You’ve probably noticed the same headlines I have: Nvidia’s chips are the new oil, datacenters are sprouting like mushrooms, and every AI company is locked in an arms race for more computing power. But while everyone’s been staring at the GPU clusters, a quieter story has been unfolding—one that reveals a trillion-dollar giant’s blind spot and a startup that turned that oversight into a $2 billion run rate.
That startup is Mercor, and it’s doing something that Amazon was perfectly positioned to own but completely abandoned: supplying the high-quality human data that makes AI actually work.
Amazon didn’t just lose a market—they handed it to the startup that understood AI’s dirty secret.
Let me explain. For years, Amazon Mechanical Turk was the go-to platform for getting humans to perform microtasks: label images, transcribe audio, filter spam. It was cheap, scalable, and the backbone of early machine learning. But then the AI boom hit, and something shifted. The demand wasn’t for five-cent tasks from anonymous workers anymore. Large language models need humans who can write detailed feedback, evaluate complex reasoning, and catch subtle errors. They need a premium, skilled workforce—not a digital sweatshop.
Amazon could have pivoted. They had the infrastructure, the trust, and the data. Instead, they let Mechanical Turk wither. New customers were silently turned away. The platform stagnated. And Mercor stepped into the void.
Here’s the part that makes you stop: Mercor’s $2 billion gross run rate isn’t a fluke—it’s a direct reflection of the gap Amazon left open. The startup built a platform that vets, trains, and manages a distributed workforce specifically for AI training. They provide the exact layer that large language model companies like OpenAI, Anthropic, and Google need to fine-tune their models. And Amazon? They’re busy selling compute on AWS, completely missing that the bottleneck isn’t just hardware—it’s the humans who teach the machines.
I saw this firsthand while researching the AI supply chain. Every foundation model company I talked to named the same pain point: finding reliable, skilled annotators. They don’t need someone to click a button for a penny. They need people who can write nuanced essays, debate ethical trade-offs, and generate high-quality training data at scale. That’s a completely different business than Mechanical Turk ever was.
The most valuable resource in the AI era isn’t GPU clusters. It’s people.
And the truly provocative twist? Amazon didn’t just miss this trend—they actively stepped away from it. In 2024, Amazon stopped accepting new Mechanical Turk customers altogether. While Mercor was raising capital and scaling operations, the world’s most valuable retailer was voluntarily exiting the market at the exact moment demand exploded.
This isn’t just a startup success story. It’s a warning for every big company that assumes their lead is permanent. When you stop listening to the market, when you let legacy products and internal silos define your strategy, you leave billions on the table for someone hungrier to grab.
So where does that leave us? If you’re an investor, stop obsessing over which model will win. Look at the hidden infrastructure—the data pipelines, the annotation platforms, the human-in-the-loop layers. That’s where the real, defensible value is being built. And if you’re a leader at a large company, ask yourself: What markets are we neglecting today that a startup will own tomorrow?
Amazon taught the world that compute was the future. Mercor proved that the future needs humans too.
FAQ
Q: Isn't this just a temporary bubble? Will demand for human data fade as AI gets better?
A: No. As AI models improve, the need for high-quality human feedback actually increases—it's what keeps them aligned, safe, and accurate. The data pipeline is a permanent layer, not a temporary crutch.
Q: What's the practical takeaway for someone investing in or building AI?
A: Stop betting exclusively on model companies. The real moats are in the infrastructure that models depend on: data annotation, RLHF pipelines, human evaluation platforms. Mercor shows there's a $2B+ opportunity in being the 'boring' layer beneath the hype.
Q: Could Amazon have made a rational choice by exiting Mechanical Turk?
A: Potentially—if they saw it as a low-margin distraction. But that rationale ignores the massive adjacent market for premium human data. Amazon could have repurposed MTurk's infrastructure for high-value work. Instead, they left a vacuum for startups. That's not strategy; that's neglect.