Snapshot

Data labeling companies sell human-annotated training data — tagged text, labeled images, rated LLM outputs (RLHF) — to AI model builders. The global market was $2.25B in 2025, growing at ~33% CAGR toward ~$9.3B by 2030. est. Scale AI (private, $29B valuation after Meta's $14.3B for 49%) dominates. Among the three tickers below, only Appen is a pure-play; Snowflake and Salesforce are adjacent data/AI platforms where labeled data is stored, processed, or consumed — not produced.

The product is labeled data — the raw material that supervised learning and RLHF-based alignment require. Every frontier lab (OpenAI, Anthropic, Google, Meta) buys it. Scale AI has ~$2B revenue est. and 90% from generative AI projects. Appen, the only listed pure-play, has $233M revenue and is loss-making. SNOW and CRM sit downstream as platforms that store and use labeled data.

The product & how money is made

Data labeling means attaching human judgments to raw data so machine learning models can learn from it: drawing bounding boxes around objects in images, rating which of two LLM responses is better (RLHF), transcribing audio, classifying text sentiment, annotating medical images. The output is a labeled dataset fed into a model training pipeline.

How money flows

• Pure-play labeling (Appen, Scale AI, TELUS International): Customer sends raw data + annotation spec. Vendor routes tasks to a crowd or managed workforce. Revenue = per-task or per-hour billing. Costs = annotator wages + platform overhead. Gross margins historically 20–40% for crowd-based work; higher for managed/expert annotation.
• Platform-assisted labeling (Labelbox, SuperAnnotate, V7): Annotation software licenses (SaaS). Customer's own annotators or third-party crowd use the tool. Revenue = subscription ARR. Margins ~70%+ but smaller revenue.
• Adjacent platforms (SNOW, CRM): No labeling services sold. Snowflake stores and queries datasets including labeled ones; Cortex AI runs inference but has no annotation tools. Salesforce Data Cloud / Einstein processes customer data for CRM AI features; no standalone labeling product. Their AI-related revenue (Snowflake Cortex consumption, Salesforce Agentforce $800M ARR) is about deploying AI on enterprise data, not producing labeled training data.

Pricing

• Basic image tagging: $0.02–$0.15 per bounding box depending on region. est.
• Semantic segmentation: $0.10–$3.00 per mask. est.
• Medical imaging: $2–$20 per image (requires domain experts). est.
• RLHF for LLMs: $25–$60+/hr in the US, $5–$15/hr in India, $2–$8/hr in Kenya/Nigeria. Expert RLHF (medicine, law, code) commands $50–$100/hr. est.
• Southeast Asia handles over 58% of global labeling tasks. est.

Source: Second Talent (2026), "Data Annotation Costs by Country"; Mordor Intelligence (2025).

Demand

Contracted and observable

• Appen FY2025 revenue: $233M, down 1% YoY (after -14% in FY2024 and -29% in FY2023). contracted
• Scale AI 2024 revenue: $870M, reaching ~$1.5B annualized run rate by year-end. 2025 estimated ~$2B. 90% of 2024 revenue from generative AI projects. est.
• Meta invested $14.3B in Scale AI (Jun 2025) for a 49% non-voting stake.
• Frontier lab RLHF spending: OpenAI, Anthropic, Google, Meta all use external labeling for alignment. Combined model-training capex across these labs exceeds $100B/yr; labeling is a small fraction. est.

Forecasts est.

• Global data annotation market: $2.25B (2025) → $2.98B (2026) → $9.27B (2030), ~33% CAGR. North America largest region.
• Data labeling tools market: Separately forecasted to reach $22B by 2027 (Mordor Intelligence); this figure includes tooling + services and uses a broader definition.
• Demand mix shift: Moving from simple bulk labeling (image tags, text classification) toward high-skill judgment work (RLHF preference ranking, domain-expert annotation for medicine/law/code). Unit prices rising for skilled work, falling for commodity tasks.

Source: Business Research Company (2026); Mordor Intelligence (2025); Sacra (2025) for Scale AI estimates.

Supply

Capacity

• Global annotator workforce: Millions of contract workers. Appen alone has 1M+ crowd contractors across 170+ countries. TELUS International (TIXT) and Amazon Mechanical Turk operate similarly large pools.
• Basic labeling: Abundant supply in Philippines, India, Kenya, Vietnam. Rates: $2–$15/hr. est.
• Expert labeling: RLHF for frontier LLMs requires annotators who can evaluate complex reasoning, code correctness, medical accuracy, or legal nuance. Qualified pool is small.
• AI-assisted pre-labeling reduces human hours per task by 50–60% (model generates initial labels, human corrects). est.

Competitive landscape

Bottleneck

For commodity labeling: none. For high-skill RLHF and domain-expert work: qualified annotators. PhD-level reviewers for code, medicine, and law cannot be scaled by adding crowd workers.

Synthetic data

AI-generated synthetic data can replace 50–80% of real labeled data for pre-training in some domains (autonomous driving simulation, privacy-sensitive healthcare). est. RLHF preference data — human judgments about which output is better — cannot be synthesized because human preferences define the ground truth. Expert domain annotation (medical, legal) similarly requires specialist knowledge. Hybrid workflows (70–80% synthetic, 20–30% human) are emerging for large-scale projects. Setup cost for synthetic pipelines: $50K–$500K+. est.

Source: Second Talent (2026); Mordor Intelligence (2025); Sacra (2025); Appen website.

The gap

Demand for labeled data is growing at ~33% CAGR. est. Supply of commodity labeling is abundant and prices are falling as AI-assisted pre-labeling and synthetic data reduce per-task cost. Supply of expert RLHF and domain annotation is constrained, and prices are rising.

The gap is bifurcated. Commodity labeling has excess supply and is being automated. Expert RLHF/domain labeling has a supply shortage and rising prices.

Source: Appen 10-K; Sacra (2025); Business Research Company (2026); Second Talent (2026).

The players

SNOW and CRM have zero data labeling revenue. Appen is the only listed pure-play. The dominant player (Scale AI) is private. TDCX (Singapore BPO, ~$1.4B market cap) has limited pure labeling exposure.

Source: SNOW 10-K FY2026; CRM FY2026 earnings release (Mar 2026); Appen FY2025 annual report; stockanalysis.com (Jun 3, 2026).

The price of exposure

Appen (APX.AX) — the only pure-play

• Owner-cash math: At ~US$185M EV and $19M FCF, the business produces about 10 cents of free cash per $1 of enterprise value. At a 15x multiple, that FCF supports ~$285M EV. Current $19M is the first positive-FCF year after three years of cash burn.
• Revenue math: At 0.8x EV/revenue, Appen trades below 1x sales. Gross margins improved from 8% (2023) to 19% (2025). Scale AI trades at ~14.5x revenue on a $29B / ~$2B basis (private). est.
• Context: Revenue has fallen ~40% from the 2021 peak (~$390M). Top customers (historically Google, Apple, Meta, Microsoft) reduced spend. Appen's crowd model is oriented toward volume work; Scale AI captured most RLHF growth.

SNOW and CRM — no labeling exposure

SNOW at $83.6B market cap (16.3x EV/revenue) and CRM at $156.1B market cap (4.4x EV/revenue) provide exposure to data infrastructure and enterprise CRM/AI respectively. Neither generates revenue from data labeling. CRM's Agentforce ($800M ARR) and SNOW's Cortex AI consume labeled data downstream; they do not produce it.

Source: stockanalysis.com (Jun 3, 2026); Appen FY2025 financials; Scale AI via Sacra (2025).

What to deep-dive next

• Appen customer concentration: What percentage of revenue comes from top 3 customers? FY2025 annual report has the breakdown.
• Scale AI post-Meta deal: Google, OpenAI, Microsoft, and xAI reportedly pulled back after Meta's $14.3B investment. Track whether these customers return or diversify to other vendors.
• Synthetic data adoption curve: Monitor whether hybrid workflows (70% synthetic + 30% human) become standard for pre-training. If so, human-labeling TAM growth may slow below the 33% forecast. est.
• Scale AI IPO timeline: If Scale AI goes public, it becomes the first high-growth listed pure-play in labeling and resets valuation comparisons.
• Appen gross margin trajectory: Improved from 8% to 19% in two years. Track whether it reaches 30%+ or stalls.
• AI self-labeling: If frontier models can reliably evaluate their own outputs (constitutional AI, RLAIF), human RLHF demand drops. Timeline uncertain.

Sources & confidence

• Appen: FY2025 annual report; FY2024 results presentation; stockanalysis.com/quote/asx/APX/. Revenue, margins, FCF from filings. contracted
• SNOW: 10-K FY2026 (ended Jan 2026); Q4 FY2026 earnings release (Feb 2026). $4.68B rev, $9.77B RPO, $1.17B FCF. contracted
• CRM: FY2026 earnings release (Mar 2026). $41.5B rev, $7.46B net income, $14.7B FCF, Agentforce $800M ARR. contracted
• Scale AI: Sacra (2025), revenue estimates. Meta investment: public reporting (Jun 2025). est.
• Market sizing: Business Research Company (2026) — $2.25B (2025), 32.7% CAGR; Mordor Intelligence (2025). est.
• Annotation pricing: Second Talent (2026), "Data Annotation Costs by Country." est.
• Synthetic data: Second Talent (2026), "Synthetic Data vs. Human Annotation." est.
• Market data: stockanalysis.com, companiesmarketcap.com (Jun 3, 2026).