We’re launching Lyria 3.5 in Google Flow Music, with advances across musicality, lyrics, vocals, and creative control
deepmind.google·6d ago
AI Engineering
OpenAI Blog·Apr 29·5 min read
Building the compute infrastructure for the Intelligence Age
TL;DR
AI model training and inference require massive computational resources that existing data centers cannot fully supply. OpenAI scaled Stargate, a large-scale compute infrastructure project, by expanding data center capacity to handle growing AI workloads.
✦ Why It Matters
Engineers can understand that compute infrastructure is now a critical bottleneck limiting AI capability advancement and deployment scale.
Key Takeaways
- Compute capacity directly constrains the size and capability of trainable AI models.
- Data center infrastructure requires planning for cooling, power, and networking at massive scale.
- Distributed data centers enable both training and low-latency inference serving globally.
- Infrastructure investment timelines must align with model development roadmaps to avoid bottlenecks.
Full Summary
Training advanced AI models demands enormous computational power measured in petaFLOPS (quadrillions of floating-point operations per second). OpenAI identified a critical bottleneck: insufficient data center capacity to support both current model training and inference serving at required scales.
The company scaled Stargate, an infrastructure initiative designed to provision specialized hardware (GPUs and TPUs—graphics and tensor processing units) across geographically distributed data centers. This expansion involved adding new physical data center facilities with optimized cooling, power distribution, and networking to support dense AI compute clusters.
The scaling effort directly addresses the computational demands of training frontier models and serving them to millions of users simultaneously. By increasing available compute capacity, OpenAI can reduce training time, support larger model architectures, and maintain service reliability during peak usage periods.