CoolIT details liquid cooling coldplate able to cool 4,000W chips

• CoolIT Systems has debuted a new prototype liquid cooling coldplate it claims can cool chips more than three times more powerful than the current generation of Nvidia GPUs.

• The company this week announced a 4,000W-ready single-phase direct liquid cooling (DLC) coldplate – more than doubling the previously accepted limits of single-phase direct liquid cooling.

• Nvidia’s Blackwell GPUs can currently operate at up to 1,200W, with the previous-generation H100s scaling up to a thermal design point (TDP) of 700W. AMD’s MI355X operates at a TDP of 1,100W.

• After around 700W, air cooling for chips becomes increasingly difficult, and liquid-cooling becomes a more suitable option.

• Single-phase DLC works by removing heat from semiconductors with water or a water-glycol mixture running through coldplates attached to the chips. Coldplates transfer heat from the device to a liquid that flows through the plate to a heat exchanger and dissipates the heat.

• Nvidia is set to announce at least some details of its next data center GPUs, Vera Rubin and Blackwell Ultra, at its GTC event later this month. Blackwell Ultra, also known as GB300, is set to operate at a 1,400W TDP.

Vertiv launches new trim cooler with up to 3MW of cooling capacity

• Vertiv has launched a new cooling unit. The company this week announced the introduction of its new CoolLoop Trim Cooler to support air and liquid cooling applications for AI and HPC deployments.

• This solution supports diverse climate conditions for hybrid-cooled or liquid-cooled data centers and 'AI factories,' with Vertiv claiming it offers up to a 70 percent reduction in annual cooling energy consumption via free-cooling and mechanical operation, and 40 percent space savings compared to traditional systems.

• Offering cooling capacity up to 2.6MW and 3MW in the air-cooled configuration, the system supports fluctuating supply water temperatures up to 40°C and cold plate functionality at 45°C (104-113°F).

• The offering can connect directly to immersion cooling systems or integrate with Vertiv’s CoolChip CDU coolant distribution units for direct-to-chip liquid cooling. With free cooling coils optimized for high ambient temperatures, the system is designed to operate in free cooling mode across more seasons and conditions to reduce electrical consumption and CO2e emissions.

Researchers say behind-the-meter geothermal could power US AI data center demands

• Geothermal power could be key to supporting the surge in AI data center demand, think tank Rhodium Group claims.

• Depending on where the data centers are located, researchers believe that the vast majority of power needs could be met by a dramatic build-out of behind-the-meter geothermal power.

• There is currently around 4GW of nameplate capacity from geothermal power plants in the US, primarily in California and Nevada. That is nearly all from conventional hydrothermal systems, which require very specific subsurface conditions.

• Newer approaches, which dig deeper, or involve hydraulic fracturing and horizontal drilling, can be explored in far more locations.

• Rhodium primarily used data center demand projections from the Lawrence Berkeley National Laboratory to forecast how much power consumption would be expected. The researchers also looked at whether data centers will be clustered, as in the pre-AI era, or will continue to spread out in search of power, as seen currently.

• Rhodium also notes that geothermal heat pumps could be used to help cool data centers, reducing the overall electric load.

Block, Inc. to deploy Nvidia DGX SuperPod with GB200s

• Financial firm Block, Inc. is planning to deploy a Nvidia DGX SuperPOD with DGX GB200 systems.

• The liquid-cooled GB200 SuperPod features 36 GB200 Superchips per rack, with each GB200 made out of one Grace CPU and two Blackwell GPUs.

• Each GB200 is connected as one with fifth-generation Nvidia NVLink, delivering 1.4 exaFLOPS of AI performance, 30 terabytes (TB) of fast memory, and 130 terabytes per second (TBps) of bidirectional GPU bandwidth.

• Before the decision to deploy a Nvidia DGX SuperPOD GB200 system, Block used the Lambda 1-Click Clusters - now available with Blackwell GPUs - to trial its hypotheses with hundreds of interconnected Nvidia GPUs.

• Block, Inc. is an American technology and financial services company. It was founded by CEO Jack Dorsey in 2009 initially a provider of point-of-sale systems. In 2024, the company began developing bitcoin mining chips, and signed an agreement with CoreScientic in July 2024 to provide the latter with hardware.

Kao plans third data center at Harlow, UK, campus

• Kao Data is expanding one of its London campuses with a new data center. The company this week announced KLON-03; a new, 17.6MW facility located at its Harlow campus, east of the capital city. Timelines for development weren't shared.

• The facility will offer direct-to-chip liquid cooling alongside traditional air-cooling. Each suite will include hybrid-cooled, hot aisle containment (HAC) systems, which can accommodate densities up to 130kW.

• Kao was launched in 2015, and its Harlow campus opened in 2018; The 61,000 sqm (656,600 sq ft) site was originally set to host four 10MW buildings.

• The Harlow campus currently has two operational data centers - KLON-01 and 02 – each offering 10MW. Work began on the site’s second data center in March 2022 and was completed in late 2023. Nvidia’s Cambridge-1 supercomputer sits within Kao’s original Harlow facility.

• Kao is owned by Infratil, Legal & General, and Goldacre. The company also operates a data center in Slough, a former Barclays data center in Northolt, and is developing a new campus in Manchester.

IREN plans 75MW liquid-cooled AI data center in Texas

• Crypto and AI data center firm IREN is planning to develop a 75MW AI data center in Texas.

• The company revealed the plans as part of its Q2 2025 earning report this week, also announcing another campus in Sweetwater, Texas.

• IREN (also known as Iris Energy) is planning to deploy a new 75MW liquid-cooled data center for AI/HPC at its Childress site in Texas.

• The facility will be known as Horizon 1 and is set for completion in H2 2025. It will be designed to support 200kW per rack via direct-to-chip cooling to host Nvidia Blackwell GPUs.

• The company said it intends to invest $300-350 million in the project. The 420-acre Childress site totals 750MW, with some 450MW currently energized. Around 675MW of the site – located outside Childress County, Texas – is currently earmarked for Bitcoin mining.

• IREN was founded in 2019. As well as the Childress and Sweetwater sites, the company operates the 30MW Canal Flats site in British Columbia, Canada; the 80MW Mackenzie site, also in British Columbia; and the 50MW Prince George site, again in BC.

JLL secures management contract with Pure DC for Abu Dhabi data center

• JLL has announced it has secured a landmark facilities management contract with Pure Data Centres for a data center in Abu Dhabi, UAE.

• Located in Yas Island, the 45MW data center is Pure Data Centres’ first facility in the UAE. The facility went live at the end of last month.

• The company said its offerings will comprise ongoing maintenance and support for Yas Island’s low-voltage and high-voltage electrical systems, including its UPS, switchgear, and HVO-powered generators, as well as its hybrid air and liquid cooling systems.

• Pure Data Centres currently has, both operational and under construction, more than 500MW of IT capacity across the UK, Europe, the Middle East, and Asia.

TSMC proposes JV with chip firms to share ownership of Intel’s foundry division

• TSMC has sought to engage Nvidia, AMD, Qualcomm, and Broadcom in talks about a potential joint venture to operate Intel’s foundry division.

• According to a report from Reuters, under the terms of the proposal being put forward by the Taiwanese chipmaker, it’s seeking a partner in order to ensure it doesn’t own more than 50 percent of Intel’s foundry operations.

• In late September 2024, Intel reportedly rejected an offer from Arm to acquire the company’s product division after the British chip company made a “high-level inquiry” but was ultimately told the unit wasn’t for sale.

• Reports first surfaced in February 2025 claiming that Broadcom and TSMC were weighing up plans to acquire some of Intel’s assets. However, while it was originally claimed that the potential acquisition was at the request of Trump administration officials, it later emerged that President Trump was unlikely to approve of a foreign company operating Intel’s US factories, given his repeated criticism of Taiwan over the issue of chip manufacturing.

• In a post on X (formally Twitter) on March 11, Intel interim co-CEO David Zinsner wrote: “Thank you POTUS, for meeting with Tech CEO Council & for your leadership in driving American innovation. Intel looks forward to collaborating with your administration to strengthen US technology and semiconductor manufacturing leadership.”

ASML plans 'new' reuse and repair center in Beijing amid ongoing US-China chip war

• ASML is planning a “new” reuse and repair center in Bejing, China, according to the company’s most recent annual report.

• In a segment discussing the chip manufacturing equipment maker’s approach to the circular economy, the company said it had a “dedicated reuse and repair organization” and was planning to “open a new” center in Beijing in 2025, “marking another important step-up in reuse manufacturing.”

• In addition to mainland China, ASML also has repair centers in South Korea and Taiwan in Asia; Wilton, Connecticut, San Diego, California, and Vancouver, Washington State, in the US; and Veldhoven in the Netherlands.

• ASML is the sole global supplier of extreme ultraviolet lithography (EUV) photolithography machines that are needed to make the most advanced 3nm and 5nm chips. Based in the Eindhoven suburb of Veldhoven, the company has been at the center of the US government’s ongoing trade war with China, with the Dutch government increasingly succumbing to pressure from the Biden administration to block exports of ASML products to China.

• The US government claims this is to stop China from developing advanced AI systems and to prevent it from having access to technology that could be used for military modernization and human rights abuses however, former ASML CEO Peter Wennink argued in 2024 that the export restrictions are largely ideologically driven rather than based on facts.

• In January 2025, the company reported that it had sold 38 refurbished systems in 2024, a ten-unit increase on 2023, and to date, has refurbished and resold over 500 lithography systems.

One hospitalized after fire at Singapore data center

• A fire broke out at a data center in Singapore this morning after a server reportedly "exploded."

• Local press including the Straits Times said one person was taken to hospital following the fire, at 750C Chai Chee Road. Emergency services were called to the scene in the early hours of Friday, March 14.

• The sprinkler system was also activated, and one person was taken to Singapore General Hospital for smoke inhalation. Power to the building was shut off.

• An email sent out by SG.GS, the cloud data center provider partner of Singapore Internet Exchange, said an “explosion” caused the fire.

• Initial checks showed that servers within the center are still in service, said the e-mail, but the company “needed some time to recover.“

CoreWeave to deploy Nvidia GB200 NVL72 cluster at Bulk Infrastructure data center in Norway

• CoreWeave has selected a Bulk Infrastructure data center in Norway to host a Nvidia GB200 NVL72 cluster deployment.

• The cluster will be located at Bulk's N01 data center Campus in Vennesla, and will have the Nvidia GB200 NVL72 configuration, interconnected by Nvidia Quantum-2 InfiniBand networking.

• The size of the cluster has not been shared, but has been described by the companies as "one of the largest Nvidia AI deployments in Europe.“

• CoreWeave began offering the Nvidia GB200 NVL72 instances in February of this year, initially from the company's US-West-01 region. The initial offering of GB200 NVL72-based instances on CoreWeave connect 36 Nvidia Grace CPUs and 72 Nvidia Blackwell GPUs in a liquid-cooled, rack-scale design, are available as bare-metal instances through CoreWeave Kubernetes Service, and are scalable up to 110,000 GPUs.