The workforce behind frontier model quality
Micro1 supplies the human layer that grades and refines AI model outputs, contracting doctors, lawyers, scientists, and other domain experts to perform reinforcement learning from human feedback, the process of scoring model responses so a lab can fine-tune toward better answers. It is also building synthetic data generation and robotics pre-training datasets, recording generalists interacting with everyday objects to give physical AI systems a training corpus that does not exist naturally at scale.
The growth rate is the headline. Micro1 went from a 100 million dollar annual run rate to 500 million dollars in eight months, and the company retains 60 to 70 percent of gross revenue as net revenue. That is not the profile of a labor arbitrage business; it is the profile of an infrastructure company that happens to route through human contractors rather than servers, and it is growing at the pace of the model training demand sitting upstream of it.
Why margins vary so much inside this business
The most revealing detail in Micro1's numbers is the margin split. Custom, contract-specific labeling work for a single AI lab client is labor-intensive and lower margin. Off-the-shelf datasets, built once and resold to multiple customers, run 80 to 90 percent gross margins, closer to a software business than a services one. That split tells you where the real value creation is happening in this market: reusable, standardized evaluation and training data, not bespoke one-off engagements.
For a buyer, that split is also a due diligence signal worth asking about directly in any vendor conversation. A vendor whose growth comes primarily from high-margin reusable datasets has built durable infrastructure that scales cleanly with demand. A vendor whose growth comes primarily from custom, single-client contract work is closer to a staffing agency wearing an AI label, with margins and quality consistency that scale far less predictably as demand grows and that depend heavily on how quickly it can recruit qualified domain experts for each new contract.
This market is bigger than one company
Micro1 is not the largest player in this market, which is itself the more important data point. Mercor reports roughly 2 billion dollars in gross annualized revenue and Handshake around 1 billion, both well ahead of Micro1's 500 million, and all three are growing at once rather than taking share from each other in a zero-sum fight. Researchers cited in coverage of the space project that future AI spending on data could eventually rival spending on compute, which reframes training and evaluation data from a line item most enterprises still treat as an afterthought into a capital category worth the same forecasting and scrutiny currently given to GPU procurement.
That reframing matters for any enterprise building or fine-tuning its own models, not just the handful of frontier labs currently driving the headline numbers. If data spend is genuinely heading toward parity with compute spend industry-wide, the vendor relationships underpinning that spend, covering contractor quality, geographic sourcing, and data provenance, deserve the same governance rigor currently applied to cloud and chip vendors during procurement, and based on how most organizations currently handle data vendor contracts, that level of scrutiny has not caught up yet.
The geopolitics hiding inside a labeling contract
Micro1's founder, Ali Ansari, drew a sharp line: the company does not sell its data to Chinese model makers, and he said some competitors sell data worth millions to what he called adversarial countries. Whether or not that framing is entirely fair to competitors, it surfaces a real question enterprise buyers rarely ask: where does the training or evaluation data your AI vendor supplies actually originate, and who else has access to the same contractor pool and dataset.
For a regulated enterprise, especially one operating in defense-adjacent, financial, or critical infrastructure sectors, data licensing terms from a training data vendor are no longer a purely commercial detail buried in an appendix. They are a supply chain security question with the same shape as questions already being asked about semiconductor sourcing and cloud data residency, and vendor contracts in this space have generally not caught up to that reality yet, leaving most procurement teams negotiating price and turnaround time while leaving licensing and resale terms largely unexamined.
What CTOs evaluating data vendors should verify
Before signing with a training or evaluation data vendor, ask directly what share of their revenue comes from reusable, off-the-shelf datasets versus custom contract work, since that ratio predicts both pricing stability and quality consistency over time. Ask who else licenses the same underlying dataset, since a dataset resold across competitors can quietly homogenize model behavior across an entire industry in ways that are difficult to detect after the fact, and ask how the vendor sources and retains the domain experts doing the actual labeling work.
Also ask, explicitly and in writing, about geographic restrictions on data resale and contractor sourcing, particularly if your organization operates in a regulated or security-sensitive sector where data provenance could become a contractual or regulatory liability later. The data layer behind AI model quality has grown into a market approaching the scale of compute spend, and it is past time procurement and security teams applied the same scrutiny to it that they already apply to cloud and chip vendors during a platform selection process.
