Amazon quietly shut down Mechanical Turk on September 30. Most people never used it, but for twenty years it was where the internet went when a computer needed a person: to transcribe a receipt, tag a photo, or, starting in 2008, to sort the millions of images that became ImageNet, the dataset Fei-Fei Li's team used to kick off the modern AI boom.
The companies that replaced it look different today. Mercor, which pays lawyers, doctors and bankers by the hour to write and grade answers for AI labs, says it's now billing at a rate of more than $2 billion a year. micro1, which started out as an AI recruiting tool, said on Friday that it will spend $1 billion over the next year buying data directly from companies, and a day earlier Mercor announced a research team stocked with people from OpenAI and Anthropic.
I went in planning to write about those two announcements. The more I read, though, the more it seemed like they only made sense as part of a bigger picture, because the market for AI training data has split into layers that are heading in very different directions. So this week’s a roundup of where the broker side of LLM model training stands and where it’s headed.
Labeling Automation
The clearest sign that basic labeling is in trouble comes from the companies that sell it. TELUS Digital, one of the biggest vendors of the drawing-boxes-and-rating-answers kind of work, wrote off C$2.1 billion of the business in July, and the explanation in its filing is about as blunt as these documents get: the work it does for big tech customers is "being automated faster than anticipated." Appen, which lost Google as a customer in 2024, saw its revenue outside China fall another 27% in the first half of this year. Sama moved to lay off more than 1,100 people in Nairobi in April after Meta ended its contract.
What's disappearing is simple work, the kind a model can now do about as well as a stranger paid a few dollars an hour. People are still labeling data, just higher up the ladder. (Appen's China business is the exception. It grew 80% in the first half, selling to Chinese model builders.)
Large Revenue, Limited Customers
The money moved one level up, to companies that hire experts instead of crowds. When a lab wants its model to get better at tax law, it needs a few dozen tax lawyers to write hard questions, answer them, and grade the model's attempts, and someone has to find those lawyers, vet them and pay them. Mercor crossed a $2 billion annual run rate in June, The Information reported. Surge, which has never raised outside money, did about $1.2 billion of revenue in 2024. Handshake's AI business is around $1 billion, and micro1 is around $500 million, according to TechCrunch.
Those are gross numbers, and the experts keep most of the money, usually 60% to 70% of what the client pays, so the businesses are smaller than the headlines suggest. They also rest on a very short customer list. The Information found that about 91% of Mercor's revenue in the first half came from companies that build foundation models, which in practice means a handful of labs.
On Friday, Brendan Foody, Mercor's CEO, posted that the average lab or AI-native company spends 5% to 10% of its revenue on expert data and put the whole market at $10 billion to $15 billion. It's also the most concrete estimate anyone in the business has offered.
Scale AI shows what happens when that short list gets shorter. After Meta bought 49% of Scale last year, OpenAI and other labs reportedly took their work elsewhere rather than route it through a company partly owned by a competitor, though Google later came back. Scale did just under $1 billion of revenue last year, according to Forbes, and in August it hired Francis deSouza, Google Cloud's former COO, as CEO, with a plan to sell more to companies and governments.
Why everyone is suddenly publishing benchmarks
If you only sell to five customers, you need a way to show each of them that your data is worth more than the next vendor's. That's the best explanation I have for something odd that happened over the last month: nearly every big broker released a benchmark. Surge put out DAYJOB, which tests whether AI agents can survive a messy day of office work, and sudo L7, a test of staff-level software engineering. Handshake's ATLAS Finance dropped agents into simulated financial firms, where the best of 11 frontier models passed 12% of the tasks. Mercor already runs the APEX benchmarks, and its new research team is built to close the loop: test a model, find where it fails, have experts write data aimed at those failures, retrain and test again. As proof, Mercor took Qwen3.5-397B, an open model, from 16% to 27% on its agent benchmark.
A benchmark, in other words, is a sales tool. It shows a lab exactly where its model falls short, and the company that wrote the test is in the best position to sell the fix. Nathan Lambert, who advises Mercor on the effort, described the problem they're all trying to solve:
The other thing all of them now sell is reinforcement-learning environments, which are simulated workplaces, like a fake Slack workspace or a cloned website, where an agent practices a task and gets graded on it. People in the business told Epoch AI that most tasks sell for $200 to $2,000 each, with the hardest software tasks going for as much as $20,000. Prices like that have drawn a crowd of startups that do nothing else. Halluminate raised $30 million this month, Preference Model announced a16z's backing last week, and AfterQuery was reportedly valued at $3.2 billion in September.
micro1 wants to own the raw material
micro1 is trying something different. Instead of hiring experts to write new data, it wants to buy the data companies have already produced: SOPs, CRM records, project histories, and records of how teams make decisions. It strips out personal information where needed, rewrites some of it, and turns it into training environments, and it has published what it will pay, from $100,000 for a qualified partnership to more than $1 million for data nobody else has.
Citi and Hercules Capital, which mostly makes venture loans, are funding the $1 billion. micro1 hasn't said how much of that is debt, but if a large share is, micro1 is betting it can resell what it buys for more than it paid. TechCrunch reported in August that some of its data can be sold to several customers, with gross margins on that "off-the-shelf" data as high as 80% to 90%. That's how a company holding inventory works.
Companies are lining up. Ali Ansari, micro1's CEO, posted earlier this month that more than 100 a day are signing up, enough that his five-person partnerships team built a sales agent named Kai to take the calls. He says Kai closed a $585,000 deal in a single day. Not every founder is biting, though.
Where this goes
Put the layers together and the data market starts to look like other markets that grew up fast. The commodity work gets automated, a few specialists capture most of the money, and the raw material, which labs used to scrape for free, starts trading at a price.
What I'm watching now is who the specialists sell to next. Garrett Lord, Handshake's CEO, predicted in August that big companies will keep using frontier models for some things and train open models on their own data for the work only they can do. Mercor publishing an open-model training recipe points the same way. If that's right, the companies selling their records to micro1 today could be paying the same brokers to train their own models in a few years.
The training market is moving fast and I’m looking forward to continuing to cover it.
Partnerships & Deals
Compute, Energy & Infrastructure
Schneider Electric agreed to buy PTC for about $22.6 billion in cash, or $205 a share, adding PTC's product-design and engineering software to its industrial AI business. The deal is expected to close by the third quarter of 2027.
Google signed a nuclear power deal with Constellation for about 3.6 gigawatts, starting with 890 megawatts over 20 years from 2028, plus a five-year deal to run Google Cloud and Gemini Enterprise across Constellation's plants.
VEIR raised a $110 million Series C co-led by Matter Venture Partners and Tyche Partners for superconducting power cables that replace copper inside AI data centers.
Etched is fielding offers that value the AI-chip startup at $40 billion to $50 billion, about double its $21 billion mark from September, TechCrunch reported.
Anthropic signed a multiyear cloud deal with IREN, the bitcoin miner turned AI host, The Information reported.
Data, Knowledge & Retrieval
micro1 committed $1 billion over 12 months to buy and license company operational data, financed by Citi and Hercules Capital, to build RL environments for AI agents.
Redpine signed BMJ Group, Sage, IOP Publishing and others to license more than 20 million peer-reviewed papers to AI agents, with each publisher setting its own price per use.
Preference Model came out of stealth backed by a16z, with a $16 million seed according to its job postings, to build RL environments for frontier labs. It also open-sourced its Karotte framework.
USA Today Co. sued OpenAI in federal court in New York for more than $250 million, asking the court to destroy models trained on articles from its 19 papers.
SAP agreed to buy TechWolf, which maps skills and work into a data graph, to ground the AI agents in SuccessFactors. Terms were not disclosed.
Models, Training & Developer Tools
Arena, formerly LMArena, raised a $200 million Series B at a $3.1 billion valuation led by Lightspeed and Khosla Ventures, nearly doubling its valuation in ten months.
TypeSafe AI, maker of the Jev model, raised $870 million at a $7.5 billion valuation led by a16z, with Sequoia and DCVC, Bloomberg reported.
Nous Research raised a $90 million Series B led by Robot Ventures at a reported $1.5 billion valuation to build Hermes for Businesses, an enterprise version of its Hermes Agent.
Harness bought Cosmos, the Auggie CLI and the Code Context Engine from Augment Code, and the team behind them is joining Harness. Terms were not disclosed.
DeepSeek is close to raising about $12 billion, with Tencent and CATL among the biggest backers, Bloomberg reported.
Amazon is in talks to buy Decart, which makes software that moves AI workloads between chips, for about $7 billion, the Wall Street Journal reported.
Enterprise Deployment & Distribution
Manus parent Butterfly Effect raised more than $500 million led by Boyu Capital and IDG Capital, its first round since China forced Meta to unwind its Manus deal.
Sierra and Meta published the Personal Agent Protocol, an open standard for AI agents logging into businesses, with Walmart, Shopify and Stripe as founding partners.
Anthropic launched its Cyber Mission with 11 partners, including CrowdStrike, Palo Alto Networks and Deloitte, to find and fix flaws in power, water and factory systems.
Physical AI & Robotics
Mecka AI raised a $60 million Series B led by Sequoia, with NVIDIA, M12, Qualcomm Ventures and Samsung, to sell human motion data to humanoid-robot builders.
Multiply Labs raised a $75 million Series B led by NantWorks, with AstraZeneca and Teradyne, for robots that automate biologic drug manufacturing.
Ultra raised $62 million, including a $50 million Series A led by Framework Ventures, for rented warehouse packing robots that run on Physical Intelligence software.