Mercor closed H1 2026 having built the densest expert-contractor network in the US AI training economy — 30,000-plus credentialed professionals paying out more than $3 million daily (per Mercor public company profile and company-cited figures) — while simultaneously weathering the most significant security incident in the company's three-year history. The revenue number is real: $1.5 billion in annualized recurring revenue as of May 2026, up from $760 million at end-2025 and zero in early 2023. The breach is real too. On March 27, a supply-chain attack via the LiteLLM Python package exposed up to 4 terabytes of internal data and contractor PII affecting 40,000-plus workers (per cybersecurity reporting by Cybernews and DoControl, April 2026), triggering an indefinite pause from Meta and five contractor class-action lawsuits filed within a single week in April. Both facts describe the same company, and understanding the gap between them is the central question for anyone watching the RLHF labor market going into H2.
What Happened
Mercor was founded in January 2023 by three college dropouts — Brendan Foody, Adarsh Hiremath, and Surya Midha — who were high school debate teammates in San Jose. The company's trajectory from that starting point to a $10 billion Series C valuation in October 2025, led by Felicis with participation from Benchmark and General Catalyst, set all three founders up as among the youngest self-made unicorn founders in US tech, per press coverage at the time of the Series C. The October round raised $350 million. The company's current contractor network — which it describes as over 300,000 credentialed professionals including engineers, doctors, lawyers, scientists, and financial analysts — serves six of the Magnificent Seven tech companies and all four major frontier AI labs: OpenAI, Anthropic, Meta, and Google DeepMind.
The revenue arc is without precedent at this scale. Mercor crossed $500 million in annualized run rate within 17 months of founding, then doubled to $1 billion in early 2026, then added a further $500 million in ARR in roughly 60 days following the March breach, per reporting from the 20VC podcast appearance by Foody on June 1, 2026. That final data point — $500 million in net-new ARR in the two months following the breach — is either the most counterintuitive metric in the H1 tech hiring story or the most clarifying one. Foody told host Harry Stebbings on June 1 that "strong client relationships are key during crises," and the ARR trajectory validated that framing, at least partially: the breach became a forcing function for Mercor to build enterprise-grade security infrastructure it had not previously needed.
The contractor pay structure sits at the top of the RLHF market. Mercor's average platform pay rate surpassed $100 per hour — a figure Foody announced via LinkedIn in late 2025 — compared to $22-$25 per hour for generalist annotation work on Scale AI's Outlier platform. The rate differential reflects Mercor's explicit positioning around domain expertise: board-certified physicians reviewing medical AI outputs earn $90-$160 per hour on the platform; credentialed attorneys evaluating legal reasoning earn in the same range; advanced engineering reviewers sit at $75-$130 per hour depending on specialization. Generalist RLHF preference annotation on Mercor runs $40-$55 per hour, per public rate data aggregated by HireFeed as of early 2026.
The company operates with 300 (estimated) internal employees against that contractor base — a ratio that defines its margin profile. It is a coordination and quality-control layer, not a staffing firm. The February 2026 acquisition of Sepal AI (confirmed via Orrick legal announcement, February 6, 2026), a data research company focused on long-horizon AI research tasks, added capability on the evaluation and agentic task side of the RLHF workflow that Mercor's existing infrastructure did not cover.
The breach itself was confirmed March 31, 2026. The threat group TeamPCP published two malicious versions of the LiteLLM open-source package to PyPI, compromising Mercor's integration with the library. Exposed data included contractor profiles, video interview recordings, source code, API keys, and what security researchers described as potentially the proprietary AI training methodologies of Mercor's lab clients. The forensic exposure was worst for contractors: 40,000-plus workers had their personally identifiable information compromised. Meta suspended all work with the platform indefinitely in early April. OpenAI opened an investigation. The combination of a public client freeze from the company's largest lab customer and five lawsuits filed inside seven days made April 2026 the most publicly difficult month in Mercor's history, as TechCrunch reported on April 9.
Why It Matters
The Mercor H1 arc is useful for the broader AI hiring market not because the breach is unusual — supply-chain attacks via open-source packages are structurally common in AI infrastructure — but because of what the post-breach ARR trajectory reveals about the RLHF labor economy's concentration.
Frontier labs have built their model training pipelines around a small number of expert-contractor platforms. Mercor and Scale AI together control the majority of credentialed-professional RLHF supply in the United States. Meta's $15 billion acquisition of Scale AI in June 2025 crystallized labeling infrastructure as a strategic moat, in the characterization used by Sacra's equity research team. When Mercor's platform went into forensic lockdown in late March, the labs that depended on it for specialized domain expertise — medical, legal, scientific — had no equivalent alternative at the same credentialing tier. That supply constraint is part of why the breach did not produce permanent client attrition: the alternatives were not interchangeable, and the labs needed the network back.
Foody's rejection of the $30 billion acquisition offer — confirmed in the June 1 interview with Stebbings — reads in this context as a bet on that concentration becoming more valuable, not less. His stated view is that Mercor is infrastructure, not a feature to be absorbed: "Token spend will exceed headcount spend in five years," he told Stebbings, framing the RLHF network as the supply chain for that token economy rather than as a temporary bridge between human labor and full automation. Whether that thesis holds depends on whether frontier labs continue to require large volumes of credentialed human feedback as models scale, or whether synthetic data and model-to-model evaluation replace human preference signals in post-training. That is the open empirical question behind Mercor's $1.5 billion ARR.
The internal hiring signal embedded in Mercor's H1 data is the contractor pay floor. Average platform rates above $100 per hour for what the company calls "a new category of work" mean that the most skilled tier of RLHF workers is now earning at the lower end of what entry-level software engineers earn at frontier labs in base salary — but doing so on flexible contractor terms, without the geographic constraint of San Francisco or Seattle, and with the ability to stack multiple platform engagements simultaneously. Physicians and attorneys who have never worked in tech are now a meaningful slice of the RLHF labor force. That is a structural shift in who the AI training economy employs, not just a rate arbitrage.
What's Next
1. Meta reinstatement timing. Meta's pause was described as indefinite, and as of late June there is no public announcement of reinstatement. Meta is also Mercor's single highest-profile client, given the company's absence from Scale AI's client roster post-acquisition by the same parent. If Meta's security team clears Mercor's infrastructure by end-Q3 2026, the ARR trajectory has room to accelerate; if the pause extends into Q4, the $1.5 billion figure becomes a ceiling rather than a floor.
2. Synthetic data competition. The RLHF platform market is projected to grow from $2.8 billion in 2025 to $18.6 billion in 2034, per Mordor Intelligence. (Note: Mordor Intelligence methodology not publicly disclosed; figure cited for order-of-magnitude context only.) But that projection assumes sustained demand for human preference signals. Every major frontier lab is running parallel research programs on synthetic data generation and model-based evaluation. If post-training workflows shift materially toward synthetic feedback in H2 2026, volume demand for credentialed-professional annotation will compress faster than the market-size figures suggest. Mercor's pivot toward longer-horizon research tasks — the logic behind the Sepal AI acquisition — is a hedge against that compression.
3. Lawsuit resolution as SOC 2 forcing function. The five class-action lawsuits filed in April require Mercor to document its data security practices in legal filings. That process, combined with the forensic investigation already underway, will either produce a credible enterprise security posture that the company lacked before March 27, or it will surface additional exposure that constrains future client onboarding. Either outcome moves faster than Mercor's Q4 2025 timeline would have suggested. The breach has effectively accelerated the compliance roadmap by 12-18 months.
Mercor's H1 2026 is a concentrated version of what the US AI training economy looks like under pressure: extraordinary revenue velocity, thin internal headcount relative to the contractor base it orchestrates, and structural exposure to the security practices of every open-source package its platform depends on. The $1.5 billion ARR survived the breach. Whether it survives the next 18 months of synthetic data competition is the question that matters more.
Global Career PlatformFind AI talent. Find your next role.
Booking is hotels. · Airbnb is apartments. · ENTRA is global careers.