Cooling & Optics: New Bottlenecks
- Datacenter thermal limits and power infrastructure are emerging as the next real constraints for big AI builds. - Grid transformers reportedly face roughly four‑year lead times, and OFC 2026 flagged pluggable optics (XPO) for bandwidth scaling. - Those capacity and thermal issues mean operators are now prioritizing power, cooling, and optical links as build gating factors ( ).
The next limit on big artificial intelligence builds is no longer just chips. Data center operators are running into power hookups, heat removal, and optical link density before they run out of demand. (eetimes.com) A data center turns electricity into computation, and almost all of that electricity ends up as heat that must be carried away. New artificial intelligence racks are far denser than older cloud servers, pushing operators from familiar air cooling toward liquid systems that pull heat off chips and optics directly. (nvidia.com) The U.S. Department of Energy said in December 2024 that data center load growth had tripled over the prior decade and could double or triple again by 2028. Berkeley Lab estimated data centers could reach 6.7% to 12% of U.S. electricity use by 2028, up from about 4.4% in 2023. (energy.gov) That surge is colliding with old grid hardware. Wood Mackenzie data published in August 2025 showed lead times of about 128 weeks for power transformers and 144 weeks for generator step-up transformers, stretching critical electrical gear into roughly 2.5 to 2.8 years before installation. (eepower.com) Some developers and suppliers now describe the longest waits as closer to four years for certain high-voltage equipment and grid upgrades, especially where utilities also need new substations or transmission work. That means a building can be ready before the power path feeding it is finished. (enkiai.com) Cooling has become a second gate. NVIDIA said cooling has historically consumed up to 40% of a data center’s electricity, and its GB200 NVL72 rack-scale systems are designed around liquid cooling rather than standard air-only layouts. (nvidia.com) Networking is the third constraint, because artificial intelligence clusters need huge numbers of optical links to move data between accelerators and switches. At OFC 2026 in Los Angeles, March 15-19, vendors centered their announcements on higher-density optics, lower power per bit, and hardware that can survive liquid-cooled environments. (yolegroup.com) One proposal that drew unusual attention was XPO, short for eXtra-dense Pluggable Optics. Arista said on March 12 that the format targets 12.8 terabits per module, 204.8 terabits per rack unit, cooling up to 400 watts per module, and about 4 times the front-panel density of 1600G OSFP optics. (arista.com) EE Times reported that more than 10 vendors showed live XPO module demos at OFC 2026 and that more than 100 companies had joined the related multi-source agreement. Photonics Spectra said the group was formed ahead of the show to address bandwidth, reliability, and density requirements in hyperscale artificial intelligence data centers. (eetimes.com, photonics.com) The shift is showing up in build plans as well as trade-show booths. At Data Center World this week, infrastructure executives told attendees that artificial intelligence capacity planning now starts with utility access, electrical distribution, and heat rejection, then works backward to servers. (techrepublic.com) That reorders the economics of expansion. In 2026, the gating question for a new artificial intelligence hall is often not how many graphics processors a buyer can procure, but how fast the site can secure megawatts, move heat, and fit the optics needed to keep those processors busy. (rolandberger.com)
