Signals Inbox·September 2, 2026·AI Infrastructure

Why is AfterQuery growing so fast?

AfterQuery is growing so fast because frontier AI labs are spending heavily on one of the hardest remaining training problems: teaching models to perform real professional work, and AfterQuery got there early with expert-built tasks, measurable model gains and a software-driven expert network.

Q1Is AfterQuery really growing this fast right now?

AfterQuery really is growing at a freakish pace right now, and the revenue numbers are more convincing than the valuation headline.

When AfterQuery announced its Series A, the company said it had passed a $100 million annual revenue run rate only 14 months after starting. A few months later, co-founder Spencer Mateega said recurring revenue had already grown to the “hundreds of millions.” Forbes’s latest reporting also says the company is profitable.

The valuation moved even faster. AfterQuery announced a $30 million Series A at a $300 million valuation. Five months later, Forbes reported that a new financing was valuing the company at about $3.2 billion, based on two people with direct knowledge of the deal. AfterQuery declined to comment on that round, so we should treat the $3.2 billion figure as reported rather than formally announced.

Still, the commercial acceleration does not depend on the valuation being finalized. “Hundreds of millions” means at least $200 million. If the founder's description is accurate, AfterQuery at least doubled a nine-figure recurring revenue run rate within a few months.

Y Combinator partner Gustaf Alströmer says the company has become the fastest startup in YC history to reach unicorn status. AfterQuery was founded in early 2025. Roughly 18 months later, investors were discussing it as a multibillion-dollar company.

That is fast even by current AI standards.

AfterQuery growth metrics

Q2What is AfterQuery actually selling to AI labs?

AfterQuery sells training experiences that teach AI models how people do difficult professional work.

A lot of the confusion around AfterQuery comes from describing it as a “data-labeling startup.” That makes the business sound like thousands of people tagging images or rating chatbot answers.

Its current work is much more involved.

AfterQuery builds supervised fine-tuning datasets, reinforcement-learning tasks, software environments, expert demonstrations, model evaluations, coding tasks and professional workflows. A model may receive source documents, a spreadsheet, a software repository or access to tools and then have to produce a finished piece of work.

For an AI coding agent, that could involve entering a repository, understanding an issue, modifying the code and passing hidden tests. For a financial agent, it could mean working through a multi-sheet model where several assumptions depend on each other. For a legal agent, it can mean reading case material, deciding what matters and producing work that lawyers can grade against expert-written rubrics.

AfterQuery recently worked with legal AI company Legora on a benchmark containing 5,161 real law-firm cases across 28 practice areas and more than 11,000 source documents. That gives us a good idea of what “AI training data” increasingly looks like these days. The product can resemble a miniature version of the job itself.

That is much more valuable to frontier AI labs than another pile of generic text.

Q3Why do AI labs suddenly need what AfterQuery sells?

AI labs need companies like AfterQuery because making models useful at real work has become one of the hardest parts of the AI race.

The first phase of modern large language models benefited enormously from feeding models more text, more code and more compute. Frontier labs have now consumed huge portions of the easily available public internet. At the same time, customers increasingly expect AI to do much more than generate plausible answers.

The goal now is software agents that fix bugs, research agents that use browsers, finance agents that build models and legal agents that finish multi-step professional tasks.

Those jobs contain thousands of small decisions that are hard to learn from reading documents. A lawyer knows which clause deserves another look. An experienced engineer notices that a technically valid fix will break something elsewhere. A financial analyst can tell when a model output looks mathematically clean but makes no economic sense.

Much of that judgment has never been written down in a form that can simply be scraped.

Reinforcement learning has also become more important in post-training. That creates demand for environments where models can attempt a task, make decisions, receive a useful score and try again.

AfterQuery happens to sell both ingredients: examples of expert work and environments where models can practice it.

The timing is unusually good because the bottleneck moved toward the exact thing AfterQuery had started building.

Q4Did AfterQuery just get lucky with the timing?

AfterQuery got lucky with the timing, but the founders also recognized the problem unusually early and changed the company around it.

Spencer Mateega and Carlos Georgescu originally entered Y Combinator planning to build AI agents for finance. While working on the product, they kept running into a basic problem: leading models could answer finance questions yet struggled when asked to perform realistic financial workflows.

AfterQuery's early FinanceQA research captured that gap. The benchmark tested assumption-based financial analysis rather than simple factual recall, and the company's Y Combinator profile said models failed around 60% of the realistic daily tasks it studied.

The founders could have kept building the finance application and hoped the underlying models improved. Instead, they moved down the stack and started making the data that could improve them.

That pivot looks much smarter with hindsight because the same weakness soon appeared everywhere. Coding agents needed realistic repositories and tasks. Computer-use agents needed actual software workflows. Professional agents needed spreadsheets, documents and decisions.

You can see that progression in AfterQuery's research history. Its work moved from FinanceQA into user-interface understanding, web-app creation, quantitative trading, IDE agents, terminal agents, office work and legal reasoning.

The company kept finding places where models looked impressive in a demo and then broke when the task became closer to a real job.

That gave AfterQuery a very clear product roadmap without having to predict which AI application would ultimately win.

Q5Why are AfterQuery customers paying humans $100 an hour or more?

AfterQuery can pay experts unusually high rates because frontier AI labs care much more about the value of the resulting training task than about minimizing the hourly wage.

Current AfterQuery contractor postings make this visible. Recent software-engineering projects have advertised ranges going from roughly $50 to $180 an hour. One coding project involving real private repositories advertised $150 an hour. Recent specialist roles have included licensed psychologists at $90 to $155 an hour and veterinarians at $95 to $145.

These are very different economics from old-school data annotation.

A frontier lab might spend billions on compute, researchers and model development. Within that budget, paying a strong engineer a few hundred dollars to create or verify a task becomes cheap if the task helps improve a model that will later serve millions of users.

The type of labor also changed. Companies increasingly need people who can spot subtle errors that a general annotator would miss.

A medical coder, for example, may need to identify an answer that sounds completely reasonable but would fail an audit. A software engineer may have to build hidden tests that distinguish a superficial patch from a correct fix. A lawyer needs to know whether an AI-produced argument would survive real professional scrutiny.

AfterQuery is effectively buying judgment rather than cheap clicks.

That shift allows much more money to flow through the expert marketplace while still making economic sense for the customer.

Q6Is AfterQuery basically another Mercor?

AfterQuery competes with Mercor for some of the same AI budgets, but AfterQuery is leaning harder into the technical question of whether the data actually improves a model.

Mercor has become the giant comparison because it built an enormous marketplace connecting AI companies with programmers, researchers and other specialists. Mercor says it went from zero to a $500 million revenue run rate in 17 months and later passed $1 billion.

AfterQuery's pitch goes further into data design and model research.

The company builds custom software to manage how experts create tasks, how their work gets checked and how difficult the resulting examples are. Mateega has described the ideal data as sitting in a “Goldilocks” range: challenging enough that frontier models still have something to learn, without being so difficult that training produces little useful feedback.

AfterQuery also employs researchers who can run post-training experiments themselves. A customer can therefore receive more than a folder of expert-created examples. AfterQuery can test whether those examples move a benchmark and adjust the dataset when they do not.

So the customer is buying something different.

The closer AfterQuery gets to selling measurable model improvement, the less its business depends on simply having access to a large pool of contractors. That is probably the most important difference between the two companies today.

Q7Does AfterQuery's data actually make AI models better?

AfterQuery now has credible evidence that its training data can materially improve AI models, with NVIDIA providing the strongest outside validation.

NVIDIA publicly used AfterQuery's Office Agent tasks while training Nemotron 3 Ultra. In NVIDIA's published warm-up experiment, performance on GDPval rose from 35.3 to 46.7 after the relevant training step. That is an 11.4-point increase, or roughly 32% relative to the lower score.

The effect was also specific. BrowseComp moved from 33.0 to 44.4, while HLE barely changed from 26.3 to 26.7. That pattern is useful because it suggests the training improved capabilities related to the tasks rather than magically lifting every benchmark.

AfterQuery has published several experiments of its own. In one, the company improved gpt-oss-20b from 3.1% to 17.0% on Terminal-Bench 2.0. Another experiment using 1,057 passing expert trajectories produced gains of up to 4.33 times on unseen τ² tasks. Its GDPval work with on-policy distillation produced a roughly 21% net win-loss advantage over the base model.

We should give NVIDIA's result more weight than AfterQuery's internal experiments because it comes from an external model developer using AfterQuery data in a real training recipe. The company-run work still tells us something useful: AfterQuery has repeatedly built datasets around different tasks and then tested whether models improve after training on them.

That makes the product much easier to believe.

Model-improvement experiments using AfterQuery data

Q8Why do AfterQuery's benchmarks help it win customers?

AfterQuery's benchmarks make the sales pitch easier because AI labs can see what the company is capable of before handing it a large training-data contract.

Buying specialist AI data is awkward. A dataset may look clean, employ impressive experts and still teach a model almost nothing useful. Customers ultimately care about what changes after training.

AfterQuery has turned that uncertainty into part of its marketing.

The research page now contains work spanning finance, user interfaces, web applications, quantitative trading, IDE agents, terminal agents, professional office work and legal reasoning. Several projects connect directly to outside organizations rather than existing as isolated company demos.

NVIDIA used AfterQuery data in Nemotron 3 Ultra. Thinking Machines Lab's Inkling was trained using data supplied by AfterQuery. Motif Technologies recently named AfterQuery its sole data partner for Motif 3. Legora worked with the company on its new legal-agent benchmark.

Those relationships cover very different customers: a giant chip and AI company, a frontier-model lab, a Korean model developer and a vertical legal-AI company.

The breadth is useful commercially. A potential customer looking at AfterQuery these days does not have to infer everything from a founder pitch. It can inspect public research and see that other serious AI teams have already trusted the company with model-development work.

Research has effectively become part of AfterQuery's distribution.

Q9How can AfterQuery handle this much work with such a small team?

AfterQuery can support a huge amount of customer work because its permanent team runs a much larger outside expert network through software.

Y Combinator currently lists AfterQuery's core team at around 30 people. At the Series A, the company said it already worked with nearly 100,000 verified practicing professionals across engineering, medicine, law, finance and other areas.

Those figures are easy to confuse. Nearly 100,000 verified professionals does not mean 100,000 people are working for AfterQuery every week. The company has never published an active-contributor number comparable with Mercor's weekly-active figure.

The network still gives AfterQuery something valuable: optionality.

A new project may require software engineers this week and licensed healthcare professionals next week. AfterQuery can search an existing pool, test people for the work and assemble a project workforce without hiring everyone permanently.

The company also keeps building software around task creation, quality control and contributor management. Recent job postings show the same model in action: contractors qualify through assessments and can then be matched to ongoing projects.

The more interesting moat is the history behind that network. After enough projects, AfterQuery can learn which engineer writes difficult but fair coding problems, which reviewer consistently finds hidden errors, which lawyer produces tasks that models can actually learn from and which experts require too much rework.

A list of professionals can be copied. Years of performance data on which professionals produce useful AI training material would be much harder to recreate.

AfterQuery is still too young for us to know how strong that advantage has become, but that is where the network could get more valuable over time.

Q10Can AfterQuery sell the same data more than once?

AfterQuery can increasingly sell reusable datasets alongside custom projects, which could make the business much more scalable than a pure expert-services company.

The Office Agent dataset gives us the clearest example.

AfterQuery created professional tasks involving files, multi-step reasoning and finished deliverables. NVIDIA later used tasks with that structure while training Nemotron 3 Ultra. AfterQuery currently offers Office Agent and GDPval-style training datasets commercially to other customers.

The company also markets data for coding agents, terminal agents and other forms of agentic post-training.

That creates an attractive cycle. A custom customer pays AfterQuery to solve a new training problem. The company learns how to design the tasks, builds tools around them and gathers expert work. Parts of that know-how can later become a repeatable dataset or environment for another customer.

The economics depend heavily on how often reuse actually happens, and AfterQuery does not disclose that revenue split.

Still, the direction is clear. A business where every new dollar requires another dollar of custom expert work behaves like a services marketplace. A business that can package some of its accumulated training material and sell access repeatedly has much better potential margins.

AfterQuery is currently trying to operate both models at once.

Q11Could Atrium make AfterQuery much bigger?

Atrium could open a much larger source of AI training data for AfterQuery by turning private company work into something businesses can sell.

AfterQuery recently launched Atrium with a simple pitch: companies can submit private codebases, CAD designs or financial models, let AfterQuery inspect them and receive a firm quote.

The roadmap goes much further. Atrium lists Jira, Asana and Linear for project-management data; Slack and Microsoft Teams for communications; Confluence, Notion and Google Docs for documents; Salesforce and HubSpot for CRM; QuickBooks and NetSuite for finance; Workday for HR; and Figma for design.

There is a huge amount of potential training material hidden inside those systems.

A public GitHub repository teaches a model how open-source software looks. A private repository can show years of real company development: bug fixes, architectural compromises, unfinished migrations and problems that never appear in tutorials.

The same applies to finance. A textbook can explain a discounted cash-flow model. A large collection of real financial models shows how analysts actually structure assumptions, make mistakes, revise forecasts and present outputs.

Atrium therefore expands AfterQuery's supply beyond newly created expert tasks. It gives the company a way to buy work that already exists.

The obvious difficulty is permission. Private repositories, emails, financial models and internal documents contain intellectual property and sensitive information. Atrium explicitly asks sellers to confirm they are authorized to share the data, but rights management will become much harder as the platform moves into broader enterprise systems.

For now, Atrium is too new to explain AfterQuery's current revenue surge. It is more interesting as a clue about where the company thinks the next scarce training corpus will come from.

Q12Is the whole AI training-data market exploding, or is AfterQuery special?

The whole high-end AI training-data market is exploding, while AfterQuery stands out mainly for reaching very large scale unusually quickly.

Mercor proves that AfterQuery's demand is part of something bigger. Mercor says it reached a $500 million revenue run rate in 17 months, passed $1 billion after that and now pays more than $2 million every day to over 30,000 weekly active contractors.

Scale AI became another giant by supplying training data and evaluation work to major AI developers, eventually reaching a strategic transaction with Meta at a valuation around $29 billion.

The market has also become crowded. A recent industry compilation counted more than 50 startups selling some combination of training data and reinforcement-learning environments to AI developers.

So customers are clearly spending enormous amounts of money on this layer.

AfterQuery's unusual feature is how quickly it entered the top tier. As seen above, the company had already crossed a $100 million revenue run rate after roughly 14 months, despite starting well after Scale and Mercor.

It changes how the story should be read. AfterQuery did not somehow convince the AI industry to invent a new budget category for one startup. The budget was already exploding. AfterQuery found a particularly valuable part of it and scaled fast enough to capture a meaningful share.

AI training-data market comparison

Q13Is AfterQuery's revenue really comparable to SaaS ARR?

AfterQuery's revenue is impressive, but we cannot treat every dollar of its run rate like a dollar of high-margin software ARR.

This is the biggest unanswered financial question around the company today.

AfterQuery pays outside experts to create and review training material. If a customer pays $1 million for a project and $600,000 ultimately goes toward contributors and delivery, the economics look very different from a software company collecting $1 million for subscriptions.

The company has not publicly disclosed its gross margin, contractor costs or precise revenue-recognition method.

Mercor shows why we need to be careful with this category. Its enormous headline revenue goes through a marketplace where more than $2 million is currently paid to contractors every day. Customer spending is real, but a substantial part of it immediately leaves the company.

AfterQuery may have better economics because it combines custom expert work with reusable datasets, research, software and training environments. Forbes also reports that AfterQuery is profitable, which is encouraging.

We still need the margin data.

Until we get it, the safest interpretation is that AfterQuery's revenue numbers prove extraordinary customer demand. They do not yet prove that AfterQuery has SaaS-like economics.

That difference could eventually matter much more than whether the current run rate is $200 million, $300 million or higher.

Q14Is AfterQuery buying its growth with venture capital?

AfterQuery does not look like a startup buying revenue with huge venture subsidies because it reached major scale after raising relatively little money.

Before the latest reported financing, AfterQuery's main announced round was a $30 million Series A. According to the company, it had already grown multiples since that round actually closed, meaning the financing announcement was partly catching up with a business that had moved ahead.

As pointed out above, Forbes now reports that AfterQuery is profitable.

We do not know how profitable, and profitability says nothing by itself about the quality of the margins. But the combination of large customer spending, a small core organization and limited previously announced funding makes heavily subsidized growth a weak explanation.

AfterQuery's structure helps.

Instead of putting thousands of experts on permanent payroll, it can assemble them around customer projects. Customers fund the work, contractors get paid for producing it, and the permanent company mainly builds research, software, operations and sales around that marketplace.

That can generate an extraordinary amount of revenue without requiring an equally extraordinary amount of venture capital.

The much more interesting financial question now is how much profit AfterQuery can keep as the business becomes larger.

Q15What could actually slow AfterQuery down?

AfterQuery's growth could slow quickly if AI labs start producing more of this training work themselves or if expert data becomes easier to buy from competitors.

Customer concentration is the clearest risk. There are only so many frontier AI laboratories able to spend tens or hundreds of millions of dollars on post-training. Landing one large program can transform a young startup's revenue. Losing one can hurt just as quickly.

Competition is also getting brutal. AfterQuery faces specialist startups, much larger data platforms and the internal data teams of the AI labs themselves. Customers such as OpenAI, Anthropic, Google, Meta and NVIDIA have no technical reason they must outsource this work forever.

AfterQuery therefore has to remain better or faster at producing useful training material.

Price competition becomes more dangerous if several providers can recruit the same lawyers, programmers and doctors and generate similar benchmark gains.

Atrium introduces a different problem. Buying private enterprise data could give AfterQuery unusually valuable training material, but the legal and security burden rises dramatically once codebases, financial models, company communications and internal documents are involved.

There is also a financial risk we still cannot quantify. If expert payouts and project operations absorb most of each additional customer dollar, AfterQuery may discover that revenue can scale much faster than underlying profit.

Finally, geopolitics could intrude. Forbes reports that AfterQuery works with Chinese AI labs as well as Western ones. Training data and frontier-model development are increasingly treated as strategic technologies, so future restrictions could affect which customers a US AI supplier can serve.

What could slow AfterQuery down

Q16Why is AfterQuery growing so fast?

AfterQuery is growing so fast because frontier AI has run into a new bottleneck: models know a huge amount, but they still need much better examples of how experts actually get difficult work done.

That explanation fits the evidence better than the simpler story that AfterQuery found a clever way to hire lots of contractors.

The demand was already becoming enormous. AI labs are spending heavily on post-training, reinforcement learning, agentic workflows and evaluations. Other companies in the same broad market have also reached extraordinary revenue levels, which confirms that AfterQuery is riding a real spending wave.

AfterQuery then picked a valuable part of that wave. Its experts create realistic coding, finance, legal, medical and office tasks. Its software helps filter and validate that work. Its internal researchers test whether models improve. Its datasets can sometimes be reused. Atrium now gives the company a path toward private enterprise data that cannot be scraped from the public web.

Recent customer work makes the story stronger. Motif chose AfterQuery as the sole data partner for Motif 3. Legora built a 5,161-case legal benchmark with the company. Thinking Machines used AfterQuery-supplied data while training Inkling. As we saw previously, NVIDIA showed that relevant AfterQuery tasks could produce a large gain on professional-work benchmarks.

We therefore have much more than a funding narrative.

The part we still cannot see clearly is the quality of the revenue. AfterQuery has not disclosed gross margins or how much customer spending ultimately goes to expert labor. That could change how impressive the economics look, but it does not change the answer to the title.

AfterQuery is currently growing this fast because AI labs have become willing to spend enormous amounts to teach models how real professionals work, and AfterQuery found a way to turn that demand into a product unusually early.

The $3.2 billion reported valuation may prove aggressive. The underlying growth is much harder to dismiss.