rationale:

Establishing the 2026 baseline relies heavily on vendor revenue disclosures and run rates, as labs do not publish exact human-data spending. The top five frontier developers likely spend an aggregate of $5–10 billion in calendar 2026. This is built from major vendor footprints: Mercor reached a ~$2 billion gross run rate by mid-2026 with ~90% tied to foundation-model labs Mercor revenue, valuation & funding | SacraMercor revenue, valuation & funding | Sacra, Surge carried a ~$1.4 billion run rate, Handshake neared $1.1 billion $1.1B/year Indeed for data labelers - Sacra, and Scale expects ~$1 billion supported by a $450 million minimum guarantee from Meta https://www.forbes.com/sites/richardnieva/2026/05/14/scale-meta-deal/Mercor vs Surge vs Scale AI: 2026 Buyer Guide. Crucially, because this market grew steeply over the course of 2026, the late-2026 exit run rate is significantly higher than the annual average. Even if nominal spending flatlines entirely from late 2026 onward, the 2029 to 2026 ratio would inherently settle around 1.2–1.5x.

The primary driver for sustained multiplier growth is the sheer scale of expanding R&D budgets paired with a shift toward costly, high-fidelity reinforcement learning (RL) environments. Labs are moving away from simple static labeling toward complex, simulated tasks with automatic scoring, which demand premium expert compensation. Rare specialists can command $500–$1,000 per hour, and complex individual tasks can cost up to $20,000 Data Annotation for AI Labs: Recruiting Guide 2026An FAQ on Reinforcement Learning Environments. Anthropic alone has reportedly discussed spending over $1 billion in a single year on RL environments Why RL Environments Became AI's Hottest Bottleneck in .... If human data holds a roughly constant fraction of aggregate computing budgets—debates among observers often place the compute-to-data split around 10:1 or 20:1 Ryan Greenblatt – What happens once AI can automate AI ...—the sheer volume of planned infrastructure commitments could easily carry human data spending up 3–5x.

However, powerful countervailing forces will likely compress spending growth as AI capabilities advance. The greatest downside risk to the multiplier is the substitution of human data generation with synthetic data and AI-built environments, reinforced by structural shifts toward AI-generated synthetic data and the capture of authentic enterprise workflows. Some observers, such as Ryan Greenblatt, argue that scaling human-expert data has not been the primary driver of progress and that AI labor is increasingly capable of building these environments Ryan Greenblatt – What happens once AI can automate AI ...Ryan Greenblatt – What happens once AI can automate AI .... Simple labeling is already estimated to be ~80% automated https://blog.pebblous.ai/report/expert-data-labor-market-2026/en/. A projected 4.6x improvement in sample efficiency would further dampen human-data budgets by allowing labs to extract more capability from each unit of data. While humans currently remain necessary for post-training checks, drafting grading rubrics, and evaluating edge cases The Human Data Market in 2026: Who Buys It, Who Makes ...Data Annotation for AI Labs: Recruiting Guide 2026, AI R&D automation could strip away much of this remaining human-in-the-loop work by 2028–2029. Additionally, as labs bring more of this work in-house to bypass vendor margins of 30–35% Mercor vs Surge vs Scale AI: 2026 Buyer GuideAn FAQ on Reinforcement Learning Environments, imperfect vendor-revenue proxies may undercount true spending.

The median estimate of 2.8x balances these explosive budget expansions against the accelerating efficiency of AI-assisted data generation, the projected 4.6x improvement in sample efficiency, the growing role of AI-generated synthetic data, and the capture of authentic enterprise workflows. This central case assumes a decelerating but robust growth path, with the structural shift away from commissioned human data limiting the multiplier relative to what infrastructure growth alone might imply. The 25th and 75th percentiles are 1.6x and 4.8x, respectively, spanning outcomes from substantial substitution toward synthetic data and authentic workflows to continued strong demand for high-fidelity human input. The 10th percentile rests at 1.1x, reflecting the reality that a drop below 1.0 requires nominal spending to collapse well below early-2026 levels, which is highly improbable given current run rates. Conversely, the 90th percentile stretches to 7.2x, capturing a scenario where human validation remains an unyielding bottleneck, forcing data spending to scale forcefully in tandem with massive compute buildouts, though sample-efficiency gains temper the upper tail.