Frontier hardware: brain interfaces and orbital compute

Reports highlight breakthroughs in brain interfaces and space-based compute. - BISC: ultra-thin wireless brain chip with 65,536 electrodes (256x256), 50 μm thickness. - BISC streams high-bandwidth neural data and decodes vision and movement in real time; tested in animals but no human trials yet. - BISC development involves Stanford, Columbia, UPenn, and DARPA funding. - Starcloud ran an H100 in orbit training a nanoGPT and running models in space; returned messages from orbital model. - Starcloud argues orbital compute offers near-limitless solar energy, cooler temperatures, and avoids terrestrial cooling/water constraints. - Companies like Google, SpaceX, and Blue Origin are exploring orbital data centers.
Key insights
BISC chip demonstrates high-density wireless neural streaming in animal tests
Researchers unveiled BISC, a wireless brain interface described as a 50 μm-thin chip with a 256x256 electrode array (65,536 channels). BISC reportedly streams high-bandwidth neural signals wirelessly at rates comparable to compressed 4K video and decodes vision, movement, and sensory cortex signals in real time; tests have been performed in pigs and monkeys with no human trials announced. The project involves multiple universities and DARPA funding.
Starcloud demonstrates training and inference of LLMs in orbit to argue for orbital data centers
Starcloud used an H100 GPU on a Starcloud 1 satellite to train and run models (nanoGPT and other models) in orbit, sending test outputs back to Earth. The company cites advantages of near-constant solar energy, cold ambient conditions, and freed land/cooling constraints to justify plans for large-scale orbital compute capacity, and notes industry interest from hyperscalers and space companies.