Large language models depend on material conditions, not computation alone: infrastructure supplies resources that shape which systems can be built, trained, and operated at scale. section 4.6
For frontier-scale systems, infrastructure includes:
- Concentrated computing capacity, specialized GPUs or TPUs, high-bandwidth networking, and reliable energy. section 4.7.2
- Access to these resources sets practical limits on model size and training duration, and affects how much uncertainty or failure organizations can absorb. section 4.7.2
- Hardware availability, infrastructure choices, model development, and deployment are coupled; cloud partnerships can bundle compute access with deployment. section 8.1.3
Access and integration
Microsoft–OpenAI and Amazon–Anthropic partnerships connect model development to cloud platforms. Hardware-specific software stacks can improve efficiency while making systems harder to move; Anthropic uses some TPUs through Google Cloud as an alternative source of compute. section 8.1.3
Infrastructure also describes a role technology can acquire: when AI assistants become embedded in everyday workflows, they can become difficult to replace without disrupting work and coordination. section 8.1.5










