Startup company plans huge AI cloud with AMD chips

Companies that want to operate large AI models have to expect waiting times: On the one hand, AI accelerators from Nvidia are not available in random quantities, and on the other, data centers are struggling with bottlenecks at power suppliers. The US companies Tensorwave and Tecfusions are now cooperating to solve both problems: they promise to quickly provide high AI computing power on servers with AMD Instinct MI300X and MI325X from 2025.

The AI clusters will also be among the first to be networked via Ultra Ethernet; AMD sells the Pensando Pollara 400 network adapter for this purpose.

According to Tensorwave, it is a partner of AMD and has secured one gigawatt of capacity in data centers from Tecfusions. However, AI servers that consume this amount of power have so far cost several billion US dollars in total.

In order to provide data centers as quickly and resource-efficiently as possible, Tecfusions buys old industrial buildings and data centers and converts them to meet the higher requirements of AI data centers with more than 20 kilowatts of capacity per rack. Tecfusions calls this "adaptive reuse". For example, the company has taken over a data center in Clarksville, Virginia, which was previously operated for the Department of Homeland Security, among others.

Elon Musk's company xAI also used this principle to build the Colossus AI data center into an existing factory building formerly used by Electrolux in Memphis, Tennessee.

However, it is not so easy to find out where Tecfusions actually sources the electricity for the data centers. The company's website refers vaguely to "on-site microgrid generation" and renewable energy sources such as photovoltaics and hydropower. Natural gas is also said to be used. In any case, 300 megawatts are immediately available.

(ciw)