Data Centre Cooling Becomes the Bottleneck as AI Racks Push Past 120 kW

The proposed decision, released on 30 September by administrative law judges at the Public Utilities Commission of Texas, found that El Paso Electric failed to adequately consider alternatives to the McCloud facility and did not issue a request for proposals to contractors. As Inside Climate News reported, the judges wrote that a Meta subsidiary, Wurldwide, "drove the process" through its prescribed timeline and its agreement to pay capital and operating costs for the first five years.

The plant would use 813 modular natural gas generators from Enchanted Rock, each rated at 450 kilowatts. Total estimated cost: $499 million. Meta's El Paso site would be the utility's largest electric customer at 1 gigawatt. El Paso sits outside ERCOT, where utilities are barred from owning generation.

Cooling moves up the stack

While Texas argues over supply, the equipment side is reorganising around the same problem. Midea Building Technologies used Data Centre World Asia, held 29 to 30 September at Marina Bay Sands in Singapore, to pitch what it calls a power-cooling hyperconverged architecture, per Thermal Control Business Update. The company claims its Magnetic CDU can cut equipment footprint by up to 70% and support PUE below 1.2, with an industrial-grade CDU rated at 2.6 MW of cooling capacity. Its Gui'an project with Keppel is claimed to deliver up to 7,654 hours of free cooling annually. SUNON, meanwhile, is taking a chip-to-rack approach at the 2026 OCP Global Summit, scheduled for 12 to 15 October in San Jose, according to TechEdge AI.

The company cites Gartner's forecast of 565 TWh of data centre electricity consumption in 2026, a 26% increase on 2025, with AI-optimised servers at 31% of that total. The IEA figure in the same piece puts 2024 consumption at about 415 TWh, roughly 1.5% of global electricity.

The silicon underneath is changing too. Upsite Technologies notes that NVIDIA's Blackwell chips carry more than 200 billion transistors and that the Vera Rubin platform is claimed to cut inference cost per token by up to 10 times. Broadcom's Tomahawk 6 supports up to 102.4 Tb/s per chip, and Cisco's Silicon One G300 offers the same 102.4 Tbps with 200 Gbps SerDes. Networking gear at that density is starting to need liquid cooling, not just the GPUs. Renesas added a component-level answer on 30 September.

Electronics For You reported the launch of a 650V gallium nitride device, the TP65H020G4PLSGBD, with 20 milliohms on-resistance and dual-side cooling in an 8 x 8 mm package. Renesas says top-side thermal impedance falls 10% versus top-side-cooled parts and the package is 57% smaller than the 10 x 15 mm TOLT it replaces. It targets the 800V DC/DC intermediate bus converter stage.

Transparency fights

In Europe, the argument is about disclosure rather than hardware. POLITICO reported that Lighthouse Reports filed a complaint under the Aarhus Convention accusing the European Commission of building a "wall of silence" around data centre energy use. Commission figures put EU data centre consumption at 20.7 TWh of electricity and more than 8 million cubic metres of water in 2025, up 26% and 52% respectively, but only in aggregate. A 2024 law banned disclosure of per-facility metrics. A compliance committee decides on admissibility in November.

None of this resolves the underlying arithmetic: more compute means more heat, and more heat means more power to move it. GOTTOGPOWER's own write-up makes the operational point plainly, noting that cooling failures rarely cause immediate outages but shorten component life and raise running costs. The El Paso decision, whenever the five PUCT commissioners vote, will test who pays for that.