Scale AI sells data for training and testing AI models, plus an agent platform for enterprises and governments. Francis deSouza took over as CEO in August. It bought ICG Solutions in April 2026 and launched Scale Post-Training with NVIDIA in October. Pricing is by contract.
Scale AI is a San Francisco company that supplies training data, model evaluations and enterprise AI agent tooling. Meta holds a minority stake, and Francis deSouza is CEO. Scale lists FedRAMP High authorization, and its About page counts more than 1,000 employees.
Founded: 2016 · HQ: San Francisco, CA, USA · Team: 1000-5000 · CEO: Francis deSouza (from 10 August 2026) · Funding: Meta: $14.3B for a 49% minority stake (June 2025). Scale publishes no total-funding figure. · Valuation: Above $29B (Meta transaction, June 2025; shown on Scale's About page)
About Scale AI
Scale AI was founded in San Francisco in 2016 by Alexandr Wang and Lucy Guo, who met at Quora and joined Y Combinator's winter 2016 batch while still students. It began by selling labeled data for self-driving perception and grew into a data and evaluation supplier to frontier AI labs. Guo left in 2018. Wang ran the company until June 2025, when Meta took a 49% minority stake for $14.3 billion at a valuation above $29 billion and Wang moved to Meta's AI effort (Meta AI). Jason Droege, previously chief strategy officer, was interim CEO from June 2025. On 30 July 2026 Scale announced that Francis deSouza would become CEO effective 10 August 2026. deSouza was most recently chief operating officer and president of security products at Google Cloud, and before that CEO of Illumina. Wang is chairman of the board and Droege is helping with the transition.
Scale's own site frames the business in three parts: data for training, evaluations and red-teaming, and applied AI for enterprises and governments. The Data Engine is the long-standing data product, covering collection, annotation, RLHF (reinforcement learning from human feedback) and model evaluation across text, image, video, 3D and sensor data. The Scale GenAI Portfolio (SGP) is the enterprise agent platform: it connects company data, hosts and orchestrates agents, adds evaluation and monitoring, and feeds human feedback back into the model. It is designed to run inside a customer's own cloud on AWS, Azure or GCP and to work across models from OpenAI, Google, Meta, Mistral and others.
The newest addition is Scale Post-Training, launched on 15 October 2026 alongside NVIDIA's release of Nemotron 3.5 Super VL. The offer is to measure the model a customer runs today on tasks from their own workflow, build training and evaluation data for it, and post-train an open-weight model to match. Scale's own figures from launch customers, which are vendor-reported and not independently audited, include accuracy matching the frontier model Skydance runs in production at about 40 times lower inference cost, and matching the model EY uses for M&A due diligence at about one-eighteenth of the cost.
On the government side, Scale Donovan is the product line for US national security customers. Scale's security page lists FedRAMP High authorization, a DoD IL4 provisional authorization, SOC 2 Type II and ISO 27001:2022. On 20 April 2026 Scale acquired ICG Solutions, a defense analytics firm founded in 2010 whose LUX product ingests, correlates and alerts on large streaming data sets, and said it would fold LUX into its agentic platform. The July 2026 CEO announcement also named BP and Mayo Clinic among recent enterprise customers. Scale's About page puts its work at 15 billion human decisions used to train AI models, more than $1 billion paid to contributors worldwide, and more than 1,000 employees. Scale is privately held.
Scale is not the right fit for everyone, and the other side of the case is worth stating plainly. Pricing is by contract, with no public rate card, so a small team that wants to label a few thousand images this month will find a self-serve tool or a smaller specialist faster and cheaper. Meta's 49% stake is a fair reason for a lab that competes with Meta to prefer a vendor with no such tie, and Scale's own pivot toward enterprise applications and government work means data labeling is no longer the whole story. Buyers comparing options should look at Surge AI for expert human data, Mercor for expert contractor marketplaces, and Encord for annotation software that teams operate themselves. Frontier labs such as OpenAI, Anthropic, Microsoft and Cohere each source training data in their own mix of in-house and outside vendors, so Scale's customer list is only one input to a vendor decision.
Mission
Develop reliable AI systems for the world's most important decisions.
Products
- Scale Data Engine (AI training data platform): https://scale.com/data-engine
- Scale GenAI Portfolio (Enterprise agent platform): https://scale.com/genai-portfolio
- Scale Post-Training (Open-weight model customization service): https://scale.com/genai-portfolio
- Scale Donovan (Government and defense AI platform): https://scale.com/donovan
Compliance
SOC 2 Type II, ISO 27001:2022, DoD IL4 Provisional Authorization, FedRAMP High
Links
Frequently Asked Questions
What does Scale AI do?
Scale AI supplies the data and testing that AI models are built on, and increasingly the software enterprises use to run agents. The Data Engine covers collection, annotation, RLHF and evaluation, the GenAI Portfolio hosts and monitors agents in a customer cloud, and Scale Post-Training customizes open-weight models such as Nemotron 3.5 Super VL. Scale Donovan serves US national security customers. Pricing is negotiated per contract, with no rate card on its site.
Who is the CEO of Scale AI?
Francis deSouza became CEO of Scale AI on 10 August 2026, as announced by the board on 30 July 2026. He previously ran Illumina as its CEO and held a senior security role at Google Cloud. Jason Droege had been interim CEO since June 2025 and is helping with the transition. Founder Alexandr Wang, who left for Meta in June 2025, is chairman of the board according to Scale's announcement.
Who owns Scale AI and how is it funded?
Scale AI is privately held. In June 2025 Meta agreed to pay $14.3 billion for a 49% minority stake, valuing the company above $29 billion, a figure Scale still shows on its About page. Scale does not publish a total-funding figure on its site, so Hokai lists only the Meta transaction rather than an aggregate. There is no public IPO filing.
Is Scale AI secure enough for government and regulated work?
It meets the usual US federal and large-enterprise checklist: the security page shows FedRAMP High, a DoD IL4 provisional authorization, a SOC 2 Type II report and an information-security certificate, with documents in its Trust Center. Hokai did not find HIPAA or GDPR attestations on the page, so healthcare and EU buyers should request those documents directly. Scale named Mayo Clinic as a customer in July 2026.
What are the alternatives to Scale AI?
The closest options listed on Hokai are Surge AI, Mercor and Encord, which cover expert data work, contractor sourcing and self-operated annotation software respectively. They differ from Scale mainly in service model: Scale sells managed programs and a government and enterprise stack on contract, while self-serve tools suit small or one-off labeling jobs. Companies such as OpenAI and Anthropic mix outside vendors with in-house data work, so a vendor choice should follow your data type and volume.
Hokai guides that mention Scale AI
- Best AI Companies in 2026: Who Is Actually Leading: Anthropic's revenue passed OpenAI's in 2026, even as OpenAI keeps the bigger valuation. See how six leading AI companies actually compare, by the numbers.