Hook
Everyone is staring at the next frontier model—GPT-5, Claude 4, some undisclosed Gemini iteration. They dissect parameter counts, watch for leaked benchmarks, obsess over inference costs. But they’re looking at the wrong end of the pipeline. The real bottleneck isn’t compute; it’s the raw, human-generated feedback loops that turn a pre-trained base into a reasoning engine. And one company, Mercor, is reportedly discussing a $20 billion valuation to capture that bottleneck. That number is absurd. It’s also probably too low.
Context
Mercor is an AI training data provider. In plain English: they hire humans to label data, write preference comparisons, and generate the RLHF (Reinforcement Learning from Human Feedback) signals that make models polite, accurate, and aligned. The company sits in the same sandbox as Scale AI (valued around $13.8 billion in 2024) and Appen (market cap ~$500 million). But $20 billion would put Mercor roughly 45% higher than Scale’s last private round. To justify that, you need to believe the data labeling market is not just growing—it’s being repriced from a cost center to a strategic asset.
Core
The data labeling market is fragmenting along a quality axis. The era of “label 100K cat pictures for $0.05 each” is dead. Today’s models demand expert feedback: board-certified doctors for medical imaging, PhD-level physicists for scientific reasoning, multilingual poets for creative writing. This is not scaling on Mechanical Turk. It requires recruitment, vetting, training, and quality control—a human infrastructure play disguised as a tech company. Mercor’s valuation implies they’ve built that infrastructure, possibly with exclusive contracts with top AI labs.
But let’s get mechanical. Based on my experience auditing DeFi protocols during the 2020 yield farming frenzy, I’ve learned to read between the lines of private market stories. A $20 billion valuation with “concerns about revenue sustainability” screams something: revenue is either highly concentrated or lumpy. If Mercor’s top three clients account for 70% of revenue, one lost contract drops the valuation by $10B overnight. That’s not a margin of safety; it’s a cliff. The market is pricing in optionality without a hedge.
Code is law, but bugs are justice. In smart contracts, a bug can drain a pool. In data labeling, a bug is a data leak or a biased annotation. The article flags “safety” as a concern. From my years in cybersecurity, I know that human-in-the-loop systems are the hardest to secure. You can audit the code, but you can audit the 10,000 annotators. Mercor’s valuation assumes they have solved this at scale. I doubt it.
Contrarian
The contrarian angle: this valuation is not about data quality—it’s about narrative. The same VCs who pushed ‘liquidity fragmentation’ as a problem to sell you their $100M TVL DeFi product are now pushing ‘data bottleneck’ to sell you a $20B multiple. The structure is identical: create a fear gap, propose a solution, charge a premium. The difference is that data is actually scarce, but the valuation is still driven by FOMO and the desire to exit before the next cycle.
Retail investors—and here I include anyone not inside the cap table—are being sold a story of inevitable growth. But look at Appen: once a darling, now a cautionary tale. Appen’s revenue slid 30% in FY2023 as clients built internal labeling teams. The same could happen to Mercor if OpenAI decides to hire 50,000 annotators directly. The barrier to entry in this business is not technology; it’s trust and scale. And trust is expensive.
Greeks don’t price data. In options, volatility is a tax on uncertainty. In private markets, valuation is a tax on narrative. The implied volatility of Mercor’s stock (if it traded) would be astronomical because the uncertainty is binary: either they sign a multi-year deal with the next GPT and become irreplaceable, or they get commoditized. There is no gentle drift between those outcomes.
NFT floor is a feeling, not a number. Similarly, Mercor’s $20B valuation is a feeling—a sentiment that data is the new oil. But oil has a spot market; data does not. Without transparent revenue multiples, we are trading on hope. And hope is a lousy basis for asset allocation.
Takeaway
When the AI ETF (yes, there are now AI ETFs) stops buying the story and starts asking for audited financials, Mercor’s $20B will either look like a bargain or a relic of a bygone bull market. The only actionable signal today is to watch for customer announcements, security incidents, and whether they file for an IPO. Until then, treat the headline as a conversation starter, not a thesis.
Rhetorical question: If data is the new oil, who gets left holding the empty barrels when the well runs dry?