ideal responses, rank multiple model outputs against each other, and flag unsafe or low-quality completions. That preference data is what gets fed into the fine-tuning loop.
• Model evaluation. Before a lab ships a new model version, someone has to test it against hard prompts, edge cases, and adversarial inputs. Scale runs structured evaluation programs for this.
• The Data Engine. Scale’s packaged enterprise product: infrastructure plus workforce plus tooling, sold to companies that need continuous data pipelines rather than a one-off labeling job.

The buyers are frontier AI labs — the source of most of Scale’s revenue and reputation — plus enterprises building their own models and government agencies. None of them can fully skip this layer. A model is only as good as the examples it was trained and corrected against, and there’s no way around needing humans in that loop somewhere, whether it’s your own team or a vendor’s.

How Scale AI makes money

Scale AI’s business model has three interlocking pieces. First, a contractor workforce, recruited and managed through platforms called Remotasks and Outlier, performs the actual labeling, ranking, and review tasks at scale. Second, a software platform that manages task distribution, quality control, and delivery to clients. Third, direct enterprise and government contracts, where Scale runs dedicated data programs for large customers rather than selling shelf product.

That combination generated roughly $870 million in revenue in 2024, and it’s why “AI training data” went from a research footnote to a real industry with venture-backed competitors in under three years. Every lab racing to ship a better model needs more, better, and more specialized data: code, multilingual text, multimodal content, expert-level domain knowledge. Building that pipeline in-house is expensive and slow. Outsourcing it to a company that already has the workforce and tooling was the obvious move, at least until ownership questions made it complicated.

The economics have a catch that rarely makes the headlines. Because so much of the cost is human labor rather than software, margins in data labeling are structurally thinner than in most enterprise SaaS. Growth means recruiting, training, and quality-controlling more people, not just spinning up more servers. That shapes everything about how the company behaves when a large customer leaves. If you’re curious how labeled data eventually turns into a working retrieval system rather than just training weights, our piece on RAG for customer support knowledge bases covers the adjacent problem of getting clean, well-structured data into a live AI product.

Who is Scale AI and why is it important?

Scale AI matters because of where it sits in the AI supply chain, a position most buyers never think about because it’s two layers removed from anything they interact with directly.

Layer	What it does	Example players
Compute	Chips and infrastructure that run training and inference	Nvidia, AMD, cloud hyperscalers
Foundation models	Trains and ships the underlying large language models	OpenAI, Anthropic, Google, Meta, xAI
Training data	Labels, ranks, and evaluates the data models learn from	Scale AI, Surge AI, Mercor, Labelbox
Applications	Products built on top of foundation models for end users	Chatbots, copilots, agent platforms

Most people who ask who is Scale AI and why is it important are really asking why a company they’d never heard of before 2025 was suddenly worth $29 billion to Meta. The answer is that Scale AI is a chokepoint. If you’re a lab racing to ship the next model, you don’t have time to build a large annotation workforce from scratch. You rent one.

That dependency is invisible to end users of any AI application, whether it’s a research assistant or the kind of retrieval-driven support tool covered in our agentic AI platforms overview. But it matters. A bottleneck in the data layer quietly shapes what every layer above it can ship, and how fast.

The Meta deal that changed everything

On June 12, 2025, Meta announced a $14.3 billion investment in Scale AI for a 49% non-voting stake, valuing the company at roughly $29 billion. The structure mattered. Non-voting meant Meta wasn’t formally taking control of the board, but a stake that large in a company that size is not a passive financial position by any reasonable reading.

The more disruptive part of the announcement was personnel, not equity. Scale AI’s founder and CEO, Alexandr Wang, left to lead Meta’s newly formed superintelligence effort. He stayed on Scale’s board, which meant the person who built the company was now simultaneously working for one of its biggest customers and, overnight, one of the biggest rivals to every other customer.

From Meta’s side, the logic is straightforward: buy access to the data pipeline and the talent that understands it best, and accelerate an internal AI push that had been lagging competitors. From everyone else’s side, meaning every other lab that had been sending Scale its most sensitive prompts, model outputs, and evaluation data, the calculus flipped overnight.

Why Scale AI’s customers walked away

Within days of the announcement, cracks showed. Google, reported to be Scale’s largest customer at around $150 million a year, moved to end the relationship. OpenAI scaled back its work with Scale significantly, though its CFO later said the company continues to work with Scale as one vendor among several rather than cutting ties completely. xAI and Microsoft were also reported to be pulling back.

The mechanism here matters more than the headline. This wasn’t a proven data breach or a documented case of Meta accessing a competitor’s proprietary information through Scale’s systems. Nothing in the public record supports that claim. It was a perceived conflict of interest, and a perceived conflict of interest doesn’t require proof to be commercially fatal. It requires plausibility.

The customers who left didn’t need evidence that Meta could see their data. They needed a scenario a competitor could raise in a boardroom. Once that scenario exists, procurement teams route around it by default, whether or not it was ever true.

Consider what actually flows through a vendor like Scale: your model’s failure modes, the prompts you’re most worried about, early access to unreleased model behavior for evaluation, sometimes fine-tuning data that encodes your product roadmap. None of that needs to leak for the relationship to become untenable. It just needs to be plausible that a rival’s board member sits one conversation away from someone who could see it. That’s not a legal argument. It’s a trust argument, and trust doesn’t survive ambiguity at this level of competitive intensity.

The fallout: layoffs and a new CEO

The revenue hit showed up fast. In July 2025, Scale cut roughly 200 employees and around 500 contractors, and restructured its generative AI unit from 16 groups down to 5. That’s a sharp contraction for a division that had presumably been staffed for a growth trajectory the Meta deal disrupted rather than accelerated.

Leadership was in flux for more than a year. Jason Droege stepped in as interim CEO after Wang’s departure to Meta. Then, according to an Axios report dated July 30, 2026, Francis deSouza was named CEO. That report is recent enough to treat as reported rather than settled fact until more sourcing accumulates. If accurate, it marks the company’s move from crisis management toward a more conventional leadership structure, and signals that Scale is trying to look stable to customers who are still deciding whether to trust it.

What does Scale AI do in defense and government?

While the frontier-lab business took a hit, Scale leaned harder into a segment largely insulated from the Meta fallout: government and defense. Scale runs Scale Donovan, a generative AI platform built for defense and intelligence analysis, and holds contracts with the Department of Defense’s Chief Digital and Artificial Intelligence Office. Reporting put an initial CDAO agreement at roughly $100 million in September 2025, later reported as expanded to a $500 million ceiling by May 2026. Those figures are worth attributing to reporting rather than treating as confirmed public financials.

Government work is structurally stickier than frontier-lab contracts for reasons that have nothing to do with data quality. Procurement cycles are long. Security clearances create switching costs that dwarf a normal vendor contract. And a defense agency isn’t worried about a competing AI lab reading its data over Scale’s shoulder, because its threat model is different. For a company that just watched several of its biggest commercial accounts walk out the door over an ownership perception problem, a customer base immune to that specific problem looks like a rational place to direct growth.

What Scale AI’s story teaches about vendor risk

Here’s the part that applies well beyond frontier AI labs. Most vendor-risk conversations focus on price, uptime, and support SLAs. Scale AI’s 2025 shows that concentration risk can flow upward from ownership, a variable almost nobody puts on a vendor scorecard, because almost nobody audits who owns their vendor’s vendors.

You probably don’t ask your CRM provider who their board members work for. You probably don’t ask your analytics platform whether a competitor holds equity in it. Scale’s customers didn’t either, until it became impossible to ignore. The lesson isn’t “avoid vendors with outside investors,” which would rule out most of the software industry. The lesson is that ownership structure is a real input to vendor risk, and it’s worth a five-minute check before it becomes a forced migration.

A short checklist for any AI vendor you’re evaluating, not just data-labeling shops:

Question	Why it matters
Who owns significant equity, and do they compete with you?	Determines whether your data ever sits one relationship away from a rival
What does it cost to leave, in time and engineering effort?	Low switching cost turns a red flag into a Tuesday. High cost turns it into a crisis
Can you export your data in a usable format, on demand?	Portability is what makes “we could leave” a credible position instead of a bluff
Is the underlying provider (model, data source) substitutable?	Single-provider lock-in concentrates risk even without an ownership issue
Where does the data physically live, and under whose jurisdiction?	Ownership changes can move data-processing terms you already signed off on

Two of those questions, exit cost and substitutability, come up constantly when teams evaluate AI vendor pricing and lock-in more broadly. We cover them in detail in our breakdown of AI agent platform pricing models and our look at no-code AI agent builders and their exit costs.

How to de-risk your own AI stack

None of this means every AI vendor with outside investors is a liability. It means concentration risk is worth designing against before you’re forced to, not after. A few concrete moves:

• Keep sensitive data on infrastructure you control. If the data involves customer PII, proprietary prompts, or competitive information, self-hosting removes the ownership question entirely. There’s no vendor board to worry about because there’s no vendor sitting between you and your data.
• Keep the model layer swappable. If your AI product is hard-wired to a single provider’s API and that provider’s ownership or pricing changes overnight, you inherit their instability. An architecture that treats the LLM as a replaceable component, not a foundation, absorbs that shock instead of transmitting it.
• Avoid vendors that own your data format. If leaving means re-labeling, re-indexing, or rebuilding your knowledge base from scratch, your choice to switch providers was never real.
• Write the exit path down before you sign. Not as a threat, as an engineering estimate. If nobody can say how long a migration would take, you don’t have an exit path.

That last point is where most teams discover the gap. Portability is easy to promise and hard to verify, and the test is boring: export everything, and see whether the export is usable somewhere else without a rewrite.

This is the pattern behind how we built AI Chat Agent, and it’s worth naming plainly rather than burying it. It’s a self-hosted AI chatbot widget, deployed with Docker Compose, where your knowledge base and chat history live in your own PostgreSQL database. It connects to five AI provider options: OpenAI, Anthropic Claude, Google Gemini, OpenRouter, or any OpenAI-compatible endpoint including Groq, Ollama, or a self-hosted model. Switching between them requires no data migration, because the data was never inside the provider’s walls to begin with. Retrieval runs on hybrid search, pairing pgvector dense search with Postgres full-text lexical search and LLM reranking, so the quality of what the model sees doesn’t depend on which provider is answering that day.

We go deeper on the architecture trade-offs in self-hosted vs SaaS chatbots, on the mechanics of running more than one model provider in building a multi-LLM chatbot, and on how the major providers actually compare for support work in our OpenAI, Anthropic and Gemini comparison. If you’re currently on a SaaS incumbent, the Intercom comparison walks through what portability looks like in practice, and the Chatbase comparison covers the same ground for knowledge-base bots.

Is Scale AI still a good partner?

The honest answer is that it depends on your position, not on Scale’s data quality. Reputationally, Scale’s labeling and evaluation work is still regarded as strong. That’s why OpenAI kept it as one vendor among others rather than dropping it outright, and why the defense pipeline is growing rather than shrinking. The underlying capability didn’t degrade in June 2025. The trust structure around it did.

If you’re an enterprise buying data services and you don’t compete with Meta in any meaningful way, the ownership question is close to irrelevant to you, and Scale’s technical track record stands on its own. If you’re a frontier lab, or any company whose competitive edge depends on model behavior a rival could plausibly learn about through a shared vendor, the calculus is different, and it should probably stay different regardless of who’s CEO. Ownership structures don’t reset just because leadership does.

There’s also a middle path that gets overlooked: use the vendor for the work that isn’t sensitive, and keep the rest in-house. Not every dataset carries strategic signal. Splitting the work by sensitivity rather than by convenience is usually cheaper than either extreme, and it keeps a second supplier warm.

We write about vendor evaluation and AI architecture decisions like this regularly. Browse the rest of the analysis on the getagent.chat blog if you’re mid-evaluation on a similar question, or start with the GDPR compliance guide if data jurisdiction is the part that worries you.

Frequently Asked Questions

The broader takeaway from Scale AI’s year isn’t “don’t work with data vendors.” It’s “know what you’re building on, all the way down.” If part of your stack touches customer data, proprietary prompts, or anything you’d rather a competitor never see, run the ownership check before you run the pricing comparison. If you want to see what a fully self-hosted, provider-agnostic setup looks like in practice, the live demo is open, or you can go straight to a one-time €79 license: no monthly fees, full source code, and your data never has to leave infrastructure you control.