The market is misjudging AI data companies because we are trying to analyze a fundamental shift in human knowledge through an old software lens.

The debate this past week shows a deep split in how people view the technology stack. Some see these businesses as temporary utilities. Others see them as the ultimate kingmakers. The financial data supports the category's scale. Frontier Labs spends close to $1 billion a year on human data. $10 billion moves through the market annually. Surge crossed $1 billion in revenue while remaining bootstrapped. Mercor is discussing a $20 billion valuation less than a year after its previous round. Meta paid $14 billion for half of Scale.

We spent the last few months looking closely at this infrastructure. Our experience with early platforms suggests these businesses will remain durable because they support the entire technology stack.

To understand the longevity of these companies, we have to look at what enterprises actually purchase. For >50 years, technology budgets targeted productivity. Software automated repetitive tasks. Today, companies buy intelligence. This requires a completely new supply chain. The primary input is human judgment converted into structured data. Demonstrations, detailed evaluations, and custom rubrics form the foundation. Synthetic data needs these human guideposts to know what success looks like. Ali Ansari at micro1 explained this by comparing the market to global labor spend. He estimated the opportunity at $1 trillion annually. The number looks big because it measures a core input to the economy.

The question for data platforms is who buys the infrastructure next.

Language models grew on a massive historic subsidy. They used a free, digitized corpus of human language covering the entire internet. Robotics platforms have no equivalent. No archive exists to show a machine how to wire a circuit board or fold laundry. Every physical action must be captured, reviewed, and structured by hand. The training data for physical intelligence requires manual construction, right as these hardware teams raise historic amounts of capital.

Enterprise agents represent an even larger wave. Every team deploying an agent hits a technical wall. Traditional software QA cannot validate a probabilistic system. Engineers cannot write a standard test script for an agent navigating a dynamic corporate workflow. Human operators have to define the failure modes, build the evaluation datasets, and capture institutional knowledge. After deployment, workflows change and software agents drift. The process looks like an ongoing subscription.

Flybridge knows the enterprise model layer well. When we backed Arcee in 2023, the team built a platform to train small, open weight models inside a customer's private cloud. The broader market ignored the idea. Most people believed every company would simply query a central frontier model. We believed enterprises would demand models they could download, inspect, and run inside their own security boundaries. Satya Nadella noted that companies using external models pay twice, once with cash and once with proprietary data. Today, private model deployment is becoming standard enterprise strategy. Every corporation running its own models needs a private data loop. They require proprietary fine-tuning data and custom evaluation pipelines. As a result, the buyer base scales from 10 labs to >1000 corporations. Most will choose to buy this infrastructure.

Skeptics see tech-enabled staffing agencies with thin margins, high customer concentration, and minimal moats. This describes the current baseline. It does not explain where we’re going.

High customer concentration is standard for early infrastructure markets. The first buyers of cloud computing were a small group of high-growth technology companies. The market diversified as adoption spread. As the work moves up the technical stack, margins change. Basic data labeling offers thin margins. Providing reinforcement-learning environments, automated evaluation tools, and custom reward models generates software-like returns. This shift is happening right now, as I write this.

That means the moats will be mainly operational. Success will require a mix of research capabilities, deep trust with frontier labs, and tight execution on complex projects. Managing these networks will teach you that operational execution creates a distinct competitive advantage. It also often takes years to build that muscle.

The long-term trajectory relies on a concept the micro1 research team calls the no last mile framework. Their public research paper demonstrates that better models do not eliminate the need for human input, they simply unlock more complex tasks. Each new tier of automation requires human judgment to supervise the next layer. Simple text became expert code, which became complex environments. The demand for human training data is growing faster than ever.

We saw this exact cycle play out with small models. Many analysts looked at the limitations in 2023 and assumed small models were a dead end. The same miscalculation is happening with data platforms today. Every major technology shift has unlocked deep demand in markets we once assumed were full. Human expertise is becoming a primary input to the global economy.

If you are a founder building infrastructure in this category, let us know.

I’d love to chat.

Jesse