Hi, and welcome to the first Grid & Silicon email. Each Friday I'll pull together the stories about power, chips and campuses that are actually moving AI infrastructure, and I'll say plainly what's still only a plan. Thanks for being here early.

This week's top 3

1. A Senate bill would make data centers pay their transmission costs

A bipartisan Senate permitting package, the Bipartisan American Affordability and Jobs Act of 2026, includes a provision requiring data centers to pay all the transmission costs tied to them. The stated goal is to keep households and small businesses from covering grid upgrades built for large new loads. The bill would also strengthen FERC's backstop authority over major interstate power lines and set deadlines for environmental reviews.

Why it matters: The idea of "who pays" is now in writing, but the details aren't. The announcement doesn't explain how costs would be split or how existing projects would be treated. It's still proposed legislation, and floor debate is expected after the election. Source: U.S. Senate announcement, Oct 2. → Read the brief

2. CoreWeave plans a 240 MW AI cloud in Navi Mumbai

CoreWeave says it will enter India through a planned 240 MW deployment at AdaniConneX's Taloja campus, spread across three 80 MW buildings. It expects to be the only tenant and to run Nvidia's Vera Rubin platform, with the first phase due in mid-2028. The halls are designed for direct liquid cooling, with a non-evaporative chilled-water system that removes heat without depending on constant evaporation.

Why it matters: The plan names a tenant, a chip platform and a cooling design, which is more detail than most campus announcements give. Still, none of the 240 MW is operating, and CoreWeave hasn't disclosed the investment, the power source, the utility connection or any customers. Source: CoreWeave company announcement, Oct 7. → Read the brief

3. Pennsylvania proposes cutting power to big compute loads first

The Pennsylvania Public Utility Commission voted 4–0 to seek comment on load-control rules for "large computational sites." Those are sites expected to draw at least 25 MW at one connection point, or 50 MW across several tied to one feeder. In a qualifying pre-emergency event, eligible sites that haven't secured capacity for their full demand would be curtailed before other customers lose power.

Why it matters: This puts data centers in a defined place in the line when supply is tight. A data center that wants to stay off that list would need to bring its own capacity. Source: Pennsylvania Public Utility Commission, Oct 1. → Read the brief

Quick hits

• Texas: ThisWay Global bought a stalled project in Bryan, Texas, and plans a 73 MW AI and research campus at Texas A&M-RELLIS. It hasn't said when the first phase would open. Brief →
• Malaysia: Bitdeer signed a 10-year services agreement for A901, a planned 67 MW liquid-cooled campus. Delivery is targeted for early 2027, and no offtake customer has been named. Brief →
• Nuclear: Google and Constellation signed a 20-year deal to fund upgrades that would add 890 MW across 11 existing nuclear units, with the first uprate expected in 2028. Brief →
• Finland: A regulator told Tuike Finland to stop major ground work at planned Google-linked sites in Muhos and Kajaani by Oct 23 while environmental reviews are finished. Brief →
• Hardware: QuantumScape showed QS PowerBlock, a solid-state battery reference design for future 1 MW AI racks. Its numbers come from prototypes and models, not from a working data center. Brief →

On the map

The map's last data review was Oct 5. It still tracks 77 U.S. campuses: 59 operating and 18 in development. No campuses were added or removed, but four records changed:

• Meta Gallatin (Tennessee): reclassified as under construction with no current AI capacity. That doesn't mean the broader campus shut down.
• Meta Temple (Texas): estimated IT power raised from 152 to 178 MW after Building 3 was treated as operational.
• Meta Prometheus (Ohio): estimated IT power revised down from 496 to 471 MW. It's still operating.
• QTS Eagle Mountain (Utah): Meta is now labeled as a possible hardware owner and user. Both labels are speculative.

These are research estimates from Epoch AI's dataset, not live power readings. Explore the map →

Explainer of the week

Two of this week's top stories come down to one question: how does electricity actually get to a data center? A power plant can generate it, but transmission lines, substations, transformers and firm capacity all have to deliver it to the site. That's where most of the waiting, and most of the cost, ends up.

→ Read: How data centers connect to the power grid (10-minute read) → Watch: How Does the Power Grid Feed an AI Data Center?

If you'd rather watch than read, the full video library is on YouTube. Subscribe here.

See you next Friday, Ryan

Got a tip? Reply to this email or write [email protected]