Magic is building frontier-scale code models on massive GPU clusters, operating the full stack from kernel optimization to trillion-parameter training. The tech stack — CUDA, XLA, Triton, Ray, Kubernetes across GCP/AWS/Azure/OCI — reflects deep infrastructure work; the project list (sandboxed execution, high-density compute, internet-scale data pipelines, inference optimization) reveals a company building toward inference at scale, not just training. Engineering dominance and heavy investment in storage, caching, and fault tolerance signal they are solving hard systems problems upstream of the model itself.
Tech Stack 28 technologies
Core StackAWS Terraform Pulumi CloudFormation C++ Go Rust GitHub Kubernetes CUDA XLA GPU TPU GCP Azure OCI AWS CDK C NVMe NFS cgroups SSD Ray Ruff LinkedIn NCCL Triton Mojo
What Magic Is Building
◆Challenges
- Maintaining diverse datasets at scale
- Navigating unknown product development
- Optimizing inference latency
- Ai training and inference process protection
- Emerging threat detection
- Red-teaming vulnerability reduction
- Highly available training
- Generating synthetic datasets reliably
- Fault detection and recovery
- Data infrastructure ergonomics
▲Active Projects
- Sandboxed execution environments
- Self-service compute platform
- High-performance storage and caching systems to support long-context inference and training
- High density compute provisioning
- Evaluate porting compute kernels to alternative hardware
- Massive-scale gpu supercomputing infrastructure
- Build out internet-scale data pipelines and crawlers
- Train trillion-parameter models on large gpu clusters
- Optimize inference throughput for novel model architectures
- Post-training dataset acquisition
Hiring Activity
25 roles · 5 in 30d
Department
Seniority
Company intelligence
