What’s funding like in the AI training data market right now?

All funding deals in the AI training data market over the last 24 months

Below is the table listing all the deals. You can find our methodology at the end of this page.

If you want a deeper understanding of the market and its current dynamics, get our report covering the AI training data market.

Is AI training data funding actually healthier right now?

Yes. AI training data funding is clearly healthier today: far more companies are raising, the typical round is bigger, and the improvement still holds once we remove the giant deals.

We counted 49 financings in the latest 12 months, compared with 22 in the previous period. The number of funded companies more than doubled from 20 to 44, while the median disclosed round climbed from $8.15 million to $14.5 million.

Known funding reached at least $1.51 billion, nearly four times the previous $404.3 million. That headline is flattering to the market because Mercor and Snorkel both raised $350 million, but the underlying numbers are also much stronger. Funding below $50 million rounds more than doubled, and first financings rose from five to 18.

The market has a much deeper bench now. A year ago, activity was thinner and relied more heavily on a relatively small group of companies. Today we see new data-collection startups, expert networks, proprietary-data platforms and physical-AI infrastructure companies entering at the same time that established businesses are raising much larger follow-ons.

Metric	Current L12M	Previous 12M	Change
Deals	49	22	+122.7%
Funded companies	44	20	+120.0%
Known/minimum capital	$1.510B	$404.3M	+273.4%
Median disclosed round	$14.5M	$8.15M	+77.9%
First financings	18	5	+260.0%
First-financing share	36.7%	22.7%	+14.0 pp
Average deals per month	4.08	1.83	+123%

Would AI training data funding still look strong without Mercor and Snorkel?

Yes. The AI training data boom survives easily without Mercor, Snorkel and the other very large rounds.

Mercor and Snorkel each raised $350 million, while micro1 disclosed a financing of more than $100 million. Those transactions clearly inflate the headline total.

Once every round above $50 million is removed, however, funding still rises from $232.3 million to $529.4 million, up about 128%. If we remove only rounds above $100 million, current funding still reaches roughly $710 million versus $404 million one year earlier.

The median round is also up 78%. A couple of monster checks cannot manufacture that change.

The billion-dollar headline exaggerates the magnitude of the improvement, but the boom underneath it is real.

Are more AI training data startups getting funded now?

Yes. AI training data funding is reaching more than twice as many companies as it did one year earlier.

We found 44 funded companies in the latest period versus 20 previously. First financings jumped from five to 18, so a large part of that increase comes from genuinely new entrants rather than companies simply returning for another round.

The new cohort is also unusually varied. Config is building robotics-data infrastructure. Origin Lab is turning licensed video-game worlds into data for embodied AI. Luel operates a contributor network for multimodal data. AfterQuery builds expert datasets and reinforcement-learning environments. Mundo AI focuses on perceptual and multimodal data.

These companies are attacking very different bottlenecks, which makes the increase in deal count more convincing than a surge concentrated in one narrow niche.

Are new AI training data startups winning, or are existing winners taking the money?

Both are benefiting, but existing AI training data companies are taking almost all of the extra dollars.

First financings rose from five to 18, yet known capital going into those first rounds barely moved, from $104.5 million to $111.3 million. The median first financing actually fell from $8 million to $6 million.

Follow-ons went in the opposite direction. Known follow-on capital increased from about $300 million to almost $1.40 billion, while the median follow-on grew from $8.3 million to $27 million.

The entry funnel is wider, but the funding curve gets much steeper once companies prove themselves. Investors are happy to seed a larger number of ideas; the serious money is going to businesses that already have customers, proprietary supply, specialist networks or operational scale.

Mercor, Snorkel, Encord, Protege and AfterQuery all fit that second group.

Metric	Current	Previous
First financings	18	5
First-financing share	36.7%	22.7%
Known first-financing capital	$111.3M	$104.5M
Median first financing	$6.0M	$8.0M
Follow-on financings	30	17
Known follow-on capital	$1.398B	$299.8M
Median follow-on	$27.0M	$8.3M

Why are AI training data follow-on rounds getting so much bigger?

AI training data follow-on rounds are getting bigger because a few business models are starting to look much more scalable than they did one year ago.

Mercor raised $350 million after building a large network of domain specialists who help train advanced models. Snorkel raised another $350 million and said its newer data-as-a-service business had grown more than 18-fold, reaching a $375 million annualized revenue run rate. AfterQuery said after its $30 million Series A that it had already passed a $100 million revenue run rate.

Protege provides another good example. It raised a $25 million Series A and then announced another $30 million led by Andreessen Horowitz only months later, as demand expanded across healthcare, media, audio and motion-capture data.

These are very different companies, but investors can now point to actual usage, revenue, repeat customers and hard-to-copy data supply. That supports much larger checks than the market was writing a year earlier.

Which part of AI training data is actually gaining the most momentum?

Physical data collection is currently the clearest broad-based winner in AI training data, while expert feedback attracts the biggest checks.

Data Collection Services grew from 9.1% to 28.6% of deals. Its median financing also rose from $8 million to $14 million. Config, Luel, Human Archive, Vision Lab, Mecka and XDOF have all raised around the idea that real-world data itself needs industrial-scale collection infrastructure.

RLHF & Expert Feedback also strengthened, moving from two deals to seven and reaching 37.3% of known capital. Mercor and micro1 have a large effect on that dollar share, but the increase in the number of funded companies is real.

Domain-Specific Training Data is growing differently. Its share of deals rose from 13.6% to 22.4%, while its share of capital stayed near 11%. That looks like a young category filling up with new companies rather than one already dominated by huge rounds.

Category	Current deal share	Previous deal share	Current capital share	Current median	Funding intensity
Data Collection Services	28.6%	9.1%	17.2%	$14.0M	0.60
Domain-Specific Training Data	22.4%	13.6%	10.9%	$15.0M	0.48
RLHF & Expert Feedback	14.3%	9.1%	37.3%	$40.1M	2.61
Synthetic Training Data	12.2%	27.3%	1.9%	$7.4M	0.16
Data Curation & Cleaning	8.2%	4.5%	0.6%	$4.7M	0.08
Licensed Proprietary Datasets	8.2%	9.1%	4.7%	$16.2M	0.57
Data Annotation & Labeling	6.1%	27.3%	27.3%	$60.0M	4.47

Is old-school AI data labeling losing ground?

Yes. Traditional AI data labeling is losing breadth quickly, even though the strongest platforms can still raise enormous rounds.

Data Annotation & Labeling fell from 27.3% of deals to just 6.1%. That is the steepest loss of deal share among the main categories we track.

The surviving leaders are moving well beyond simple labeling. Snorkel now talks about expert-created tasks, reinforcement-learning environments, evaluation rubrics and a broader data factory for frontier AI. Encord's $60 million Series C was pitched around AI-native data infrastructure for physical AI.

Fewer investors appear interested in funding another generic annotation business. Platforms that can handle difficult workflows, proprietary data, evaluation and large-scale production still attract plenty of capital.

As seen above, the category's $60 million current median financing comes from that smaller group of scaled companies rather than from broad startup formation.

Why is expert AI training data suddenly worth so much?

Expert AI training data is becoming expensive because frontier models now need judgment and specialized workflows that cheap crowdwork cannot easily provide.

Mercor describes doctors, lawyers, bankers and other specialists teaching models how professional work is actually done. AfterQuery builds datasets and reinforcement-learning environments across areas such as finance, software engineering and enterprise workflows.

Snorkel is pushing in the same direction. Its latest work focuses on expert-created agent tasks, environments and evaluation criteria that can take hours or days to design properly.

The valuable output may now be a complete workflow, a difficult decision process, an interactive environment or a scoring system for an agent. Producing those assets requires expensive people and more operational infrastructure.

That helps explain why RLHF and expert-feedback companies are attracting much more capital per deal than many conventional labeling startups.

Why is robotics data getting so much investor attention now?

Robotics data is attracting much more funding because physical AI does not have an internet-sized dataset waiting to be scraped.

A language model can learn from vast quantities of existing text, images and code. A robot needs demonstrations of hands manipulating objects, factory processes, motion, sensor data and interactions between people and machines.

Companies therefore have to create that data deliberately.

XDOF is building production-scale robotics datasets and released an open teleoperation dataset with more than 130,000 episodes across 200 manipulation tasks. Config says high-quality training data is the main bottleneck for general-purpose robot foundation models and has built collection operations across several locations. Luel collects video, audio, sensor streams and other multimodal data through a global contributor network.

The money is following the work required to produce those datasets. Config's $27 million round attracted Samsung Venture Investment alongside investment arms connected to Hyundai, LG and SK Telecom. Human Archive raised $8.2 million around sensor-rich real-world human activity data.

Robotics data has started to look like a standalone infrastructure business.

Is synthetic AI training data falling out of favor?

Yes, at least relative to the rest of the AI training data market. Synthetic data has lost a lot of its previous funding momentum.

Synthetic Training Data represented 27.3% of deals in the previous period and only 12.2% now. Its share of known capital fell from 6.7% to 1.9%.

There are still good reasons to generate synthetic data. Secludy uses it to work with sensitive enterprise information, while companies building reinforcement-learning environments can create interactions that would be difficult to collect naturally.

What has weakened is the broad idea that synthetic generation on its own is enough of a venture moat. Investors currently seem more interested when synthetic data solves a specific problem such as privacy, simulation, agent training or lack of access to real-world examples.

Is AI training data funding moving earlier stage or later stage?

AI training data is getting more Seed-heavy in the number of deals while the bulk of the money moves toward later rounds.

Seed rose from 40.9% to 53.1% of deals. Series A moved down from 36.4% to 22.4%. That is a clear increase in very young company formation.

Capital tells another story. Later-stage companies captured 63.6% of known dollars, and Series C plus Series D+ alone represented roughly half of current funding.

Round sizes are climbing too. The median Series A increased from $8.15 million to $25 million, while the median Series B reached about $50 million versus $16 million previously.

The market is becoming younger at the bottom and more capital-intensive at the top.

Stage	Current deal share	Previous deal share	Current capital share	Current median	Previous median
Seed	53.1%	40.9%	12.7%	$8.1M	$5.0M
Series A	22.4%	36.4%	18.6%	$25.0M	$8.15M
Series B	6.1%	13.6%	13.3%	$50.3M	$16.0M
Series C	4.1%	0%	27.2%	$205M	-
Series D+	2.0%	4.5%	23.2%	$350M	$100M
Growth Equity	0%	4.5%	0%	-	$72M
Unknown	12.2%	0%	5.1%	$18M	-

Are AI training data startups raising again faster now?

Some AI training data startups are raising again extremely quickly, but the evidence is still too thin to call faster fundraising a market-wide trend.

Among the seven current follow-ons where we could establish a reliable previous-round date, the median gap was about 4.9 months. Four raised within a year and six within 18 months.

Protege is a good recent example. Its $30 million financing followed a $25 million Series A announced only months earlier. David AI also moved quickly through a $5 million Seed, a $25 million Series A and then a $50 million Series B.

The comparison sample from the previous year contains only three reliable observations. Announcement dates also create noise: some rounds were made public long after they actually closed.

We can confidently say that selected winners are able to return to market very quickly these days. A broader claim about the average company would go beyond what the evidence can support.

Do AI training data companies need more proof to raise a big round now?

Yes. The biggest AI training data rounds today usually come with clear evidence of customer demand, proprietary supply or a production system that would be difficult to reproduce quickly.

Snorkel said its newer data service had grown more than 18-fold before the $350 million round. AfterQuery reported a revenue run rate above $100 million after its Series A. Protege said adoption was expanding across healthcare, media, audio and motion capture before its latest financing.

Physical-AI companies show a different kind of proof. Config is already working with large enterprise customers and operates a full data-to-model adaptation loop. XDOF works with robotics labs and companies on production-scale datasets. Luel has built a global contributor network rather than simply a software interface.

Investors seem comfortable funding experiments with smaller Seed checks. The bar becomes much higher once the round gets large.

Is North America still dominating AI training data funding?

North America still dominates AI training data dollars, although Asia is producing a much larger share of funded companies than it did a year ago.

North America accounts for 65.3% of current deals and about 91% of known capital. The giant Mercor, Snorkel and micro1 financings explain a large part of that dollar dominance.

Asia-Pacific increased from 13.6% to 22.4% of deals. Its reported capital share looks much smaller, but that comparison is unusually unreliable because every current financing without an exact disclosed amount comes from APAC.

That missing-dollar problem affects seven deals, so we should not read the reported 3.2% APAC capital share literally.

Europe remains active but much smaller than North America in funding dollars. The previous period was also helped by Toloka's $72 million transaction, which makes the year-on-year decline look sharper.

Are more investors entering AI training data?

Yes. More investors are showing up in AI training data, and we are starting to see repeat investors build positions across several companies.

We identified 54 distinct named lead or strategic investors in the latest period versus 30 previously. Named lead participations rose from 22 to 45.

General Catalyst, Lightspeed, Framework Ventures, Wing Venture Capital and Oriental Fortune Capital each appeared more than once in the current period. We did not see that kind of repeat lead activity in the previous sample.

Investor disclosure is actually weaker today, so those counts probably understate the current breadth.

The more interesting change is specialization. Some investors are coming back repeatedly around expert data, physical AI or proprietary multimodal data rather than treating training data as one generic category.

Are corporate investors becoming more important in AI training data?

Corporate investors look more selective today, with the clearest activity concentrated around physical AI and strategically scarce data.

We found named strategic investors in seven current financings versus eight previously. The visible participation rate therefore fell from 36.4% to 14.3%, although weaker investor disclosure makes that comparison imperfect.

Config shows why the corporate rounds that remain are interesting. Its financing brought together Samsung Venture Investment, Hyundai's ZER01NE, LG Technology Ventures and SKT America around robotics-data infrastructure.

Those investors have obvious industrial reasons to care about better physical-AI training data.

There is less evidence of broad corporate participation across the whole market today, but the strategic logic is strong in deals tied directly to robotics, manufacturing and proprietary data access.

What are AI training data companies spending all this money on?

AI training data companies are increasingly spending their funding on industrial-scale data production: specialist networks, multimodal collection, reinforcement-learning environments and physical-data operations.

Mercor said its $350 million round would expand its talent network, improve matching and speed up delivery. Snorkel is using its capital to scale a data factory built around expert tasks, environments and frontier-model development.

Luel raised $31.2 million to expand a worldwide contributor network collecting rights-cleared multimodal data. Wirestock's $23 million Series A is supporting a platform built around human-made image, video, 3D and other multimodal data. Config is investing in collection infrastructure and robot foundation models, while XDOF operates further down the physical stack with robotic systems, teleoperation and production-scale datasets.

More of the funding is moving into operations: people, hardware, collection networks and production systems that continuously generate scarce data.

What really changed in AI training data funding over the last 12 months?

AI training data funding became substantially healthier. We found more than twice as many deals and companies, while the median round increased sharply. The acceleration remains strong after removing the biggest financings.

The biggest structural change is the split between new companies and proven ones. Many more startups are receiving first checks, but those initial checks are relatively modest. Most of the incremental capital is flowing into follow-ons, where the median round has more than tripled.

Money is also moving toward a different kind of data. Real-world collection, professional expertise, reinforcement-learning environments and proprietary multimodal datasets have gained momentum as model developers look beyond easily available web data.

Traditional annotation is becoming a narrower business. Large platforms can still raise very large amounts, but new-company formation has moved elsewhere. Synthetic data has also cooled considerably compared with the rest of the market.

Physical AI may be the most important new funding theme. Robotics companies need large quantities of real-world interaction data that do not already exist on the internet, and investors are now financing the infrastructure needed to produce it.

The misleading headline would be that AI training data funding simply jumped 273%. The more useful conclusion is that the market has become broader at entry, much more aggressive when backing winners, and increasingly focused on data that is expensive or operationally difficult to create.