Edwin
Chen

Edwin Chen's apartment in San Francisco. May 2020. The city is locked down. Every major AI lab is racing to build smarter models. And this former data scientist from Google, Twitter, Dropbox, and Facebook has a hypothesis nobody wants to hear: the models aren't the bottleneck. The data is.

He launched Surge AI from that apartment. No co-founder, no pitch deck making the rounds, no parade of VC coffees. Just a platform designed to connect AI labs with high-quality human annotation work - the kind that required real intelligence, not just clicks.

Five years later, Surge AI is pulling in over $1.2 billion a year. With under 110 full-time employees. No traditional sales team. Anthropic uses it. Google uses it. OpenAI uses it. Meta uses it. The company that nobody in tech circles talked about built more revenue than almost anyone who talked about nothing but.

The math is obscene: roughly $11 million in revenue per employee. For comparison, a healthy SaaS company aims for $200,000. What Surge built is not a startup that scaled. It is a different shape of company altogether.

The people who know this industry call Chen "the Michael Jordan of post-training data." He would probably find the comparison excessive. He would also probably be right. Michael Jordan had competitors.

Chen did not just build a data labeling company. He documented how bad everyone else's data was.

In July 2022, Surge published research showing that 30% of Google's GoEmotions Reddit dataset was mislabeled - including emotion labels that were the exact opposite of the correct answer. This was a dataset from one of the world's most sophisticated AI research labs, used by researchers globally to train models that detect human emotion.

Five months later, Surge found that 36% of HellaSwag - one of the most widely-cited benchmarks in AI evaluation - contained errors. HellaSwag at the time was considered a gold standard for measuring model common sense reasoning. It was not.

The implications were not subtle. If the benchmarks used to measure AI progress were themselves riddled with errors, the leaderboard was fiction. Models trained on flawed data would learn to mimic flaws. Quality control in AI data, Chen argued, was "often an adversarial problem, similar to email spam" - requiring active ML infrastructure to catch, not just careful humans.

This is not the kind of finding that makes you friends in an industry that has declared its own benchmarks gospel. It is the kind of finding that builds a $1.2 billion business, because it turns out the problem is real and most people prefer to ignore it.

The conventional Silicon Valley wisdom says you need scale to generate revenue. You need growth at all costs. You need to burn capital to build market position. Then you raise your Series B. Then your C. Then your D.

Chen raised a Series A in 2020 ($25M) and then - stopped. Not because he couldn't raise more. Because Surge did not need it. The company grew on revenue. Every new contract funded the next. No dilution, no board seats, no quarterly conversations about burn rate with people who had not touched the product.

The result is a financial structure that looks impossible from the outside: roughly $11 million in revenue per full-time employee. Apple, one of the most profitable large companies in history, runs around $2M per employee. Netflix runs around $3M. Surge runs at a ratio that sounds like a rounding error but is simply what happens when your product is expensive, your quality is non-negotiable, and your customers keep coming back without being asked.

Chen holds approximately 75% of Surge's equity. At a valuation in the $24-30 billion range estimated by analysts, that stake puts his net worth in Forbes territory - the youngest new entrant on the 2025 Forbes 400. The apartment in San Francisco was a good place to start a company.

Chen talks about AGI arriving within a decade - not as a threat, but as a design challenge.

His concern is not that AI becomes too powerful. It is that AI becomes powerful while trained on the wrong things. An AI that is optimized to produce content that feels satisfying to humans, rather than content that is actually true or genuinely valuable, will produce a very sophisticated version of what we already have: systems that tell you what you want to hear.

The phrase he comes back to: AI that is "rich and warm and creative" and "inherently human." Not in the science-fiction sense - not a robot that passes for a person. In the craft sense. AI that writes the way a careful, thoughtful person writes. AI that reasons the way a well-trained expert reasons. AI that knows the difference between a good answer and a comfortable one.

That requires training data that captures actual human expertise, judgment, creativity, and values - not just patterns from whatever was on the internet in 2021. Surge's entire infrastructure exists to produce that data.

The architecture of his MIT education - mathematics for rigor, computer science for systems, linguistics for meaning - was not an accident. It was a twenty-year setup for a company that is, at its core, a translation project: taking the best of human intelligence and making it legible to machines.