Amazon plans to move about $8B of Nvidia Grace Blackwell accelerators into a special purpose vehicle that will issue debt and sell a 10% equity stake to outside investors, then lease the hardware back, the Financial Times reported. The chips sit across five Amazon data centers, and an estimate of $60,000 to $70,000 per accelerator implies roughly 114,000 to 133,000 units. Moving the chips and their debt off the balance sheet would improve Amazon's debt-to-equity ratio and could lower its borrowing costs. A hyperscaler is now financing its own accelerators the way neoclouds finance theirs, which links the cost of Amazon's capital to the resale value of Blackwell hardware. (SiliconANGLE)
Public Markets
TeraWulf$15.49▲ 3.9%Mkt Cap: $7.7B
TeraWulf doubled contracted power at its Muskie campus in Kentucky to 1 GW and moved the second 500 MW phase from 2030 to 2029, pending state approval. No tenant is named; the power is secured ahead of a lease. (TeraWulf press release)
onsemi$84.89▲ 6.0%Mkt Cap: $33.1B
onsemi cut its Synaptics offer to $5.7B in cash from about $7B in stock after an unsolicited third-party proposal, with closing expected by mid-2027. The CEO said Synaptics adds to onsemi's AI data center business. (onsemi press release)
Emerging
Chiplet GPUs: A paper from Mario Ibáñez Bolado, Borja Pérez Pavón, Jose Luis Bosque Orero and Julio Ramón Beivide, Divide and conquer: Scalable performance and energy in MCM GPUs, compares multi-chip-module GPU designs by interconnect topology as chiplet counts grow. A 16-chiplet configuration with 256 streaming multiprocessors reaches 2.40x the performance of comparable MCM architectures at 4.45x lower energy. Interconnect topology, and not chiplet count alone, sets how far disaggregated GPUs scale, which puts die-to-die link design at the center of GPU scaling. (arXiv)