AfterQuery's $3.2B: expert judgement is the AI moat
AfterQuery reportedly went from a $300M Series A to a $3.2B valuation in under five months by paying doctors and lawyers to teach AI. Here is what that signals for Sri Lankan professionals.
The AfterQuery $3.2 billion valuation is the clearest price tag anyone has put on human expertise this year. TechCrunch reported on 1 September 2026 that the AI training-data startup has raised at $3.2B, roughly five months after announcing a $30M Series A at $300M.
That is not a story about a lucky round. It is a story about where the money in AI has quietly moved, and the answer should interest anyone in Sri Lanka who has spent ten years getting good at something specific.
📊 What the numbers actually say
Here is the timeline as reported, with nothing added:
| Milestone | When | Number |
|---|---|---|
| Y Combinator batch | Winter 2025 | ~18 months before this round |
| Series A announced | April 2026 | $30M at $300M valuation |
| Annualised revenue run rate | April 2026 | $100M |
| Latest round (reported) | September 2026 | $3.2B valuation |
| Founders' ages | — | 22 and 23 |
Two things jump out. First, a 10x valuation jump in under five months is not normal even by 2026 standards. Second, the $100M run-rate figure is from April, so any multiple you calculate off it is already stale. At April's numbers the company would be priced at roughly 32x revenue; if revenue kept climbing, the real multiple is lower. I don't know the current figure and neither does anyone outside the cap table.
YC partner Gustaf Alströmer called it the fastest any startup has gone from launch to unicorn in the accelerator's history, according to TechCrunch's report.
🧠 The product is judgement, not data
This is the part worth sitting with. AfterQuery's described business is employing professionals, including doctors and lawyers, to train models on how tasks are actually performed. TechCrunch quotes the framing as encoding the patterns, decisions and reasoning of the world's best practitioners.
Notice what is not being sold:
- Not raw text scraped from the web. That well is long since drained.
- Not cheap bounding-box labelling at scale, the model that built Scale.
- Not synthetic data generated by another model.
What is being sold is the thing a model cannot get from the open internet: how a competent professional decides. A lawyer's reasoning for rejecting a clause. A doctor's differential when two symptoms conflict. That reasoning is almost never written down, because nobody writes down the obvious parts of their own job.
The named customers are telling too. Nvidia, Legora, and Motif Technologies, a Korean AI lab. Those are labs and infrastructure companies, not end-user apps. The buyers are the people building models, and they are paying a premium because this input has no substitute.
🌐 Why this matters more in Colombo than in San Francisco
Sri Lanka has a specific asset here that we usually undersell: a large pool of English-fluent, formally qualified professionals whose hourly rate is a fraction of the US equivalent. Companies like AfterQuery and Mercor exist because that arbitrage is real and buyers will pay for it.
The realistic opportunities, ranked by how likely they are to actually pay you:
- Expert task authoring. Writing problems and reference solutions in your own field: accounting, law, medicine, civil engineering, actuarial work.
- Preference and quality rating. Judging which of two model answers is better, with a written reason. Boring, steady, pays hourly.
- Code trajectory recording. Solving real bugs while your keystrokes and reasoning are captured. Strong fit for the SL software crowd.
- Domain review. Catching where a model is confidently wrong in a field you know cold.
If you land any of these, the money arrives in USD and the deductions are not trivial.
Set the number before you negotiate, not after. The freelancer hourly rate calculator works backwards from your target income, unpaid time and expenses, and the Sri Lanka freelancer tax calculator tells you what IRD expects on foreign-currency income.
A $30/hour contract does not mean 30 × the mid-market rate lands in your account. Between the transfer spread and the tax you owe here, it never does.
⚠️ The uncomfortable half of this story
Now the part these funding announcements skip. In this model, the expert is an input cost, and the input cost does not participate in the exit.
| Who | What they contribute | What they capture |
|---|---|---|
| The professional | Years of trained judgement | An hourly rate |
| The platform | Coordination, QA, contracts | The $3.2B |
| The lab | Compute and distribution | The end product |
I don't say that as a complaint about capitalism. I say it because it determines what you should do with the work. Treating expert annotation as a career is a mistake. Treating it as funded research time is not.
Key takeaway: The demand is for what you know, not for your hours. If you sell only hours, you are competing on price with every qualified person on earth. If you use the work to learn exactly where models fail in your field, you are building something the platform cannot bill you for.
The second uncomfortable point: this work is teaching models to do the reasoning parts of professional jobs. Anyone taking these contracts should be honest that they are compressing the timeline on their own field's automation. That is not a reason to refuse. It is a reason to pay attention to what the tasks are actually about, because the task list is a free preview of which parts of your profession get automated first.
🛠️ What I'd do with this information
Concretely, for a Sri Lankan engineer, student or small-team builder:
- Write down your domain reasoning. If you can articulate why you make a call, you have the exact asset these companies buy. Most people cannot, which is why the rates are decent.
- Keep a failure log. Every time a model gets something wrong in your specialty, record the prompt and the correct answer. After six months that log is a product, a paper, or a hiring argument.
- Don't quit your job for annotation work. Contract volume is lumpy and the platforms change terms.
- Watch the buyer list, not the valuation. Nvidia and a Korean lab buying expert data tells you more about 2027 than the $3.2B does.
💡 What this means for you
Strip out the unicorn framing and one fact remains: in September 2026, the scarcest input in AI is a qualified human explaining their own reasoning. Not compute, not architecture, not scraped text. That scarcity is why two founders aged 22 and 23 are reportedly sitting on a $3.2B company eighteen months out of Y Combinator.
For anyone here holding a real professional qualification, that is a genuine, if narrow, window. Take the contracts if they come, price them properly, and treat every task as intelligence about your own field rather than as a paycheque. The valuation belongs to San Francisco. The knowledge of where these models break is still yours to keep.
Written as commentary on TechCrunch's reporting. Figures are as reported there; I have not independently verified the round.