Scale AI vs Surge AI: the comparison turns on ownership
For most of the last decade Scale AI was the default answer for frontier training data, and Surge AI was the quieter alternative that labs used for preference work. In June 2025 that arrangement changed, when Meta took a 49 percent stake in Scale and several of Scale's largest customers reconsidered. Surge was widely reported as the beneficiary. This comparison works through what each company actually sells, what each publishes about its own operations, and where each one is still the better answer, because the reporting on this rivalry has outrun the facts in both directions.
Key takeaways
- The structural difference is ownership. Meta holds 49 percent of Scale. Surge has taken no outside capital and its founder holds roughly 75 percent, which makes its independence verifiable rather than asserted.
- Scale publishes the strongest compliance posture in the category, including FedRAMP High and DoD IL4. Surge publishes substantially less, so expect to run your own security assessment.
- Surge is text, code, and multimodal digital work. It publishes nothing about real-world or physical collection, so for sensor and robotics capture neither of these is the right vendor.
- Both publish substantial open research. Scale has the larger dataset footprint; Surge's benchmarks are cited by OpenAI, Anthropic, and Microsoft in dated model releases.
- Neither has a usable third-party review base. Scale has two Gartner reviews, Surge has one on G2.
- Scale is the only one of the two with government authorizations, and for defense work that settles it.
About Scale AI
Founded around 2016 by Alexandr Wang and Lucy Guo, based in San Francisco. In June 2025, Meta invested roughly 14.3 billion dollars for a 49 percent stake at a valuation above 29 billion. Wang left to lead Meta's superintelligence effort while staying on Scale's board, and Francis deSouza became chief executive in August 2026.
Scale sells a Data Engine for annotation and curation across text, image, video, and 3D and LiDAR; a generative AI portfolio covering RLHF, prompt-response generation, and red-teaming; public model leaderboards; and Donovan, a decision-support platform for defense and enterprise. Its positioning has shifted markedly toward government work, which is also the most verifiable part of its record because federal contracting data is public: Army contract W911QX20C0051 shows 105.5 million dollars total obligated, an Army indefinite-delivery contract signed in May 2026 carries a ceiling near 99.5 million, and a February 2026 Air Force award of roughly 5.1 million integrates Donovan into an airborne command platform. The much-cited 500 million dollar Department of Defense agreement is a ceiling, not obligated spend.
About Surge AI
Legally Surge Labs, Inc., founded in 2020 by Edwin Chen, based in San Francisco. The notable fact is financial: Surge has raised no outside capital and has been profitable from the start. Bloomberg reported 1.2 billion dollars in 2024 revenue, and Forbes reports Chen holding roughly 75 percent of the company. Reuters reported in July 2025 that Surge was seeking up to a billion dollars at more than 15 billion, and Bloomberg reported talks at 25 billion or above later that month. No round has closed, so Surge has no priced valuation and the figures in circulation are estimates.
Surge sells nine product lines organized around post-training: RL environments and agents, rubrics and verifiers, RLHF, supervised fine-tuning, human evaluation, expert professional domains, internationalization, multimodal work covering image, audio, and video, and off-the-shelf data. It describes its workforce as a global community including a claimed 200,000-plus PhDs across 500-plus disciplines. Named customers on its own site are OpenAI, Anthropic, Meta, and Google, and its blog documents dated engagements with Anthropic and Microsoft.
What Surge does not publish anywhere is physical or real-world data collection. Multimodal in its usage means digital image, audio, and video work performed by remote contractors, not sensor capture, teleoperation, or on-site collection.
What buyers report
Neither company has a usable review footprint, which for two businesses of this scale is itself worth noting.
Scale's GenAI Platform holds 4.5 out of 5 on Gartner Peer Insights from two reviews, citing model quality and tooling, with pricing structure flagged as a concern. Surge has a single G2 review, no Gartner Peer Insights listing, and no Capterra presence found as of October 2026. G2's own profile text says there are not enough reviews to provide buying insight.
Both sell top-down to a small number of large customers rather than through software-buyer channels, which explains the absence. The practical consequence is the same for both: run a paid pilot, because no third-party source will tell you what you need to know.
The comparison
| Scale AI | Surge AI | |
| Ownership | Meta holds 49 percent since June 2025 | No outside capital; founder holds roughly 75 percent |
| Model | Managed data engine plus government software | Expert workforce delivering finished post-training data |
| Modalities | Text, image, video, 3D and LiDAR, multimodal | Text, code, image, audio, video. No physical collection published |
| Compliance | SOC 2 Type II, ISO 27001, DoD IL4, FedRAMP High. No HIPAA found | Substantially less published |
| Public terms | No IP indemnification published | No IP indemnification published |
| Government | Verifiable federal contract history | Reported US Army and Air Force work; no authorizations published |
| Open research | 36 datasets and 5 models on Hugging Face | 13 benchmarks, 8 arXiv papers, 3 Hugging Face datasets, all shipped in 2026 |
| Workforce disclosure | No headcount, geography, or pay rates published | Some rates published; no geographies published |
| Third-party reviews | 4.5 out of 5, two reviews | 5.0, one review |
Independence, which is the live question
Scale's ownership is the fact every lab evaluates first. A frontier lab sending unreleased-model data to Scale is sending it through a pipeline part-owned by a competitor. The market reacted immediately: within weeks of the June 2025 investment, Google, OpenAI, and xAI were reported to have pulled back or halted projects, and Scale laid off roughly 200 data-labeling staff. By August 2025, reporting indicated that Meta's own AI unit was favoring Surge and Mercor over Scale, with internal sources describing Scale's data as low quality, a characterization Meta disputed.
Surge's independence is the mirror image and it is unusually well evidenced. A company that has taken no outside capital has no investor whose strategic interests could conflict with a customer's, and Forbes reporting on Chen's ownership stake makes the structure checkable rather than a claim in a sales deck. For a lab whose primary concern is that no competitor sees its data, this is the single strongest argument in Surge's favor, and it is structural rather than contractual.
The caveat is that independence is not the same as governance. Which brings us to the sharpest contrast between these two.
Published research, where both are genuinely strong
This is the dimension where both companies outperform the rest of the category and where the comparison is close.
Scale publishes 36 datasets and 5 models on Hugging Face, the largest open footprint of any services vendor, including MCP-Atlas, EnigmaEval, PRBench, and MultiChallenge, and it runs public model leaderboards.
Surge shipped 13 benchmarks across 2026 with 8 accompanying arXiv papers, including Riemann-bench, GDP.pdf, ComplexConstraints, Chartography, and the DAYJOB healthcare and finance sets. The stronger signal is uptake: OpenAI cited GDP.pdf in its GPT-5.6 release, and Anthropic cited GDP.pdf and Riemann-bench in a system card. Benchmarks that the labs themselves cite in model documentation are a harder credential than download counts.
For a buyer, published benchmarks are useful evidence because you can inspect the task design and judge whether the evaluation rigor on display resembles what you would get on your own project.
Which to choose
Choose Scale when you are deploying into defense or government, where FedRAMP High and IL4 decide it; when you need 3D, LiDAR, or sensor annotation at throughput; or when you want published certifications you can hand to your own security review before a sales conversation starts. Weigh the Meta stake according to whether Meta competes with you.
Choose Surge when independence is your primary concern and you want it structurally rather than contractually guaranteed; when the work is text, code, reasoning, or digital multimodal evaluation; and when benchmark-grade evaluation design is what you are actually buying. Go in expecting to negotiate security and IP terms from scratch, and expect to ask for the current enterprise agreement rather than reading the public one.
Choose neither if what you need is data that does not exist yet in physical form. Scale annotates sensor data you have collected; Surge does not publish any physical collection capability at all.
Why HumanSignal is the better choice when you need the data created
This comparison turns on ownership, and that is the right axis, but it is worth noticing what both sides of it have in common. Scale and Surge are both organized around processing or judging material that already exists. Independence determines who you can safely hand your data to. It does not help when the data has never been recorded.
HumanSignal Services is organized around the other problem. We design the protocol, recruit against it, and run instrumented capture in our studios, on site, or in the field, covering egocentric and multi-camera video, LiDAR and sensor streams, multi-channel audio, and biometric collection, then verify the result against the spec before delivery. That is an operations capability rather than a workforce, which is why it does not substitute for either company here and why neither substitutes for it.
On the independence question this piece is built around, our answer is contractual and checkable. Our customer agreement states that the customer exclusively owns all right, title, and interest in customer data and in insights produced specifically for them, which covers labeled output. We publish contributor pay ranges rather than describing our sourcing in adjectives, and we hold SOC 2 Type 2 and HIPAA compliance, documented on our security page. Quality workflows run in Label Studio Enterprise, so agreement metrics and reviewer decisions stay visible to you throughout.
We built this through the acquisition of Erud AI, a physical data collection company, in November 2025, which is what turned a labeling and evaluation company into one that also creates the underlying data. If your bottleneck is capture rather than judgment, start at physical AI data services.
Did Surge AI actually gain business from Meta's investment in Scale
Reporting supports it directionally. TechCrunch reported in August 2025 that Meta's own AI unit was favoring Surge and Mercor over Scale, and Google, OpenAI, and xAI were reported to have pulled back from Scale in June 2025. Surge has published no revenue figure since 2024, so the size of any shift is not publicly quantified.
Is Surge AI really bootstrapped
Yes, and it is one of the better-evidenced facts about the company. Forbes reports Surge has not raised outside funding and that Chen holds roughly 75 percent. Surge sought outside capital in 2025 at valuations reported between 15 and 25 billion dollars, but no round has closed, so there is no priced valuation.
Can either company collect physical or robotics data
Scale annotates 3D and LiDAR data and lists collection among its services. Surge publishes no physical collection capability. If you need teleoperation, egocentric capture, or on-site instrumented recording, both are the wrong starting point and you should look at collection specialists instead.