New Hardware Tackles AI Data Center Demands
Huawei unveiled its Xinghe AI Fabric 2.0 solution, including an industry-first liquid-cooled 51.2T switch, to handle massive AI workloads. Meanwhile, Ampace is promoting new power systems to manage the volatility and high-density power demands of GPU clusters.
The insatiable demand for computing power in AI is fundamentally reshaping data center architecture, moving beyond general-purpose hardware to highly specialized, integrated systems. This shift requires facilities engineered for extreme power density and cooling, where hardware availability, not just innovation, dictates the pace of scaling. * [Huawei's introduction](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQER2GznU3HK9LtZHYgJYZGqREXkiwrqHJT1uMXuFiKvwSB4p6K_j4E_438zIMER7bUJMhrK3Dc0Iay-6Qk-qx1Kyn0R2xBDtIV55kMN8OEcuN_CsBcPOjwywBUs-agcW2MpZdmMBRJua8lklOw76up1ogtWu937UhUHYf--scsR8V8A6by9AzLv32Nj35KzpW44tfzth5abtsm2C8StnLdOCI7I7L4BUi7oOQ19Ww==) of a 51.2T liquid-cooled switch is a direct response to this thermal challenge. [High-speed components](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQHqe8MjqeYQBwc5HZksYrEcWWRnCQFgtazoOnIqNbXfWcAtXTug8owB30QU0KuqPJdsgxTHKvNtTaUSOXknZXNykVfm48seCJR3Z-953Sdm1LW9Xjo4SUZItKTGvhy-c9ytmKFHcnfPdtPJyzZJtCV2TKLhyHHFHSSWu09-c_ZC4ylkw805npiOxHyM), like the 64x 800 Gbps optics in a 51.2T switch, can generate over 1,000 watts of heat from the front-end components alone. [Liquid cooling offers](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGT58K_7qcMwKXm7VVMkxERBfdZX33kNIN7oDwll2GvQnx9MB-SrR3pA1f8qZFsQ-wfea9hgkLC9VNcQrXI2HJqt8XeOO9A6GaM0vrmV5VNyVFluEpgBfxSRoc0lo0YFbKXtH6_yOhE4EwNfGh8ZsvtTO2KBZ6KIWWnf4VRgK90tL8T9vFN1NcdyWGbmQcNSasBTbOGrvlUtzIIxd5bAGoUIgpq) up to 3,000 times the heat absorption capacity of air, making it a necessity to manage the thermal loads of next-generation hardware. [The global market for](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGAXgL6nIPB5lpcsLiKzuQCVe1np5Tfqve3it_0LrNwNdlaSy5c1j43YMKYgc-q6-t0CKsc7d2OgRlCgSNFAIv3aeeE-kTlmomZYhIwx7qHqyn_7OoMwN9P-dFD2pCXkU-XBk0ly3HTjnYE5mGRhJCPnCYPONyt14IESenSZl3JkTVbnYoocTc7cur3c25UcarJU8WxavAY) data center liquid cooling is projected to grow from $6.65 billion in 2025 to over $29 billion by 2033. * This hardware evolution is driven by the sheer scale of AI models. Since 2019, the power requirements for leading AI supercomputers have doubled annually. A high-end AI GPU like Nvidia's H100 can consume 700W, nearly double its predecessor, the A100. Extrapolating these trends, a leading AI supercomputer in 2030 could require 9 gigawatts of power, equivalent to nine nuclear reactors. * [On the power delivery](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQGcV7w4dyo1-PfEV3niouDhSV_P1Q6PEjZ5WYq6Lnab5zJIjFWae76a1YV_iv3ieAIRW2nbbYrKpQTkbw7ZrdTtdw2GPe7bk5gmhCiLE_82_PStI_hpWfNVw3NFE4ScJQAXVbVYtek4lBK928HX2y7WVOux9D6kNt-ykGfsDJDx2yQdsCgsIJQbkrp_65AP7wdtUon9Thts5eAk4oPwFA==) front, AI workloads create unprecedented volatility. [GPU clusters introduce](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEiC3L3MlHvUHMNBeI45KD_rcfY29g6j1KZSplEpwmpHNrd-dl2_M-EhraQTaiOghdp8Usrb7BDPo1v1WwJNei-udcbku0oEqH5f00UWDSzSdTpsvbj6t1jl_Z4yw7xG-49SypTANxgF8j5m_pKqhBNmzMMpWxtuGWQ7u5MuTitJdv3iZgUxE8u-K3465SUiQGE1pnhpp7kG2ISWS5ZhBdLjv5GkuqIFRidNweqeE_HoX9iu_e-Z734) rapid and unpredictable load swings, stressing traditional power systems. [Companies like Ampace](https://vertexaisearch.cloud.google.com/grounding-api-redirect/AUZIYQEiC3L3MlHvUHMNBeI45KD_rcfY29g6j1KZSplEpwmpHNrd-dl2_M-EhraQTaiOghdp8Usrb7BDPo1v1WwJNei-udcbku0oEqH5f00UWDSzSdTpsvbj6t1jl_Z4yw7xG-49SypTANxgF8j5m_pKqhBNmzMMpWxtuGWQ7u5MuTitJdv3iZgUxE8u-K3465SUiQGE1pnhpp7kG2ISWS5ZhBdLjv5GkuqIFRidNweqeE_HoX9iu_e-Z734) are developing battery energy storage systems (BESS) with semi-solid-state technology designed to absorb these sudden spikes and provide stable, uninterrupted power, a concept they term "AI Continuity Infrastructure." * These advancements are creating a new set of dependencies for startups. Infrastructure is no longer about sourcing interchangeable parts but deploying tightly integrated, often liquid-cooled, hardware stacks. This reality means that for software engineering teams at high-growth companies, understanding the physical constraints of power, cooling, and high-speed interconnects is becoming as critical as the software itself for achieving performance and scalability.
