Scale AI Vs Surge AI: Which RLHF And Training Data Partner Fits?

Meta Description: Scale AI vs Surge AI compared for RLHF, SFT, evals and red teaming: ownership, workforce models, benchmarks, security, pricing and which one fits your team.

If you are buying human data for LLM post-training, two names come up first: Scale AI and Surge AI. Both supply the preference rankings, expert-written demonstrations, evaluations and adversarial testing that frontier labs use to tune their models. Both are headquartered in San Francisco, and both count the largest AI developers among their customers. Beyond that, they are built very differently.

Scale is a broad AI data and applications company with outside investors, including a large minority stake from Meta, a public sector business with federal authorizations, and data operations that reach well past language models into computer vision, LiDAR and robotics. Surge is a founder-controlled company that, by its own account and press reports, grew without venture funding and has focused almost entirely on high-skill human data for language and reasoning work.

The Scale AI vs Surge AI decision is less about which vendor is “better” and more about which operating model matches your project. This comparison covers ownership, workforce, services, benchmarks, security, pricing and fit. For a wider field of vendors, see our list of the top RLHF and LLM training data companies and the broader ranking of top data annotation and labeling companies.

What Scale AI Is

Scale AI was founded in 2016 by Alexandr Wang and Lucy Guo through Y Combinator. It started as a labeling provider for autonomous vehicle and computer vision teams and now describes itself as having two core businesses: supplying the data used to develop frontier AI models, and building AI applications for enterprises and governments.

Ownership and leadership have changed significantly since mid-2025. In June 2025, Scale confirmed a major investment from Meta that TechCrunch reported valued the company at $29 billion. Reports put Meta’s investment at about $14.3 billion for a 49% stake, which The Information described as non-voting. As part of the deal, Wang left to join Meta’s superintelligence effort while staying on Scale’s board, and chief strategy officer Jason Droege became interim CEO. On July 30, 2026, Scale announced that its board had appointed Francis deSouza, formerly COO of Google Cloud and before that CEO of Illumina, as chief executive effective August 10, 2026, with Droege supporting the transition.

The period after the Meta deal included real disruption. In July 2025, Scale cut about 200 full-time roles, roughly 14% of staff, plus around 500 contractors, mostly in its generative AI data unit, and reports said OpenAI and Google were winding down work with the company. Scale says it has since rebuilt momentum: in a January 2026 post, Droege wrote that the data business was profitable and performed best in the second half of 2025, and that Scale signed over $1 billion in new business that year.

What Surge AI Is

Surge AI, legally Surge Labs Inc., was founded in 2020 by Edwin Chen, a former engineer at Google, Meta and Twitter, who started the company out of frustration with the quality of crowdsourced labeling. Its focus is reinforcement learning from human feedback (RLHF), RL environments and language data, with a workforce built around credentialed experts rather than general crowd workers.

Surge’s defining business trait is its funding model. Reuters reported in July 2025 that Chen had bootstrapped the company, that it had been profitable, and that it brought in over $1 billion in revenue the previous year, compared with about $870 million for Scale over the same period. Reuters and Bloomberg both reported that Surge was in talks for its first outside raise of up to about $1 billion, with Bloomberg citing a valuation of at least $25 billion. Forbes reported in September 2025 that Chen owned about 75% of the company. Surge’s site names OpenAI, Anthropic, Meta and Google as partners.

Scale AI Vs Surge AI At A Glance

Factor	Scale AI	Surge AI
Founded	2016, San Francisco	2020, San Francisco
Ownership	Venture backed; Meta holds a reported 49% non-voting stake	Founder controlled; reported as bootstrapped, Chen holds about 75% per Forbes
Leadership	CEO Francis deSouza since August 2026	Founder and CEO Edwin Chen
Workforce model	Outlier (expert GenAI work) and Remotasks (crowd labeling), plus vendor partners	Curated expert network; site cites 200,000+ PhDs across 500+ disciplines
Core post-training services	RLHF, prompt and response generation, red teaming, evaluation	RLHF, SFT-style demonstrations, human evaluation, safety and alignment data, coding
Data types beyond text	Image, video, LiDAR and 3D sensor fusion, robotics data	Primarily language, reasoning, coding and agent tasks
Public benchmarks	SEAL leaderboards, now under Scale Labs (Humanity’s Last Exam, SWE-Bench Pro, MCP-Atlas)	Surge benchmark suite (Hemingway-bench, Riemann-bench, DAYJOB, sudo L7)
Security credentials	SOC 2 Type II, ISO/IEC 27001:2022, FedRAMP High, DISA IL4	SOC 2 listed on its RLHF page
Government work	Large U.S. defense and international public sector business	Not a stated focus; U.S. Army reported as a customer
Pricing	Not published; sales led via demo	Enterprise priced per task; self-serve sign-up for individuals

Workforce Models Compared

How Scale Sources Contributors

Scale runs two main contributor platforms. Remotasks, launched in 2017, is a crowd work platform originally built for computer vision and autonomous vehicle labeling, with operations in Southeast Asia and Africa. Remotasks ended operations in Kenya, Nigeria and Pakistan in early 2024, according to reports at the time.

Outlier, created in 2023, handles generative AI work such as LLM fine-tuning data, evaluation and RLHF. Outlier’s site says it is powered by Scale AI and lists more than 900,000 verified experts in 52 countries and more than $600 million paid out to contributors. Scale’s homepage says 25% of its contributors hold advanced degrees. Scale has also tested an Expert Match feature that lets customers search, screen and invite contributors to projects. The practical upshot is breadth: Scale can staff very large programs and many task types, from bounding boxes to graduate-level reasoning prompts.

How Surge Sources Contributors

Surge positions its workforce as a smaller, more selective expert network. Its workforce page cites 200,000+ PhDs, 80+ languages and 500+ disciplines, and lists open roles with explicit credential bars, for example lawyers with Supreme Court clerkship or appellate experience and consultants at partner or senior manager level at top firms. One posted journalist role lists pay of $200 to $400 per hour. Business Insider reported that Surge worked with about 1 million annotators in 2025, so the network is large, but the sales pitch centers on depth of expertise rather than volume.

In 2023, The Verge and New York Magazine reported that Surge appeared to operate several separately branded work platforms, including DataAnnotation. Buyers rarely interact with these platforms directly, but it is worth knowing where your tasks may be routed.

Worker Classification Lawsuits

Both companies have faced contractor classification claims, which matter to buyers mainly as a supply continuity and reputational question. Scale was hit with several California suits from late 2024 alleging contributors were misclassified as independent contractors and underpaid, as TechCrunch reported; Scale said it complies with the law. A U.S. Department of Labor investigation into Scale was dropped in May 2025, according to TechCrunch. Court records in McKinney v. Scale AI show a $12.5 million class settlement that received preliminary approval in 2026, with Scale denying liability. Surge faced a proposed class action filed in May 2025 alleging annotators were misclassified, as the Los Angeles Times reported; Surge did not respond to the paper’s request for comment.

Services Compared

Post-Training Data

For language model work, the service menus overlap heavily. Scale’s Data Engine lists RLHF, generation of complex prompt and response pairs, red teaming with prompt injection and other techniques, evaluation against diverse prompts, and data curation and annotation. Surge’s RLHF page lists RLHF, human evaluation of model quality, alignment and safety data, InstructGPT-style fine-tuning data, tool use, summarization, STEM labeling and programming assistant data. Surge’s site also highlights post-training runs, off-the-shelf datasets, an enterprise offering and a dedicated software engineering track.

Data Types And Breadth

This is where the two separate most clearly. Scale also covers image, infrared and electro-optical imagery, full motion video and LiDAR or 3D sensor fusion, and its January 2026 update said it delivered more than 150,000 hours of physical AI data and onboarded 10 new robotics customers in 2025. Scale also builds AI applications, including the Donovan agent platform for government. Surge does not market computer vision, LiDAR or robotics labeling; its work is concentrated on text, reasoning, coding and agentic tasks. If you need multimodal or physical AI data, see our guide to top computer vision annotation companies.

Delivery Model

Scale sells primarily through managed programs scoped with its team. Surge offers native APIs and an SDK alongside a managed service with a dedicated data team, and allows individuals to sign up and launch a labeling project on their own.

Evaluation And Benchmarks Compared

Scale SEAL And Scale Labs

Scale launched its Safety, Evaluation and Alignment Lab (SEAL) in 2023 and published its first SEAL leaderboards in May 2024, using private, expert-built datasets so models cannot be tuned against the test set. SEAL produced benchmarks including Humanity’s Last Exam, MCP-Atlas and SWE-Bench Pro. Scale says SEAL introduced 15 benchmarks and published more than 450 evaluations across 50+ models in 2025. In March 2026, Scale folded this work into Scale Labs, a broader research unit covering evaluation, agentic systems, post-training methods and enterprise deployment. Its leaderboard page lists more than 20 benchmarks and more than 100 evaluated models.

Surge Benchmarks

Surge publishes its own benchmark suite, weighted toward professional work. Examples include DAYJOB: Tax, in which agents complete 50 tax assignments across 1,789 files, and sudo L7, a set of 60 coding tasks that test senior-level engineering judgment. Others include GDP.xlsx, DAYJOB: Healthcare, Riemann-bench and the Elo-style Hemingway-bench writing ranking. Surge also reports a Human Judgment Score that averages twelve dimensions, such as factuality, reasoning and harmlessness, each scored by expert annotators.

For buyers, both benchmark programs serve as public evidence of evaluation design skill. Scale’s are older and more widely cited in model release notes; Surge’s lean toward realistic, long-horizon professional tasks.

Security And Government Work Compared

Scale

Scale’s security page lists a SOC 2 Type II report, ISO/IEC 27001:2022 certification, a DISA Impact Level 4 provisional authorization and FedRAMP High authorization, with reports available through its trust center. Donovan is offered in GovCloud, classified and air-gapped environments. Scale reports nearly $200 million in U.S. Department of War contracts in 2025 and leads the Pentagon’s Thunderforge program with Anduril and Microsoft. It also has an expanding Middle East public sector business.

Two buyer considerations follow from Scale’s ownership. Some AI labs reportedly reduced work with Scale after the Meta deal; OpenAI denied that concerns about Meta were the reason. Separately, Business Insider reported in 2025 that Scale had used public Google Docs that exposed some client project material. Ask any vendor how project data is isolated between competing customers.

Surge

Surge lists SOC 2 compliance on its RLHF page and describes its service as “private, secure, and trusted by the largest AI enterprises.” We did not find published FedRAMP or ISO 27001 credentials. Inc. reported in July 2025 that a Surge internal guidelines document had been left publicly accessible. Surge’s independence from any single model developer is a selling point for labs that want to avoid a vendor with a competitor on its cap table.

Pricing Compared

Neither company publishes rate cards for enterprise post-training work. Scale’s pricing page routes buyers to a one-to-one intro call. Surge’s pricing page says enterprise projects are priced per task after an intro call, while individuals can sign up for free and launch a project themselves.

In practice, the main cost drivers for both are the same: contributor expertise (a credentialed lawyer or physician costs far more per hour than a generalist rater), task length and complexity, quality assurance layers such as multiple reviewers, turnaround time, and security requirements. Surge’s posted contractor rate of $200 to $400 per hour for one expert role gives a sense of what top-tier specialist labor costs before vendor margin. For budgeting context across the full model build, see our breakdown of AI development cost.

Which One Fits Your Team?

Choose Scale AI When

• You need one vendor across text, image, video, LiDAR or robotics data.
• Your program requires FedRAMP High, DISA IL4 or ISO 27001 credentials, or deployment in government environments.
• You want very large contributor capacity and a vendor that can also build and deploy applications.
• You value long-standing public benchmarks, such as SEAL’s, as part of your evaluation strategy.

Choose Surge AI When

• Your work is concentrated on RLHF, expert demonstrations, coding and agent evaluation for language models.
• You need narrow, high-credential experts, such as lawyers, physicians or senior engineers, and want the vendor to screen for them.
• You prefer a vendor with no major model developer among its shareholders.
• You want API or SDK access alongside a managed team.

When To Use Both Or Neither

Large labs often split work across several vendors to reduce dependency and compare quality. Running a paid pilot with both on the same task set is the most reliable way to decide. If your budget is smaller, your data is mainly computer vision, or you need a specialist in a regulated domain, other providers may fit better; our list of top Scale AI alternatives covers the wider market.

Frequently Asked Questions

Does Meta Own Scale AI?

No. Meta acquired a reported 49% stake in June 2025, described by The Information as non-voting. Scale says it remains an independent company, and it is now led by CEO Francis deSouza.

Is Surge AI Bigger Than Scale AI?

By one measure, yes. Reuters reported that Surge brought in over $1 billion in 2024 revenue, compared with about $870 million for Scale. Scale has more employees, more business lines and a large government business, so size depends on which metric you use.

Has Surge AI Raised Outside Funding?

Surge was reported in 2025 to be bootstrapped and profitable. Reuters and Bloomberg reported in July 2025 that it was in talks for its first outside raise, with Bloomberg citing a valuation of at least $25 billion.

Which Is Better For RLHF?

Both deliver RLHF at frontier lab scale. Surge is the more specialized choice for expert-heavy language work. Scale suits programs that combine RLHF with multimodal data, red teaming at scale or government security requirements.

Do Scale AI Or Surge AI Publish Prices?

No. Scale sells through intro calls and custom quotes. Surge prices enterprise projects per task after an intro call, and offers free self-serve sign-up for individuals.

Conclusion

Scale AI and Surge AI both sit at the top of the post-training data market, but they solve different problems. Scale offers breadth: multimodal and physical AI data, federal security credentials, an application business and a long-running benchmark program, alongside the ownership and leadership changes that followed Meta’s 2025 investment. Surge offers depth: a founder-led, reportedly bootstrapped company focused on expert human data for language, reasoning and coding, with no major model developer on its cap table.

Run a scoped pilot on your own tasks and ask both vendors about contributor vetting, data isolation and quality assurance before committing.