AI infrastructure works around the bottlenecks
Subject: Grid & Silicon: Routing AI around power limits
Preview: Five developments in fiber, rack power, cooling and grid capacity—and the tests that still separate plans from working systems.
This week’s coverage followed several ways AI infrastructure could work around scarce power and physical capacity. Companies proposed new fiber, direct rack power, cooling and workload-routing designs, while PJM revised a plan to buy generation capacity. Most are plans or tests. The next proof will come from deployment, regulatory approval and measured performance.
Tata plans a new fiber backbone for India’s AI hubs
Tata Communications plans to invest $400–450 million in about 10,000 kilometers of new long-haul fiber across India. The first phase would link Mumbai, Pune, Hyderabad and Chennai, with later phases extending to other data-center regions. Fiber can move training data and inference traffic between sites, but the announcement gives no construction schedule or completion date. The next useful evidence will be route delivery, customer commitments and operating performance.
Read the full briefing → · Primary source
Enphase tests a shorter path from the grid to AI racks
Enphase says a lab system in Austin converted 13.8-kilovolt alternating current directly to 800-volt direct current while responding to changing simulated rack loads. The 132-module setup is designed to replace several conversion stages between the utility feed and computing equipment. Enphase plans a 34.5-kilovolt test and megawatt-scale demonstration before the end of 2026, followed by customer pilots in 2027. Those larger tests and field trials will determine whether the lab result translates into a dependable data-center system.
Read the full briefing → · Primary source
Aligned claims more cooling from the same footprint
Aligned Data Centers launched Delta Cube 3.0, an air-cooling unit developed with Munters. Aligned says the design raises cooling capacity and air-cooling energy efficiency by 10% while keeping the previous footprint. It also says the closed-loop system uses no water for operational cooling. The release does not provide test conditions, measured energy use or a named customer deployment, so the figures remain company claims awaiting comparable field data.
Read the full briefing → · Primary source
Emerald and Lumen explore routing AI toward available power
Emerald AI and Lumen are developing a pilot that would route AI inference requests among locations according to power availability, electricity cost and network conditions. The idea treats computing demand as movable: when one site faces tighter power conditions, the network could send suitable requests elsewhere while accounting for latency and security. The companies have not named a pilot site, schedule, workload or measured result. Those details will determine how much demand can move without hurting service.
Read the full briefing → · Primary source
PJM revises its plan to procure capacity for large-load growth
PJM filed an amended one-time procurement plan with federal regulators after its 2028–2029 capacity auction fell about 6,800 megawatts short of the target. PJM proposes opening bids on Mar 1, 2027, and allocating costs to zones according to firm large-load growth between the 2026 summer peak and the 2028–2029 forecast. The target would be reduced for qualifying bilateral contracts. FERC has not approved the amendment, and PJM says the final procurement target will be recalculated with updated information.
Read the full briefing → · Primary source
Go deeper
The long road from a data-center plan to a grid connection
PJM’s proposal is one response to a broader timing problem. This explainer follows the studies, equipment and approvals that stand between a large new load and an operating grid connection.
The common thread is coordination: bits, power and heat all have to arrive at the right place at the right time. This week brought several proposed paths around those constraints; deployment data and regulatory decisions will show which ones become durable infrastructure.
Sources linked below.
Primary source · Primary source · Primary source · Supporting source · Primary source