Magic trains frontier-scale code models aimed at an autonomous coding coworker rather than an assistant. Its stack spans four clouds (GCP, AWS, Azure, OCI) plus TPU, CUDA, XLA, NCCL and Triton, pointing to a hardware-agnostic training operation. Pain points cluster around long-context memory, KV-cache scaling and cross-device communication, and a senior-heavy engineering hiring mix backs custom GPU kernel and cluster work.
Magic is a privately held AI company headquartered in San Francisco, founded in 2022, with 51–200 employees. It develops large code models intended to work as a software engineering coworker. Its engineering runs on Python, C++, Rust, Go, Mojo, Kotlin and TypeScript, with infrastructure managed through Terraform, Pulumi, AWS CDK, CloudFormation and Kubernetes. Distributed workloads use Ray across GPU and TPU hardware. Hiring is concentrated in the United States, mainly in engineering with smaller ops and security functions.
Magic uses Python, C++, Rust, Go and Mojo, with CUDA, Triton, XLA, NCCL and TPUs for model work, Ray for distributed compute, and Kubernetes, Terraform and Pulumi across GCP, AWS, Azure and OCI.
Magic is building custom high-performance GPU kernels, large-scale GPU clusters for training and inference, synthetic and post-training datasets, and evaluating porting its kernels to alternative hardware.
Yes, hiring leans toward senior engineering roles in the United States, with additional openings in ops and security; overall hiring velocity is currently minimal.
Other companies in the same industry, closest in size
Magic's technology stack, projects, and hiring signals are inferred from public hiring and company data — career pages, public listings, and company web presence — then clustered and de-duplicated. Figures are estimates that refresh over time. Read our full methodology →
This is not an official vendor or customer list. It is a technology-adoption signal inferred from public data, intended for B2B research.
