WEEKLY DIGEST · November 14–20, 2026

Five constraints shape AI infrastructure

Subject: Grid & Silicon: Five constraints shaping AI infrastructure

Preview: This week: GPU memory choices, faster grid studies, integrated cooling, a Loudoun approval pause and Google’s water commitments.

This week’s coverage shows why AI infrastructure cannot be measured by processor counts alone. AMD added a memory-capacity choice to its next accelerator family; PJM proposed a faster study path for generation and large loads; Ecolab announced partnerships spanning cooling, power and design models; Loudoun County paused some final votes while it rewrites data-center rules; and Google set out new water commitments. Each development addresses a different constraint, and most still depend on future filings, designs or measured results.

AMD adds a memory-capacity choice to MI400 GPUs

AMD introduced the MI450X with 288 GB of HBM4 memory while positioning the MI455X with 432 GB for workloads needing more local memory. The company says both retain the family’s core compute, memory-bandwidth and connectivity capabilities. The distinction is about workload fit, not a universal speed ranking: capacity determines what can remain close to the processor, while actual performance still depends on software, data movement and the full system. AMD’s reported inference results are company tests under specified conditions, not results for every model or deployment.

Source date: Nov 12, 2026

Read the full briefing → · Primary source

PJM proposes a one-year grid-study process

PJM’s proposed Fast Energization process would study new generation and transmission service for large electricity users together, reducing its projected study timeline from two years to one. PJM also proposes earlier readiness and financial requirements and clearer visibility into transmission costs driven by large loads. The filing still requires Federal Energy Regulatory Commission review. A faster study would give developers earlier information, but it would not build the generation, substations or transmission upgrades identified by that study.

Source date: Nov 16, 2026

Read the full briefing → · Primary source

Ecolab links power, cooling equipment and design models

Ecolab announced collaborations with Siemens, Alfa Laval and Dassault Systèmes around high-density AI infrastructure. The work connects electrical distribution and automation, heat exchangers and coolant-distribution equipment, and models of the technology-cooling and facility-water loops. The announcement describes partnerships and a simulation showcase rather than measured performance at a named operating facility. The useful next evidence will be comparable tests with defined workloads, local water conditions and clear energy and redundancy boundaries.

Source date: Nov 17, 2026

Read the full briefing → · Primary source

Loudoun pauses some final data-center votes

Loudoun County supervisors paused Board action on certain legislative data-center and substation applications while the county completes new standards and location policies. Staff review, new submissions and Planning Commission hearings can continue, so the action delays final Board decisions rather than freezing every application. The county says 33 active legislative applications could be affected, with exceptions for the Route 28 Tax District, administrative applications, grid reliability and approaching statutory deadlines.

Source date: Nov 18, 2026

Read the full briefing → · Primary source

Google sets out water safeguards and funding commitments

Google outlined five commitments covering replenishment, public infrastructure, cooling choices, disclosure and alternatives to freshwater. It says it has announced more than $100 million in community initiatives across 14 states this year and committed more than $500 million cumulatively to water, wastewater and reuse infrastructure. Those figures describe commitments, not completed projects. Google’s 2030 replenishment goal and more than 700 project submissions under review are also forward-looking; local designs and annual reporting will show what is implemented at individual sites.

Source date: Nov 19, 2026

Read the full briefing → · Primary source

Go deeper

Why AI accelerators need high-bandwidth memory

This new explainer separates memory capacity from bandwidth, follows HBM’s stacked construction and explains why faster arithmetic cannot rescue a workload that is waiting for data.

Living Next to a Data Center: Noise, Water and Your Electric Bill

This Grid & Silicon video connects project-level water, noise and electricity questions to the community decisions highlighted by the Loudoun and Google stories.

Explore the U.S. AI data center map

Use the map to explore publicly documented U.S. sites, their development status and the research behind each record. The inventory is not exhaustive.

The common thread is systems thinking. More memory, a shorter study, a coordinated cooling design, a paused vote or a water commitment can remove one obstacle without finishing the entire project. The next useful evidence is specific: measured accelerator performance, an approved grid process, operating cooling data, adopted county rules and completed local water projects.

Sources linked below.

Primary source · Primary source · Primary source · Primary source · Primary source · Living Next to a Data Center: Noise, Water and Your Electric Bill