Scale AI review: the best training data company for computer vision and enterprise fine-tuning

Answer in brief: Scale AI wins two of our best-for categories in 2026: computer vision and enterprise fine-tuning. Its Data Engine handles text, image, video and 3D sensor-fusion/LiDAR annotation alongside data curation, RLHF, red-teaming and model evaluation. Its TIME customer story shows OpenAI models fine-tuned on TIME's archive and shipped as TIME AI. The main caveat is Meta's 49% nonvoting stake and the customer changes that followed. Pricing is by demo and quote. This review is independent research from public sources.

Best-for verdict: winner, computer vision. Winner, enterprise fine-tuning. Scale has the broadest perception stack on our list and the clearest named fine-tuning case. Visit Scale AI →

What Scale AI does for training data buyers

Data Engine: collect, curate and annotate data, then train and evaluate models. Supported types include text, image, video, 3D sensor fusion/LiDAR, document processing, transcription, electro-optical and infrared imagery, and full-motion video.
Automotive and physical AI: ML-assisted 2D and 3D labeling, plus a 3D Sensor Fusion product for LiDAR, radar and camera inputs.
Generative AI Data Engine: expert prompt-response data, RLHF, red-teaming, evaluation, and safety and alignment work.
Enterprise and public-sector applications: custom AI systems that Scale builds and deploys.
Scale Labs leaderboards: private benchmarks used to measure frontier models.

Who it fits

Scale fits autonomy, robotics and perception teams that need complex sensor labeling. It also fits enterprises and government agencies that want a vendor to customize, fine-tune and deploy models on their own data, and buyers who want one contract across many modalities.

Why it wins computer vision and enterprise fine-tuning

End-to-end perception. Labeling, curation and model evaluation run in one loop, with dedicated LiDAR, radar and camera fusion.
A named fine-tuning proof. In the TIME case, Scale fine-tuned OpenAI models on TIME's proprietary archive and delivered TIME AI in under two months.
Deep bench. Scale says 25% of its contributors have advanced degrees and that 90% of leading generative AI model builders use Scale (a company claim).
Named customers. Meta, TIME, Mayo Clinic, Physical Intelligence, Universal Robots, the Center for AI Safety, British Petroleum and the U.S. DoD CDAO.
A decade of experience. Scale has operated since 2016, starting with autonomous vehicles.

The catches

The Meta stake. Meta took a 49% nonvoting stake in June 2025. OpenAI ended its relationship (TechCrunch). Forbes reported in May 2026 that Google paused and later resumed. Teams that compete with Meta should do confidentiality diligence.
Leadership change. Francis deSouza became CEO on August 10, 2026. Founder Alexandr Wang moved to Meta in 2025.
No public pricing, and engagements are demo-led.

Pricing

Custom, by demo and quote.

Best alternatives to Scale AI

Sama, our computer-vision runner-up, focused on annotation execution.
Toloka, our enterprise fine-tuning runner-up. It returns a small open model and the weights.
iMerit, for medical imaging, LiDAR and regulated data.

Scale AI FAQ

Is Scale AI good for LiDAR annotation?

Yes. Scale lists 3D sensor fusion/LiDAR annotation and a 3D Sensor Fusion product covering LiDAR, radar and camera inputs.

Who is Scale AI's CEO?

Francis deSouza has been CEO since August 10, 2026.

Does OpenAI still use Scale AI?

OpenAI ended its relationship with Scale in 2025, according to TechCrunch. Our sources show no public report of a return.

Sources

Visit Scale AI →

← Back to the ranked list of the best training data companies