Palo Alto–based Mythic has developed Analog Processing Units (APUs) that collapse compute and memory into a single plane, achieving around 120 TOPS-per-watt—roughly 100 × the energy efficiency of top-tier GPUs—even when moving data. The chips, already in production and validated by the U.S. DoD, major auto OEMs and defense contractors, target four trillion-dollar markets: data centers, automotive, robotics and defense, while also scaling to 1-trillion-parameter large language models. Internal benchmarks show up to 750× more tokens-per-second-per-watt and as much as an 80× cost advantage per million tokens versus NVIDIA’s highest-end GPUs. Beyond standalone APUs, Mythic is commercializing “Starlight” sub-1 W sensing modules that integrate its compute-in-memory technology directly into image sensors for 50× performance gains in low-light scenarios. The company offers a mature CAMP SDK that supports ONNX, PyTorch and TensorFlow and has been rated best-in-class among compute-in-memory peers. Led by former NVIDIA executive Taner Ozcelik, Mythic’s roadmap includes enabling ChatGPT-class models to run locally on handheld devices at roughly 1/100th today’s cost. The recent $125 million raise gives the firm substantial runway to scale manufacturing and expand its product portfolio.

Total raised

$287M

Funding rounds

6

Latest round

Series D

Latest activity

Dec 2025