Hey there. Welcome to the second edition of this manmade newsletter. The goal here is to make it just a bit easier for all of us to understand what the hell is happening in the training data market. Who is buying? What data? Why?
Benchmarket-ing and hitting the RL Gym
Every major industry has a rating agency. They develop in the presence of information asymmetry and they are typically oligopolies where a few groups lock arms…
Bonds: Moody’s, S&P, and Fitch
Consumer Credit: Experian, Equifax, and TransUnion
Audit: Deloitte, PwC, EY, and KPMG
Sometimes the rating agency vertically integrates into a Testing, Inspection, and Certification (TIC) company. These companies are full of highly nuanced, hard-to-regulate conflicts of interest, but boy are they profitable!
This is what is happening right now with many data vendors. They’re vertically integrating from data processing to collection to benchmark, evaluations, RL environments, and soon handling deployments.
In case these are new terms, here’s what these terms mean:
Data processing: we clean, label, and annotate your data
Data collection: we collect any kind of data you want
Benchmarking: we evaluate and score your model
Reinforcement Learning Environments: we give you a training area, or a RL gym, where your model can improve on our benchmark
This march to vertically integrate and do ALL THE THINGS makes sense. When you have a customer on the hook with money to burn and a trillion-dollar IPO on the line, the answer is definitely to land and expand more business lines with them.
“Hey, remember how we sold you the data? Well, now we’re going to quiz you on that data. Damn, sorry you didn’t do so hot. You can pay us to improve your test results! You’ll need to buy more data and this RL gym membership”
A data lab can go to market now by “benchmark-eting” with an absolutely beautiful, unprecedented, fear-based sale. The pressure on labs to catch up with peers is just suffocating and unrelenting, especially when made legible with a viral bar graph.
Welcome to the era of benchmarks and RL gyms. Thousands of neolabs will form, just like SaaS niched down, in increasingly specialized areas.
China Read AI-2027 Too
In April of 2025, a team of AI researchers wrote AI 2027, an incredible super forecast of what the decade of AI development could / might look like. It’s been my favorite resource to send to friends and family to melt their brains.
Anyways, they were right on the money with their quote here of what China will do in mid-2026…
And now here’s a blurb from China Daily covering the announcement on Tuesday that the Chinese government has created a state-owned data company to unify a national data market and accelerate AI dominance!
The establishment of the company was a major strategic decision made by the Communist Party of China Central Committee with Comrade Xi Jinping at its core, and will help China seize opportunities in AI development, accelerate the building of a digital China, and forge new competitive advantages.
A few second- and third-order effects here:
I imagine China-owned labs are going to be getting all the training data they need from their friendly fellow Chinese companies, and we’ll see a wild jump in benchmark-eting from Chinese labs.
Any US companies trying to sell training data into Chinese labs will have their pricing power hobbled. This has always been geopolitically contentious, so if the money isn’t there, it’ll slow down.
These AI accelerationist moves from China are likely prompting the US government to consider doing more. They took a tiny baby step with the Genesis Mission in November of last year to unify datasets across US Federal research labs. Haven’t seen or heard anything in months in terms of progress or impact here.
Could the US take an equity position in a US lab? Could the US buy training data and give it to US labs? Could we see a Fannie Mae or Freddie Mac Government-Sponsored Enterprise? Will Don Jr or Eric Trump start a data lab 😂 ?
A few more major China benchmark-eting posts and you’ll see Trump pull a big training data lever.
This is parody
Quote of the Week
“Almost every company in this category is selling hours. Maxinsights is the only one that treated quality control as an engineering problem rather than a staffing cost, which is what lets them operate in real life. That decision is hard to copy after the fact, and it’s why the labs that tried other vendors keep coming back to them.”
Thanks for reading! You can reply directly to this newsletter with what I got wrong, what you want to see next week, or any other dank training data memes. See you next week 🫡