A 1 megawatt solid-state transformer sits in Lenox as idle Macs start serving model requests
North Carolina State University's 1 megawatt transformer is on a Lenox field test while Amperesand, Heron Power and DG Matrix have raised more than 280 million dollars. Eigen Labs pays idle Macs, and SCX.ai adds a Sydney data layer.
Artificial Intelligence··Morning
A 1 megawatt box sits in Lenox until September
Conventional power transformers are still hand-wound to a nineteenth-century design, are custom-built for each substation and can take several years to arrive, so AI data centres are backing solid-state transformers assembled from silicon carbide switching electronics. Amperesand, Heron Power and DG Matrix have together raised more than 280 million dollars over the past year to commercialise them. North Carolina State University's 1 megawatt unit has sat in a shipping container at the Electric Power Research Institute's site in Lenox, Massachusetts since May 2026 and returns to the university in September after a summer of testing; it comes out of work with the New York Power Authority that began in 2018. Srdjan Lukic, the NC State professor leading it, says data centres are what is pulling the technology forward, because a solid-state unit converts grid alternating current straight into the direct current the racks want and removes a separate conversion stage. Engineering and commercialisation work still lies ahead.[1]
Eigen Labs pays idle Macs on OpenRouter
Seattle company Eigen Labs has turned its Darkbloom distributed inference network into a paid provider on OpenRouter, paying owners of Apple Silicon machines for capacity they are not using. The company puts typical earnings at 120 to 200 dollars a month for a Mac and says roughly 250 machines are online at any moment. Eigen Labs engineer Gajesh Naik says the network has served 4.5 billion tokens and reached annual recurring revenue of 102,000 dollars, with more than 900 registered providers and 42 million inference requests since an April 2026 experiment. The documentation puts the Mac mini forward as the machine best suited to the job and the terms bar renting cloud instances for the purpose. Inference runs inside a single hardened Swift process on the Apple Silicon graphics processor, with debugger attachment refused at the system call level, and requests are routed by a coordinator running in a confidential virtual machine.[2]
SCX.ai adds a data layer at Equinix SY5 in Sydney
SCX.ai and DDN said they plan to pair SCX's chip-accelerated inference infrastructure with DDN's Infinia data platform to expand what the two companies describe as Australia's first fully domestic enterprise AI cloud. The first node runs at the Equinix SY5 data centre in Sydney and a second is targeted for the end of 2026. SCX runs air-cooled SambaNova SN40L processors and claims 2.5 to 5.6 times the performance per watt of graphics-processor systems, a vendor comparison offered with no stated workload or measurement boundary. The company raised 40 million dollars in its listing and reported contracted annual recurring revenue of 6.5 million dollars at the end of July 2026, with more than 400 active users. DDN's chief technology officer Sven Oehme frames the pairing as removing input and output bottlenecks so that inference can run at sub-millisecond latency.[3]
Related columns
For more information on this topic, you can read the related columns.
