AI is rewriting how power flows through the datacenter (www.theregister.com)

AI is transforming power management within datacenters, driven by increased rack densities and the rising energy demands of GPUs and other accelerators. With high-performance computing now requiring over 1 MW per rack, traditional electrical distribution systems are strained. Notably, Nvidia's recent DGX GB200 system demands 120 kW, illustrating the urgency for innovative solutions. The shift towards solid-state transformers (SSTs) is emerging as a key advancement, as they are more compact, efficient, and capable of handling high power loads with silicon carbide components. Infineon expects SSTs to capture a sizable segment of the datacenter transformer market, valued at up to $1 billion by 2030. The implications of these technological changes extend to power distribution architecture as well. Companies like Nvidia advocate for 800 VDC supplies to reduce conversion inefficiencies while managing rising power requirements. The integration of advanced semiconductor materials like silicon carbide and gallium nitride is critical for these developments. Furthermore, voltage regulator modules (VRMs) are evolving to minimize energy loss, with a future goal of embedding them directly within chip substrates. These innovations signal a dramatic shift in datacenter infrastructure, underscoring the growing importance of power semiconductors alongside traditional computing hardware.