Snorkel AI's $3.5B Round Puts a Price on Enterprise AI Training Data — and Asks Whether Innodata Is the Only Public Way to Play It

Snorkel AI raised $350 million at a $3.5 billion valuation in a round disclosed to Reuters on September 22, 2026, led by Insight Partners and S32 with Addition, Greylock, and Wells Fargo returning. The mark is almost three times the $1.3 billion post-money it set on a $100 million Series D in May 2025, and the company crossed a $350 million annualized run rate — up from about $20 million a year earlier, a roughly 17x expansion in twelve months.

Put alongside its rivals, Snorkel's mark is neither the cheapest nor the most expensive on the table — it is the first one with a published price that matches revenue scale. Surge AI is the bigger ARR story at $1.2 billion in 2024, Scale AI is the most expensive at roughly $29 billion for about $1.5 billion in implied revenue, and the robotics-data pure-play Mecka AI is the cheapest by stage, valuing $100 million of projected 2026 ARR at about $500 million.

Snorkel's growth did not come from selling more human hours. The company launched a data-as-a-service offering in September 2025 and now sells finished datasets and reinforcement-learning environments to frontier AI labs, hyperscalers, enterprises, and the US federal government. The shift from labor billing to product billing is the same one Surge AI rode to $1.2 billion in 2024 revenue on the back of roughly twelve frontier AI labs.

Micro1 illustrates the other end of the same shift: gross run rate of $500 million masks roughly $150–200 million of net run rate, with 80–90% gross margins only on the small slice of synthetic, off-the-shelf data. Snorkel is positioning to capture that higher-margin slice before Micro1 can fully convert.

Training data is a three-tier stack and the Snorkel round prices the middle. Specialist networks and labeling platforms sit upstream; the data-development platform sits in the middle; frontier AI labs, hyperscalers, and the US federal government sit downstream. Every dollar of ARR at Snorkel is a dollar not spent at Scale AI, Surge AI, or Mercor — and the upstream specialist labor pool is shared across all of them.

The downstream split is the part that breaks Scale AI's old lock. After Meta paid $14.3 billion for 49% of Scale AI in June 2025, Google — Scale's largest customer — began splitting away, OpenAI pulled back, and xAI cut back. Frontier labs migrated to Surge AI for RLHF and to Mercor for expert data; Snorkel is the third neutral, and the only one with a fresh, larger, published mark.

Only one listed company is a clean AI training-data pure-play, and the Snorkel round validates it. Innodata reported record Q2 2026 results on August 6, 2026: revenue of $92.1 million, up 58% year over year, with adjusted EBITDA of $25.4 million at a 27.5% margin. The release called out the same trend Snorkel is selling — off-the-shelf datasets and high-value pre-training programs — and ended the quarter with $250.4 million in cash, up $168.2 million since December 2025.

The hyperscalers and the federal-AI complex sit one step further out. Microsoft, Alphabet, and Meta are direct buyers of Snorkel-style finished datasets; Meta is also the 49% owner of Scale AI, which makes the Snorkel mark a quiet pressure point on Meta's own stake. Palantir is the cleanest federal-AI proxy — Q2 2026 US government revenue grew 90% year over year to $809 million, and Snorkel's GSA MAS presence positions the two to compete, and partner, on the same defense RFPs.