JPMorgan Report: Anthropic Commits About 15 GW of Power, OpenAI Targets 30 GW by 2030
The author of the JPMorgan Private Bank’s Eye on the Market report, Michael Cembalest, published the report on September 10, 2026, stating that Anthropic has committed to about 15 GW of additional power grid interconnection, estimating cloud, chip, and data center lease costs at around $500 billion, up from $100–200 billion. OpenAI aims to reach 30 GW by 2030, with data center spending of about $750 billion. Most of the power would come from grid or behind-the-meter natural gas generation; on-site power is about 75 GW, almost entirely gas equipment, including transactions with Amazon AWS worth $10 billion / 5,000 MW. The report also notes that OpenAI has not released any proof for its model claims regarding the Navier–Stokes Millennium Prize problem; it is currently under mathematical review. There are rumors of progress on the Hodge conjecture, while Anthropic has made progress on the Birch and Swinnerton-Dyer conjecture. GPT-6 Astra scores 62% on the ARC-AGI-3 standard harness and 99% on the Provider Adapter. François Chollet points out that progress toward AGI is faster than expected.
The author of the JPMorgan Private Bank’s Eye on the Market report, Michael Cembalest, published the report on September 10, 2026, stating that Anthropic has committed to about 15 GW of additional power grid interconnection, estimating cloud, chip, and data center lease costs at around $500 billion, up from $100–200 billion. OpenAI aims to reach 30 GW by 2030, with data center spending of about $750 billion. Most of the power would come from grid or behind-the-meter natural gas generation; on-site power is about 75 GW, almost entirely gas equipment, including transactions with Amazon AWS worth $10 billion / 5,000 MW. The report also notes that OpenAI has not released any proof for its model claims regarding the Navier–Stokes Millennium Prize problem; it is currently under mathematical review. There are rumors of progress on the Hodge conjecture, while Anthropic has made progress on the Birch and Swinnerton-Dyer conjecture. GPT-6 Astra scores 62% on the ARC-AGI-3 standard harness and 99% on the Provider Adapter. François Chollet points out that progress toward AGI is faster than expected.