Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Infrastructure, Capstone Case

30 GW target by 2030

OpenAI has set an aspirational goal to reach 30 gigawatts of compute capacity by decade's end, split between research and products, to ensure researchers remain unconstrained in exploring new models.

Revenue tracks compute capacity

At frontier labs, revenue functions as a lagging indicator of compute capacity and utilization, with OpenAI tripling both compute and revenue year-over-year for the past three years.

Inference dominates future workloads

Over 80% of compute will shift to inference workloads, encompassing not just product usage but also post-training RL and synthetic data generation as scaling laws evolve across the entire AI lifecycle.

Three-dimensional cost optimization

OpenAI simultaneously pushes to make tokens cheaper, make each token more intelligent, and reduce the number of tokens required per task to maximize accessibility while managing infrastructure costs.

Gigawatt-scale orchestration

Building compute at scale requires synchronizing an entire supply chain including chips, memory, networking, power generation, cooling systems, data center construction, and land acquisition to align simultaneously.

Operational fragility post-deployment

The hardest work begins after contracts are signed, as modern GPUs are extremely brittle and sensitive to power fluctuations and cooling variations, requiring constant engineering to maintain performance at scale.

Grid stability risks

Large synchronized training jobs can cause hundreds of megawatt fluctuations on regional grids within minutes, creating blackout risks that require redesigned infrastructure to prevent collateral damage to national energy systems.

Energy diversification imperative

OpenAI is actively derisking supply chains by relocating manufacturing, decoupling from grid dependency through natural gas and nuclear energy investments, and pioneering infrastructure innovations with broader societal applications.

From chatbots to agents

AI workloads have evolved from simple one-shot chatbot inference to reasoning models, and now to agentic systems that autonomously execute actions, spin up VMs, and iterate through tasks rather than merely suggesting answers.

Complex compute graphs

Agentic AI requires dramatically more complex compute architectures that close the loop between thinking and doing, involving multiple tool integrations and iterative processing rather than single-response generation.

CPU resurgence

CPUs are making a comeback specifically for agentic workloads, contributing to Intel's recent momentum alongside supply constraints that benefit the only remaining leading-edge American manufacturer with fabrication capabilities.

7 terawatt vision

The long-term vision involves universal GPU access for 7 billion humans, requiring approximately 7 terawatts of compute—two orders of magnitude beyond current targets—suggesting infrastructure scaling has barely begun.