The Frontier

Goodwin founded Fractile in 2022 to build specialized inference chips focused on scaling test-time compute. The company now operates with a lean team of 150 people to manage the entire design process in-house.

Goodwin notes that hyperscaler custom chips rely heavily on Broadcom, a $2 trillion company, to translate high-level architectural designs into physical layouts and TSMC-ready packaging.

Fractile abandoned its early SRAM-based architecture after realizing that expanding model context lengths and parameter sizes made SRAM unscalable. The company pivoted to a DRAM-based design to secure higher capacity at a lower cost.

Goodwin points out a massive hardware bottleneck, noting processing power scaled a million-fold over 20 years while memory bandwidth grew only 40-fold. Fractile designs chips with 25 times more bandwidth than current HBM-based systems.

Goodwin projects that fully automated chip prototyping from initial architect intent to physical GDS2 layouts will arrive within two and a half years. This target is four times faster than the conventional industry timeline.

Goodwin argues that the primary value of ultra-fast inference is enabling long-running agentic workflows rather than snappier chatbots. Running massive models at thousands of tokens per second acts as a direct capability elevator for complex reasoning tasks.

Goodwin states that while increasing sparsity in Mixture of Experts models reduces overall compute requirements, current GPU memory bandwidth limitations prevent these architectures from running efficiently at high sparsity levels.

Goodwin warns that frontier labs face extreme strategic danger if they rely solely on proprietary, in-house silicon. Sudden shifts in machine learning architectures could render custom hardware obsolete, leaving those labs months behind competitors.