Terakraft, an AI datacenter operator, and Neurophos, an AI chip company, have entered into a collaboration agreement to provide sustainable, high performance and energy efficient AI infrastructure. The companies plan to host a pilot as part of a commercial early access program in 2027 for Neurophos’ accelerated AI inference platform. The project will provide a real-world proving ground for sustainable, ultra-efficient compute — complementing today’s GPU-driven systems with experimental next-generation photonic hardware. Neurophos’ optical processing unit (shown) will be integrated into Terakraft’s green energy-powered data centers to improve energy efficiency and compute power. Courtesy of Neurophos. Neurophos' technology leverages the efficiency of photonics and the scaling properties of metamaterials and miniaturizes optical modulators by a factor of 10,000×, allowing future optical processing units to deliver the compute power of 100 GPUs while consuming the equivalent of 1% of the energy, the company said. According to the company, its proprietary optical processing units are 100× more energy efficient than leading GPUs for the same workload. A spinout of Duke University and metamaterials commercial technology incubator Metacept Inc., Neurophos raised $7.2 million in December 2023 to commercialize metamaterial and optical AI inference chips targeting data center applications. The company's technology combines an optical metasurface for silicon photonic computing with a compute-in-memory (CIM) processor architecture fed by high-speed silicon photonics for fast, efficient matrix-matrix multiplications. Additionally, the company uses a standard CMOS process, Patrick Bowen, CEO of Neurophos, said at the time of the financing. Located in Norway, Terakraft’s AI infrastructure is powered entirely by renewable hydropower and cooled with natural lake water, earning the company a Power Usage Effectiveness rating below 1.1.