SCALE AI

Alexandr Wang grew up in Los Alamos, New Mexico — the kind of place where your parents are nuclear physicists and competing in math olympiads is basically a team sport. He got into MIT on that trajectory, started studying computer science and mathematics, and then dropped out in 2016 at 19 to join Y Combinator with a co-founder named Lucy Guo.

The original idea was straightforward to the point of being almost boring: AI companies had a data problem. Not a lack of data — a lack of labeled data.

A self-driving car model doesn't just need millions of images of roads. It needs millions of images where someone has carefully drawn boxes around every pedestrian, cyclist, traffic cone, and pothole.

That work was being done manually, badly, slowly, and expensively by whoever the AI company could find.

Wang and Guo decided to build the infrastructure to do it properly. Scale AI started as a platform that connected AI companies with human contractors who could label images, transcribe audio, annotate text, and categorize anything that needed categorizing.

It wasn't the flashiest pitch in the YC batch. It was, however, immediately useful — and immediately profitable.

In the early days, the customers were self-driving car companies. Waymo, Lyft's autonomous division, Cruise — all of them needed annotated road data at scale, and none of them wanted to build the annotation pipeline themselves.

Scale AI charged per task, delivered faster than the alternatives, and quietly became infrastructure for an entire industry before most people had heard of it.

Lucy Guo left the company in 2017. Wang kept building.

By the time the broader AI wave hit in 2022 and 2023, Scale AI had already spent six years doing the work nobody wanted to talk about — and was perfectly positioned to become the company everyone suddenly needed.

Scale AI sits between AI companies and the data they need to train their models. The core product is data labeling — but that description undersells what's actually happening.

When an AI lab is building a language model, an image recognition system, or a code assistant, it needs training data that has been annotated by humans. That means humans reading text and rating whether a response is helpful or harmful.

Humans looking at photos and drawing precise bounding boxes around objects. Humans listening to audio and transcribing edge cases the automated systems get wrong.

At the scale these companies operate, you need thousands of contractors doing this work consistently, accurately, and quickly.

Scale AI provides the platform that manages all of that. It recruits and manages a global workforce of data annotators (called the Scale Data Engine's contributor network), builds the tooling they use to do the work efficiently, runs quality checks on the output, and delivers clean, structured training datasets to the customer.

Customers pay per task, per project, or through enterprise contracts. The bigger customers — think major AI labs, defense agencies, and Fortune 500 companies building internal AI tools — tend to work on long-term contracts worth tens of millions of dollars.

The government side of the business has become increasingly important. Scale AI has major contracts with the U.S.

Army, Air Force, and DARPA for AI model evaluation, training data, and what the company calls 'AI readiness' work. The Pentagon is one of their biggest clients.

That's not an accident — Wang has been deliberate about positioning Scale as critical national security infrastructure, which is a very different moat than 'we have a good product.'

Scale AI's main product is the Scale Data Engine — the platform that manages the end-to-end pipeline for collecting, annotating, and quality-checking training data. Enterprise customers use it to submit labeling tasks, set quality standards, and receive structured datasets back.

It supports image annotation, video annotation, document processing, natural language tasks, audio transcription, and increasingly, RLHF-style model evaluation work.

Rapidly becoming their flagship product is Scale Evaluation — a suite of tools for testing and red-teaming AI models. Companies use it to stress-test their models before deployment, find failure modes, and get human expert evaluations on model outputs.

Given how much scrutiny AI companies are now under from regulators and the press, model evaluation is becoming a must-have, not a nice-to-have.

Donovan is Scale's defense-focused platform, purpose-built for government and military AI applications. It's designed to integrate data from multiple classified and unclassified sources and provide AI-assisted decision support for military operations.

The U.S. Army and Air Force are among the known users.

It is, depending on your perspective, either critical national security infrastructure or a deeply uncomfortable reminder of where AI deployment is heading.

Scale also runs Spellbook, an AI data labeling tool specifically for legal document review — a narrower vertical product that shows Scale's ambition to go deeper into specific industries rather than remaining purely horizontal infrastructure.

Scale AI's growth story is less about a single viral moment and more about being the only serious option at exactly the right time.

The first smart move was targeting self-driving cars before the general AI wave. Autonomous vehicle companies had enormous data needs, near-unlimited budgets, and a desperate urgency to hit milestones.

Scale AI built deep relationships with Waymo, Cruise, and Lyft's AV division early — which meant when those companies recommended a data partner to other AI teams, Scale AI's name came up first.

The second move was the enterprise pivot. By 2020, it was clear that the self-driving car timeline had slipped by years.

Scale AI was already in conversations with defense contractors and U.S. government agencies, and Wang leaned in hard.

Getting on government contract vehicles — the procurement frameworks that agencies use to fast-track approved vendors — created a revenue stream that is almost impossible to compete away. Once you're on a government contract vehicle, the switching costs for the customer are enormous.

The third move was the bet on RLHF. Reinforcement Learning from Human Feedback is the technique that made ChatGPT coherent — it's the process of having humans rate AI outputs to tell the model what good looks like.

Scale AI had been doing human evaluation of AI outputs for years before RLHF became the hot acronym. When OpenAI needed to scale that process, Scale AI was the obvious call.

Wang has also been unusually aggressive about public positioning for a B2B infrastructure company — writing op-eds on AI policy, testifying before Congress, and being vocal about the U.S.-China AI race. That visibility has opened government doors that pure product companies would never reach.

Scale AI's biggest challenge is existential and structural: the better AI gets, the less it needs Scale AI's core product.

Data labeling is a human-in-the-loop business. Humans label data because AI can't yet do it reliably enough.

But the whole point of training AI on labeled data is to make AI that can eventually handle more and more of that work itself. Multimodal models are already handling image annotation tasks that required human contractors three years ago.

Every capability improvement in foundation models is a potential threat to a labeling revenue line.

Wang is aware of this, which is why Scale has been aggressively repositioning from 'data labeling company' to 'AI development platform.' The pitch is no longer just 'we'll label your data' — it's 'we'll help you evaluate your models, find their failure modes, and build the feedback loops to improve them.' That's a stickier value proposition, but it's also a much more competitive space that puts Scale up against well-funded teams at the AI labs themselves.

The second challenge is customer concentration. OpenAI, Anthropic, and a handful of defense agencies likely represent a substantial chunk of Scale's revenue.

If any of those relationships sour — or if OpenAI decides to build an in-house annotation operation — the impact would be significant.

The third is political exposure. Wang has publicly argued that China is winning the AI race and that U.S.

government action is needed. That's useful for winning defense contracts.

It also makes Scale AI a target if political winds shift, and it creates complications if the company ever wants to operate internationally at scale.

Scale AI's investor list reads like a who's-who of people who understood early that the AI wave would need a foundation layer — and that the foundation layer was going to be unglamorous data work.

Accel led Scale's early rounds and has remained a core backer. Tiger Global came in during the growth stage.

Amazon, Google, and Meta have all participated in funding rounds — which is notable because they're also customers, which creates both alignment and interesting conflict-of-interest dynamics.

The $325 million Series E in 2021 valued the company at $7.3 billion and was led by Tiger Global. By 2024, the valuation had reached $13.8 billion as the AI boom made Scale's position even more obviously critical.

Reports in 2024 suggested Scale was exploring further fundraising at valuations north of $14 billion.

Y Combinator was the original backer — Wang and Guo went through the YC program in 2016. That early validation mattered: YC's network helped Scale land its first enterprise customers and gave Wang access to the Silicon Valley relationships he needed to close deals with Waymo and Lyft before the company had much of a track record.

The defense investment angle has also attracted government-adjacent capital. Scale's positioning as critical AI infrastructure for national security has made it attractive to investors who see the U.S.

government AI budget as a multi-decade growth story — which, given how much the Pentagon has committed to AI modernization, isn't an unreasonable bet.