Conference Presentation, Interview
Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Infrastructure, Capstone Case
- Intel is expected to benefit from supply constraints and a CPU resurgence driven by AI agents, with stock performance anticipated to improve within approximately two weeks and high confidence placed in leadership execution despite remaining operational challenges.
- Frontier lab revenue is predicted to remain a lagging indicator tightly correlated with compute capacity, tripling alongside it, while OpenAI forecasts meaningful double-digit growth for Codex roughly two weeks after version 5.5 releases.
- Future compute allocation is projected to be 80% or more dedicated to inference, including research and synthetic data generation, with OpenAI aiming to make every token cheaper through hardware, software, and model intelligence improvements.
- OpenAI intends to utilize a 30-gigawatt compute target for both research and products with no idle capacity, while general hyperscaler compute planning is expected to reach 100 gigawatts, consuming double-digit percentages of U.S. grid capacity.
- Gigawatt-scale data centers pose significant energy risks, including grid failures or state-wide blackouts due to fluctuations of hundreds of megawatts, though the speaker envisions a long-term future requiring 700 terawatts of compute if every human possesses a GPU.
- Agentic workloads are expected to evolve into complex compute graphs involving iterative task attempts and VM spinning, with economic advantages favoring concentrated gigawatt-scale clusters over distributed 50-megawatt edge clusters due to labor and latency constraints.
- Latency optimization is critical, as shaving 30 to 50 milliseconds from current 400 to 500 millisecond generation times (dominated by the "pre-fill" phase) will drive higher engagement and revenue, while human users may become the primary bottleneck in AI workflows.
- Infrastructure will require heterogeneous computing involving CPUs, GPUs, and accelerators like Cerebras, with software orchestration identified as the primary short-to-medium term innovation driver and new memory architectures becoming critical in the medium to long term.
- TSMC's wafer allocation strategy is expected to force diverse chip utilization, while AI models are predicted to design future chips and software during training, potentially reducing the standard three-year design cycle; however, fab capacity across TSMC, Samsung, Intel, Micron, and SK Hynix remains a single structural choke point.
- Value flow in the AI stack is expected to shift from infrastructure to platform and application layers similar to the mobile revolution, with foundational component builders (transformers, batteries, cooling) accruing long-term value while traditional "model wrappers" and apps are predicted to be replaced by outcome-delivering systems.
- OpenAI plans to continue investing in frontier model intelligence as open-weight models are expected to take six months to catch up, with the speaker predicting OpenAI could enable unlimited usage through daily or weekly token limit resets and Nvidia becoming the first company to reach a 10 trillion market cap.
- Investment decisions are shifting toward "time-to-compute" as the primary metric, favoring larger concentrated chunks of compute for operational speed, with the speaker predicting Nvidia will be the first to reach a $10 trillion market cap.