AI training data funding hit $450 million in a single day on September 22, 2026. Snorkel AI raised $350 million at a $3.5 billion valuation, per TechCrunch. Hours earlier, Forbes reported Micro1 closed over $100 million at $4 billion — 8x its September 2025 mark. Both sell the same thing: human expertise, packaged as data, to a handful of frontier labs.
The model layer gets the headlines. The data layer is where the money moved this week.
Two companies most consumers have never heard of added roughly $5.7 billion in paper value between them in twenty-four hours. Neither trains a frontier model. Both sell the raw material that makes one work.
How much did Snorkel AI and Micro1 raise?
Snorkel AI raised a $350 million Series E at a $3.5 billion valuation. Micro1 raised more than $100 million at $4 billion. Combined, that is over $450 million of AI training data funding announced on a single day, against a combined valuation of $7.5 billion.
Snorkel AI’s $350 million Series E
According to TechCrunch, the round was led by Insight Partners and S32, with Addition, Lightspeed, Greylock, GV and Wells Fargo participating.
The markup is the story. Snorkel’s previous round — $100 million at a $1.3 billion valuation — closed 17 months earlier. That is a 2.7x step-up in under a year and a half.
TechCrunch reports the company now runs at $375 million in annualized revenue, an 18-fold increase over the previous twelve months. Snorkel commercialized in 2019 after four years of Stanford research, and is led by co-founder and CEO Alex Ratner.
Micro1’s $100 million round at $4 billion
Forbes reports Micro1 raised over $100 million at a $4 billion valuation, up from $500 million in September 2025 — an 8x markup in twelve months.
Founder Ali Ansari is 25 and runs the company solo. Forbes traces the revenue line: $7 million ARR at the start of 2025, past $100 million by December 2025, and a $500 million gross annual run rate by August 2026.
Investors in the new round include two frontier AI labs and two xAI cofounders, according to Forbes. Microsoft, Amazon and robotics firms including 1X are listed as customers.
What do the numbers actually look like side by side?
Four companies now dominate the AI training data market. Their disclosed revenue and valuation figures differ enough that the multiples only make sense once you separate gross billings from what the vendor keeps.
| Company | Latest valuation | Latest raise | Reported revenue run-rate | Implied multiple |
|---|---|---|---|---|
| Micro1 | $4B (Sep 2026) | $100M+ | $500M gross / $150–200M net | 8x gross, 20–27x net |
| Snorkel AI | $3.5B (Sep 2026) | $350M | $375M annualized | ~9.3x |
| Mercor | $20B (in talks, Jul 2026) | $500M | $2B gross annualized | ~10x gross |
| Scale AI | ~$29B implied (Jun 2025) | $14.3B from Meta | 2026 guidance: just over $1B | ~29x guidance |
Why is AI training data funding exploding now?
Because pre-training data ran out and post-training data did not. Labs have scraped what the public web offers. The marginal gain now comes from expert-generated reasoning traces, reinforcement learning environments and domain evaluations — none of which exist until someone pays a specialist to produce them.
That shifted the business from labeling to sourcing scarce human judgment. The work now looks like this:
- Expert contracting. Micro1 recruits engineers, doctors and lawyers to generate domain data, per Forbes.
- Reinforcement learning gyms. Simulated work environments built from real enterprise data, including Slack and email archives from defunct companies.
- Dataset acquisition. Forbes reports Micro1 attempted to buy Spirit Airlines’ bankruptcy-estate operational data for $12.5 million.
- Robotics annotation. Paying $50 to $90 an hour for the physical-world data humanoid developers cannot scrape.
- Synthetic generation. Snorkel’s hybrid model pairs synthetic data with subject-matter experts, which is what let it pivot from software to data-as-a-service.
The pivot matters more than the pitch decks suggest. Snorkel spent years selling labeling automation software. It got to $375 million by selling the output instead of the tool.
Who profits from the AI training data boom?
Founders and early investors, most visibly. But the cash flow tells a more complicated story: most of the revenue these companies book leaves again as contractor pay.
TechCrunch notes that competitors in this category typically pay out 60% to 70% of top-line revenue to the domain specialists doing the work. Snorkel books those payments in cost of goods sold rather than netting them out of revenue — which is why its $375 million figure is not directly comparable to Micro1’s $500 million gross.
Micro1’s own net run-rate sits between $150 million and $200 million, TechCrunch reported in August. Synthetic and off-the-shelf datasets carry 80% to 90% gross margins; human-sourced work does not.
Strip that out and a $4 billion valuation on $150–200 million of retained revenue is a 20x to 27x multiple — priced like software, delivered like staffing. That is the single most aggressive assumption in this week’s AI training data funding.
Compare it with the markups we tracked at Crusoe’s Series F and the ARR multiple behind the Cohere-Aleph Alpha merger, and the data layer is no longer the cheap part of the stack.
What could go wrong with these valuations?
Customer concentration. These vendors sell to perhaps a dozen buyers worldwide, and the largest cautionary tale in the sector is only fifteen months old.
The Scale AI precedent
Meta paid $14.3 billion for a 49% non-voting stake in Scale AI in June 2025, implying roughly $29 billion. Scale did $870 million in 2024 revenue and reached about a $2 billion run rate in 2025.
Its 2026 guidance is just over $1 billion. Google, OpenAI and xAI pulled back or exited within weeks of the Meta deal closing, unwilling to route proprietary training pipelines through a vendor half-owned by a competitor.
Revenue in this sector can halve on a single ownership decision. That is unusual for a business trading at a software multiple.
Investor-customers cut both ways
Forbes reports two frontier AI labs invested in Micro1’s round. That secures demand and creates the exact conflict that cost Scale three flagship accounts.
A data vendor partly owned by one lab is a harder sell to its rivals. The more successful these rounds are at locking in strategic money, the narrower the addressable customer list becomes.
The labs can build it themselves
Nothing about expert recruiting is proprietary. If post-training data becomes the primary differentiator, in-housing it is the obvious move — and the labs have the capital to do it.
Why this matters for the AI market
The data layer is now a genuine profit pool, not a cost line. Roughly $7.5 billion of value was repriced in one day, and the sector’s disclosed run-rates — $2 billion at Mercor, $500 million gross at Micro1, $375 million at Snorkel, $1 billion at Handshake — imply an addressable market that has grown several times over in eighteen months.
For investors, three things follow.
- Capex is not the only AI spend. Alongside chips and datacenters, labs are routing billions into human-generated data with far less disclosure.
- Gross versus net is the whole argument. Any comparison of these companies that ignores contractor payout ratios is comparing different businesses.
- Private marks are moving faster than exits. As we noted in our look at the 2026 AI IPO market, markups are outpacing the liquidity available to realize them.
The European comparison is instructive too: Mistral’s Series D valued a full-stack model developer at €21 billion. Two data suppliers are now worth more than a third of that between them.
This post is reporting and analysis, not financial advice.
Frequently asked questions
How much did Snorkel AI raise in September 2026?
Snorkel AI raised a $350 million Series E at a $3.5 billion valuation, led by Insight Partners and S32, according to TechCrunch.
What is Micro1’s valuation?
$4 billion, per Forbes, following a round of more than $100 million. That is 8x its $500 million valuation from September 2025.
What is the difference between gross and net revenue for AI data companies?
Gross run-rate is total billings. Net is what the vendor keeps after paying contractors — typically 30% to 40% of the total, since 60% to 70% flows to domain specialists.
Who is the largest AI training data company?
By disclosed revenue, Mercor, which said it crossed $2 billion in annualized revenue in June 2026. Forbes reported it was in talks for $500 million at a $20 billion valuation.
What happened to Scale AI after Meta invested?
Meta paid $14.3 billion for 49% non-voting in June 2025. Google, OpenAI and xAI reduced or ended their business, and Scale’s 2026 revenue guidance fell to just over $1 billion from an approximately $2 billion run rate.
Why do AI labs need human training data if synthetic data exists?
Synthetic data carries 80% to 90% gross margins but cannot originate expert judgment in law, medicine or engineering. Vendors use hybrid models — synthetic scale, human ground truth.
Is the AI training data market a bubble?
The revenue is real and growing fast. The risk is concentration: a dozen buyers, high contractor costs, and multiples set as if these were software companies.
The bottom line
AI training data funding has repriced the least glamorous layer of the AI stack in a matter of months. Snorkel’s 18x revenue growth and Micro1’s 8x markup are not speculative bets on a future market — they are responses to revenue that already exists.
What is speculative is the multiple. Paying 20x to 27x retained revenue for a business that passes most of its billings to contractors assumes those margins improve as synthetic generation scales. They might.
Watch two things next. Whether Mercor’s $20 billion round closes at that number, and whether any frontier lab announces it is bringing expert data generation in-house. The first would validate the repricing. The second would end it.
Sources
- TechCrunch — Snorkel AI triples valuation to $3.5B as demand for AI training data booms
- Forbes — This 25-Year-Old Raised Over $100 Million For His AI Data Startup At A $4 Billion Valuation
- TechCrunch — AI data startup Micro1 reaches $500M gross run rate
- Forbes — AI Data Labeler Mercor In Talks To Raise $500 Million At $20 Billion Valuation
- Value Add VC — Scale AI: Meta Paid $14.3B, Now Guiding to Just $1B
Leave a Reply