Happy Wednesday!

Three U.S. senators — Elizabeth Warren, Chris Van Hollen, and Richard Blumenthal — are probing whether AI data centers are driving up household electricity bills.

In letters to major tech firms, they warned that billions in grid upgrade costs tied to AI power demand may be getting passed on to ordinary consumers.

As AI power demand surges, Washington is now asking a blunt question: Who should pay for the infrastructure that keeps the servers running?

Let’s dive into today’s issue.

The U.S. has formally launched Pax Silica, a new alliance designed to secure the full AI and semiconductor supply chain — from critical minerals and energy to chips, data centers, and advanced manufacturing.

The initiative debuted at a closed-door summit convened by the State Department, bringing together countries that collectively anchor the world’s AI, chipmaking, mining, and infrastructure capacity.

Who’s in: Japan, South Korea, Singapore, the Netherlands, the UK, Israel, the UAE, and Australia — with participation from Taiwan, the EU, Canada, and the OECD.

Why it matters: This marks a shift from reactive export controls to proactive industrial coordination. Washington is now treating AI supply chains as strategic infrastructure — on par with defense and energy.

Texas’s power grid is facing an unprecedented surge in demand as artificial intelligence–driven data center projects flood the state. Grid operator ERCOT says large-load power requests have crossed 230 gigawatts this year, nearly quadruple last year’s levels. Most demand is coming from hyperscale AI data centers seeking gigawatt-scale connectivity.

What’s happening: The Electric Reliability Council of Texas (ERCOT) says large-load interconnection requests have jumped to 220+ gigawatts in 2025, nearly 4× last year’s level. More than 70% of the requests come from data center developers, many proposing gigawatt-scale sites comparable to large gas power plants.

What’s next: Texas regulators are moving to impose stricter rules. A proposal from the Public Utility Commission of Texas would require tighter coordination for any project seeking 75 MW or more, marking the state’s first formal effort to rein in the data center rush.

PJM Interconnection is weighing major changes to how data centers connect to the grid as surging electricity demand pushes prices higher and raises the risk of power shortages across the mid-Atlantic and Midwest.

The grid operator, which serves 13 states, is nearing a decision on new interconnection rules after data center growth has begun to outpace new power generation. The outcome could reshape data center development and electricity costs for millions of households.

Key Details: PJM Interconnection is considering sweeping changes to how data centers connect to the grid, including requiring large facilities to fund or bring their own power before receiving firm access. The proposals would also tighten demand-response and curtailment rules, allowing PJM to cut data-center load during periods of grid stress.

Why now? PJM Interconnection is taking action because data center demand is rising far faster than new power plants can be built. With generation projects taking 5–7 years to come online, PJM risks supply shortfalls, higher capacity prices, and potential reliability events as early as 2027.

PJM Interconnection is expected to decide on its new data-center interconnection framework by the end of December 2025. If approved, the revised rules would begin rolling out in 2026, with implementation phased to align with FERC review and PJM’s upcoming capacity auctions.

The House of Representatives is preparing for a key vote on the SPEED Act, a bill backed by major tech firms that would overhaul federal permitting rules for AI data centers, chip plants, and related energy infrastructure.

Why it matters: Companies like OpenAI, Meta, and Microsoft argue that slow, unpredictable approvals under the National Environmental Policy Act are now a competitive liability as the U.S. races China to scale AI infrastructure.

The details: The bill would tighten review timelines, limit lawsuits that delay permits, and shrink the window for legal challenges to 150 days. Supporters say this would unlock faster buildouts of data centers, transmission lines, and power plants needed for AI.

Foxconn plans to invest about $510 million (T$15.9 billion) to develop a new headquarters complex in Kaohsiung, reinforcing its long-term presence in southern Taiwan. Construction is expected to start in 2027, with completion targeted for 2033.

The project underscores Foxconn’s strategic shift away from low-margin consumer electronics toward higher-value AI servers, cloud infrastructure, and software — areas now driving the bulk of its growth.

By the numbers:

GE Vernova expects to end 2025 with an 80-gigawatt gas turbine backlog, extending deliveries into 2029, as demand from utilities and data center developers accelerates. CEO Scott Strazik said turbine reservations could be sold out through 2030 by the end of 2026.

GE Vernova said the current quarter is on track to be its largest ever for direct electrical equipment orders from hyperscaler technology companies.

Why this matters: GE Vernova, Siemens Energy, and Mitsubishi Power are the three dominant original equipment manufacturers in the global gas turbine market, collectively supplying roughly two-thirds of worldwide capacity.

All three are now reporting record order backlogs, with lead times stretching up to three years for new turbines—underscoring how tight the market has become as data centers, grids, and utilities race to secure firm power.

Nvidia is considering increasing production of its H200 artificial intelligence chips after demand from Chinese customers surged beyond the current supply, according to a Reuters report. The move follows the U.S. government's decision to allow exports of the H200 to China, subject to a 25% fee.

What’s driving the surge? Chinese technology companies, including Alibaba and ByteDance, have approached Nvidia seeking large H200 orders as soon as export approvals were signaled. The H200 offers a major performance leap over China-approved alternatives, making it critical for training and deploying advanced AI models. Domestic Chinese chips still lag the H200 by a wide margin.

Big picture: H200 production is currently limited, as Nvidia is prioritizing newer Blackwell and upcoming Rubin chips. Any output increase would depend on additional capacity at TSMC, where advanced chipmaking slots are already tight.