Scale AI processes reinforcement learning from human feedback across a global contractor workforce exceeding 240,000 people, managed through its Remotasks and Outlier platforms — making it the single largest employer in the AI training supply chain (Sacra, Scale AI revenue and workforce data, 2026). The data labeling and AI training market reached $3.77 billion in 2024 and has continued expanding (AI Training Data market analysis, WEF 2026 Future of Jobs Report). Industry estimates for the total US-based AI training and data annotation contractor pool — aggregating Scale AI, Mercor, Surge AI, DataAnnotation, Alignerr, and smaller platforms — now range as high as 500,000 workers (ENTRA Intelligence market estimate, September 2026; corroborated by WEF's explicit classification of "Data Annotator" as a growth job category in its 2026 report). Every major frontier model released this year was shaped, in part, by that workforce. None of them received a return-to-office memo.
September 2026 closes today as the final month of the "RTO Wars" era — a twelve-month cycle in which Anthropic, OpenAI, Meta, Amazon, and Apple each ratcheted up in-office requirements for their full-time employees. The workforce that actually generates the training signal for those labs' models operated, the entire time, from kitchen tables in Austin, spare bedrooms in Chicago, and co-working spaces in Des Moines. October opens the AI Trainer Economy theme. The two stories are not separate.
The Workforce Powering RLHF
Scale AI is the structural anchor of the US AI training market. Meta's June 2025 purchase of a 49% non-voting stake, at a $29 billion valuation and a reported $14.3 billion price tag, cemented Scale's position — and immediately complicated it (ValueAddVC, "Scale AI: Meta Paid $14.3B, Now Guiding to Just $1B," 2026). Founder-CEO Alexandr Wang departed to join Meta; Jason Droege now leads as CEO. The neutrality loss was immediate: OpenAI, Google, and xAI began routing training work toward Mercor and Surge AI within weeks of the deal closing.
Scale's contractor footprint remains the largest in the industry. Its Outlier platform, which handles higher-skill RLHF tasks including code evaluation and domain-expert annotation, employed 21,485 workers in 2026 — down 6.2% from 2025, reflecting both automation of simpler annotation tasks and a shift in project mix (Revelio Labs, Outlier AI headcount data, 2026). Forty-two percent of Outlier's workforce is based in North America.
Mercor has emerged as the platform most aggressively capturing volume displaced from Scale. As of late August 2026, Mercor's advertised average contracted rate was $109 per hour — down from a peak of $141 per hour in early June 2026, after the company relabeled the figure from "headline rate" to "average contracted rate" and reset the baseline in late June (Mercor rate tracking, UsagePricing.com, August 2026). Approximately 30,000 independent contractors work through Mercor's platform, with the median earner sitting in the $40–$75 per hour band for standard RLHF and code evaluation work (Mercor, AI Training Compensation Economics & Pay Trends, 2026).
Surge AI manages approximately 50,000 contractors globally, with a reputation for high-curation specialist work — including domain expert projects paying medical fellows $250–$450 per hour and senior executives $500–$1,000 per hour for evaluation tasks requiring C-suite or clinical expertise (AI Training Jobs 2026, TheAIRankings.com). xAI cut roughly 500 of its approximately 1,500 generalist data annotators in September 2025 and announced it would "surge our specialist AI tutor team by 10x," a formulation that telegraphed the structural shift before it became consensus (public xAI announcement, September 2025).
The RTO Irony
September 2026's dominant narrative in US AI hiring was the in-office enforcement season. Anthropic locked 92% of open roles to San Francisco, New York, or London. Meta went five days globally from February 2026. Amazon held its five-day mandate through the year, sustaining the ML attrition documented in this publication across Q1 through Q3. The word "mandate" appeared in more CHRO communications this quarter than any comparable period since 2022.
The AI trainer workforce did not register this policy cycle, because it could not. The work is delivered through digital platforms — Scale AI's Remotasks interface, Outlier's task dashboard, Mercor's contract portal. There is no office to return to. The geographic distribution is structural: a data annotation task requires a human with domain expertise and reliable internet access, not a badge swipe at a San Francisco campus. Outlier's North American workforce of roughly 9,000 is spread across all 50 states. Mercor recruits verified experts by credential, not by ZIP code.
This creates the defining irony of the current moment. The frontier labs enforcing the most aggressive in-office policies for their full-time employees are simultaneously running training pipelines that depend entirely on a permanently distributed contractor layer. Anthropic's Constitutional AI refinement, OpenAI's instruction-following evaluation, Google's Gemini preference data — each requires thousands of human judgments generated by workers who have never been to 575 Market Street or 3180 18th Street. The RTO debate consumed HR bandwidth at every lab this year. The training workforce was not part of that conversation, because it was never included in it.
Platform Economics: The Pay Stack
Pay rates in the AI trainer economy span four orders of magnitude, sorted by task complexity and credential verification.
Generalist annotators — text classification, image labeling, sentiment tagging — earn $15–$25 per hour across Scale AI's Remotasks, DataAnnotation, and similar platforms (TheAIRankings.com, AI Training Jobs 2026). Language specialists with verified bilingual skills command $25–$28 per hour. Coding and CS-focused RLHF evaluators — the fastest-growing segment — earn $50–$65 per hour at Scale AI and Outlier, reflecting demand from every lab for feedback on code generation quality. At the specialist tier, Mercor's most active segment pays $75–$200 per hour for verified domain experts, with board-certified physicians, credentialed attorneys, and FAANG-alumni engineers occupying the upper band (Mercor, AI Trainer Salary: Hourly Rates & Ways to Increase Pay, 2026).
The platform with the clearest compensation-transparency record is Mercor: its published rates fluctuated visibly across Q2 and Q3 2026, moving from $99/hr to $141/hr to a reset at $80/hr before climbing back toward $109/hr by late August. That volatility reflects real supply-demand dynamics — Mercor absorbed a significant share of the project volume that previously ran through Scale AI's Outlier platform after Meta's investment, and the rate movement tracked that influx (UsagePricing.com, Mercor rate history, June–August 2026).
AI agent training has emerged as the highest-growth segment in 2026. Teaching models to complete multi-step autonomous tasks requires richer, longer-horizon feedback than single-turn annotation, and the pay reflects it: agent training roles on Mercor and Surge now open at $80 per hour with specialist rates above $150 per hour (AIGigJobs.com, "The Rise of AI Agent Training," 2026).
What Q4 2026 Signals
Global AI spending is forecast at $2.5 trillion for 2026, a 44% year-over-year increase (Gartner, 2026 AI spending forecast). That number includes compute, software, and services — but human feedback sits inside the services bucket, and the direction of the budget is toward higher-value evaluators, not larger volumes of generalist annotators.
Three indicators will define the AI trainer economy through Q4:
Specialist-to-generalist ratio at Scale AI. With 2026 revenue guidance revised down to approximately $1 billion from an earlier $2 billion projection, Scale AI's project mix under Jason Droege's leadership will clarify whether the Meta alignment has permanently redirected generalist annotation volume or only temporarily disrupted neutrality-sensitive accounts (ValueAddVC, Scale AI revenue analysis, 2026).
Mercor's rate trajectory. The platform's advertised rate has been declining since its June 2026 peak. If Q4 frontier model training cycles — particularly Anthropic's and OpenAI's next-generation releases — pull evaluation demand into the market, rates should stabilize or recover. A continued decline would indicate that automation is compressing the generalist tier faster than specialist expansion can offset.
xAI's specialist training ramp. The company's September 2025 commitment to scale its specialist AI tutor team by 10x from a base of roughly 1,000 implies a Q4 2026 specialist pool of approximately 10,000 — a number that, if achieved, would make xAI the second-largest employer of specialist AI trainers in the US after Scale AI. Verification will come through LinkedIn headcount signals and Levels.fyi job posting data across October and November.
October's AI Trainer Economy coverage in this publication will track all three.
Methodology: Scale AI global contractor headcount from Sacra and ValueAddVC revenue analyses, 2026. Outlier AI headcount and geographic distribution from Revelio Labs, 2026. Mercor contractor count and rate history from Mercor.com published resources and UsagePricing.com rate tracking, June–August 2026. Pay rate benchmarks aggregated from TheAIRankings.com AI Training Jobs guide, Mercor AI Trainer Salary resource, and AIGigJobs.com, all accessed September 2026. Data labeling market size ($3.77B, 2024) from WEF 2026 Future of Jobs Report. Global AI spending forecast from Gartner, 2026. xAI annotator headcount change from public xAI announcement, September 2025. ENTRA's 500,000 US AI training contractor estimate aggregates platform-disclosed figures with US geographic share data from Revelio Labs; it should be treated as an order-of-magnitude estimate, not a census figure.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.