A Chip Built Around Causal Reasoning, Not Matrix Math

A three-unit architecture with hybrid compute-in-memory, targeting B300-class cloud parts and Thor-class edge parts.

Tongxin Intelligence, founded in April 2026 and incubated by the Beijing Institute for General Artificial Intelligence under Zhu Songchun, confirmed on 12 November 2026 that it had raised several hundred million yuan in seed and angel funding, with backers including the Tsinghua alumni fund, the institute itself, Songlin Technology and Pingzhi Information. Its architecture pairs Vector and Tensor units for CNN and Transformer work with an AGI Core aimed at graph-based logical, causal and value reasoning, using hybrid compute-in-memory against the memory wall and Chiplet packaging to span cloud to edge; the first generation benchmarks against Nvidia’s B300 in the cloud and Thor at the edge, with FPGA validation scheduled for early December. The commercial read is that this is a bet on inference-era economics — the pitch is cost per token, not peak FLOPS, and the December prototype is the gate that decides whether the story survives.