Liquid Cooling Becomes Default for AI Data Centers
The intense energy demands of AI are forcing a fundamental shift in data center design, with liquid cooling emerging as the new standard. Traditional air cooling maxes out at rack densities of 41kW, while liquid cooling can handle over 200kW. While retrofitting costs $2-3 million per megawatt, it delivers a 40% energy saving, a crucial trade-off as hyperscalers pour billions into new AI-focused facilities.
The performance demands of modern AI chips are a key driver behind the liquid cooling trend. NVIDIA's H100 GPUs, for example, can draw up to 700W each, and their upcoming Blackwell B200 chips are expected to have a Thermal Design Power (TDP) of 1,000W. A single rack of servers equipped with these powerful processors can generate heat loads far exceeding the capabilities of traditional air cooling. * [Two primary liquid cooling](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHuH1kc_lpEZ0qFOa7larRcIzp6tBgbK36sN2pzasVjUVotfb0-qWDjmiFdViHc4yXkTNjdSEjh_Cgx4b3BnqEKdde8RfNP5JY7szzvA1vyk_eQdQPlqVsA_hubj2CDb0hH5TdByj_lIMOQ4_VbthMP_Aie6RJfAZhvFU_MUbrpLFcVUd8HMWqQWpvawQNqROx6aqugVsK_dj-QoP4KBHb4x-OuuKw-9CG0OaXsWw==) methods are being widely adopted: direct-to-chip (DTC) and immersion cooling. [Direct-to-chip systems](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF92EKc26AIeYTBlxbjehnVGrFEm-XFZjCAniqUPQDUUtYQTcBIcVx3-w47_ukkl24r6DEHUyfExS7v6D3He7VuGMrEdoS75Cwq8ifGAPOIRWPG45b5-oTJrEuikoONNjTPo0e9BulDaGt7xT656W5bmrL-eE6BB9N_MXK8IqpXUs5GCl430CS9TdouCG7i5c7bsYBiKIquAA==) use cold plates to circulate coolant directly over hot components like CPUs and GPUs, while immersion cooling involves submerging entire servers in a non-conductive, dielectric fluid. [While DTC can be more](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHq-cqfQrJcY3yOyKwKlfdAOWHLRraHQU3MQCcavYEtfZ0F9TxsVEO5MDTFT3gHLTQuHEY6S8oGmW8OQq5yS7UBPgH63SWEPveZzor9rqHt8YCdmKYKX_ycTjoZk1j1iq58p-faxKctUbxtFWMQRjwgGBPn2NTs3OB-A32YeTeL7yPVtjxytyuZUptmcVBZ7vVydO9I5Adr) easily retrofitted into existing data centers, immersion cooling offers more uniform heat dissipation. * This technological shift is creating opportunities for both established players and startups. Companies like Vertiv, Schneider Electric, and CoolIT Systems are major providers of liquid cooling solutions. At the same time, a new wave of startups, including Iceotope, LiquidStack, and Submer, are innovating with specialized liquid cooling technologies tailored for AI infrastructure. * The move to liquid cooling also has significant environmental implications, particularly concerning water usage. [While some liquid cooling](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHDGUEVoWAes2yxplhPDiaTI23TgPafBF24NeTTzJjckO8nA-UO9P3XC1jTg96Y-UbKGPNC-BPzAu-bh2Pva_Uw2h2BAEemrDxoLbkQIqLTn1coJP7nHaa8MFQGVFkS71L4HLsvdw0Ocz9L7ZJLPkvgKQLGqzsVU1nh84jflCIm5w==) systems are closed-loop, many traditional data centers rely on evaporative cooling, which can consume millions of gallons of water per day. [Advanced liquid cooling](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQF63U1QphWQxyhGZrV3qXg0StFOeJR1PprwXa0K9puKuqhVeaANaMB3f4P1b6fUpO8iXAB9_ELpJTy_2HWGSOMSEsmY3Lh9GeHPeqQGwC3r0a_dzsfmjLXNmtbFMZbYziVhfh8TpB2_06E-vFYj5ifK_IiSieqyol5g7Qe2yH-1JDLj8EDObuLoJNDX0Dughw4MrkkBops=) techniques can dramatically reduce this water consumption, with some direct-to-chip solutions cutting water needs by up to 95%. * Waste heat from liquid-cooled data centers is increasingly being viewed as a resource. The captured thermal energy can be repurposed for applications like district heating, providing a sustainable energy source for nearby communities and further improving the overall energy efficiency of the data center. This creates a circular energy economy, turning a data center byproduct into a valuable asset.
