Customer Insourcing Threatens Datacurve

The core risk is that Datacurve is selling a capability, not a durable system of record, so a few frontier labs changing build versus buy decisions could remove a large share of demand at once. Datacurve provides high quality coding data for training and evals, but that category is increasingly shaped by labs with huge internal research teams, growing post training infrastructure, and incentives to keep sensitive data generation in house.

• The closest analogs show how concentrated and fast moving this market can be. Scale grew from $215M ARR in 2022 to $760M in 2023 and $1.5B in 2024 by catching the LLM data wave, while Invisible reached $134M in 2024 on major RLHF contracts. That speed cuts both ways when budgets shift.
  2 sacra 5 sacra
• The workflow itself is easy for a big lab to internalize. Instead of buying expert written coding tasks from an outside vendor, a lab can hire researchers, build eval harnesses, generate synthetic examples, then use humans mainly for validation. Interviews across the ecosystem show synthetic data is scaling, while human input is moving toward narrower expert review.
  6 sacra 7 sacra 8 sacra
• Specialization still matters, but it narrows the addressable budget. Prolific is positioned around vetted humans for nuanced evaluation, and Surge and Scale are much larger multi product platforms. Datacurve is younger, founded in 2024 with $17.7M in funding, which makes customer concentration more acute because it has less product breadth to absorb buyer pullbacks.
  1 sacra 2 sacra 3 sacra 9 sacra

The next phase of this market favors vendors that become part of the customer workflow, not just a line item in the training budget. Datacurve is best positioned if it moves from selling raw coding datasets toward ongoing evals, expert validation, and audit ready data work that frontier labs will keep outsourcing even as more generation moves in house.