While water cold plates continue to dominate current liquid-cooling adoption, the industry has also turned its attention to different approaches, with two-phase cold plates in particular becoming a promising alternative.
Choosing whether to train a model from scratch or fine-tune an existing one comes down to the use case and cost — with hardware utilization remaining an important cost factor.
Enterprises deploying AI inference need to choose carefully to limit costs and protect their data.
Although cloud platforms often offer the lowest cost for AI inference, on-premises deployment may be preferable due to application architecture, data locality and control requirements.
The cost of AI inference varies widely depending on deployment model, utilization and hardware. This costing tool compares on-premises, colocation and managed AI platforms on a like-for-like basis.
Results from Uptime Institute's 2026 AI Infrastructure Survey (n=1,141) focus on the data center infrastructure currently used or being planned to use to host AI Training and AI Inference, as well as future industry outlooks on the usage of AI. The…
The shortage of DRAM and NAND chips caused by demands of AI data centers is likely to last into 2027, making every server more expensive.
Nvidia CEO Jensen Huang's comment that liquid-cooled AI racks will need no chillers created some turbulence — however, the concept of a chiller-free data center is an old one and is unlikely to suit most operators.
Cybercriminals increasingly target supply chains as entry points for coordinated attacks; however, many vulnerabilities have been overlooked by operators and persist, despite their growing risk and severity.
Data4 needed to test how to build and commission liquid-cooled high-capacity racks before offering them to customers. The operator used a proof-of-concept test to develop an industrialized version, which is now in commercial operation.
Currently, the most straightforward way to support DLC loads in many data centers is to use existing air-cooling infrastructure combined with air-cooled CDUs.
Large-scale AI training is an application of supercomputing. Supercomputing experts at the Yotta 2025 conference agree that operators need to optimize AI training efficiency and develop metrics to account for utilized power.
By raising debt, building data centers and using colos, neoclouds shield hyperscalers from the financial and technological shocks of the AI boom. They share in the upside if demand grows, but are burdened with stranded assets if it stalls.
The data center industry will benefit from the race between Intel and AMD for technical supremacy, but the outlook in terms of power efficiency remains challenging.
Many operators report that they trust AI to draft their MOPs, EOPs and SOPs. But this potentially error-prone approach demands meticulous review by an appropriate member of staff, or operators risk increasing the likelihood of costly downtime.
Jacqueline Davis
Daniel Bizo
Dr. Owen Rogers
Max Smolaks
Paul Carton
Anthony Sbarra
Laurie Williams
John O'Brien
Peter Judge
Rose Weinschenk