Surge AI has become one of the most significant yet understated players in the AI infrastructure space. Founded in 2020 by former Google and Twitter engineer Edwin Chen, the company surpassed $1 billion in revenue last year by providing high-quality, human-labeled data to leading AI developers, including OpenAI, Google, and Anthropic.
While competitors have often grown through aggressive fundraising, Surge remained bootstrapped for five years. Chen explained this decision on a recent appearance on the No Priors podcast: “Raising immediately always felt silly to me,” he said. “If you didn’t know how Silicon Valley worked, why would you do that?”
Surge is now preparing for its first external capital raise, reportedly up to $1 billion at a valuation north of $15 billion, signaling a shift in posture as demand for model training and evaluation scales rapidly.
The company’s approach is focused on human-generated data as both product and differentiator. Whether it’s code completions, creative writing, or multi-step reasoning, SurgeAI looks to position its data as higher in quality, and more contextually rich than alternatives. “We don’t just want a poem that checks boxes,” Chen said. “We want the kind of poetry that Nobel laureates would write.”
Defining and Measuring Quality
This emphasis on quality underpins Surge’s critique of how the industry evaluates AI performance. Benchmarks like LMS (Learning from Model Scoring), where annotators compare two model outputs side-by-side, are, in Chen’s view, fundamentally flawed. “They’re literally just vibing with the model,” he said. “They don’t care whether it hallucinated: they’re just picking the one that looks better or has more emojis.”
Instead, Surge says it provides deeper evaluations to its clients, including failure analyses and loss pattern reports. The company treats evaluation as an integral part of its service. Still, the reliance on subjective human judgment introduces its own challenges: consistency across domains, evaluator bias, and limits on scalability. Much of Surge’s methodology remains proprietary, making external validation difficult.
The company is also investing in more advanced forms of training data, including simulated reinforcement learning environments. These aim to reflect real-world workflows, such as a salesperson managing leads across email, chat, and spreadsheets, complete with time-evolving scenarios and external events. “You can’t just synthetically generate this,” Chen said. “You need environments that evolve and reflect reality.”
While synthetic data plays a growing role in the ecosystem, Chen is skeptical of its effectiveness in isolation. “We’ll get customers who say they’ve collected 10 million pieces of synthetic data,” he said, “and end up keeping 5% of it. Sometimes, a thousand well-curated human examples are more useful than the rest.”
SurgeAI says that human feedback will remain critical: not just for quality, but for ensuring models are aligned with intended behavior. But this emphasis on humans also brings the company into a broader conversation about how the AI industry treats the labor behind its systems.
Operational Trade-Offs and Growing Scrutiny
In May 2025, Surge was hit with a class action lawsuit in California alleging it misclassified its annotators as independent contractors, required unpaid training, and imposed time constraints that resulted in reduced pay. The suit frames annotators as central to Surge’s operations, providing the backbone for its billion-dollar business, without receiving the protections of full-time employment.
The company has described the claims as meritless. Still, it indicates growing tension in the data-labeling sector, where contract work is widespread and oversight remains inconsistent. Competitors like Scale AI have faced similar legal challenges, suggesting a systemic issue in how AI training labor is sourced and managed.
The lawsuit wasn’t the only controversy. In 2024, Surge left an internal safety training document publicly accessible online. The document outlined how annotators should handle sensitive topics, such as hate speech, sexual content, and illegal activity, when training chatbots. Some examples drew criticism for being ethically ambiguous or overly permissive. Surge removed the document and stated that it was a research artifact meant to simulate difficult edge cases.
These incidents underscore the complexity of the work Surge facilitates: shaping the rules that define what AI systems are allowed to say, and who decides. They also highlight the challenge of scaling a company built on careful, context-dependent decisions while maintaining operational security and public trust.
Internally, Chen’s views on company-building also depart from typical startup playbooks. He discourages early hiring of product managers and data scientists, arguing that startups should focus on making large leaps in product design rather than small optimizations. “Data scientists are great when you want to improve something by 2%,” he said. “But early on, you’re trying to swing for 10x or 100x.”
That stance may resonate with engineers, but it may also oversimplify the needs of less experienced founders: especially those without access to the networks or safety nets that Surge’s founding team had. As the company becomes more public-facing, its internal practices and assumptions are likely to come under closer examination.
Surge AI has grown by betting on human data as a long-term asset, and by prioritizing control and precision over scale-at-any-cost. As demand for more sophisticated AI systems rises, the value of that data, and the systems that deliver it, will only grow. But so too will the expectations around how it’s sourced and governed.
