The memory-bandwidth bet behind Fractile's inference chips

Key points

Key takeaways from a No Priors interview with Fractile Founder/CEO Walter Goodwin (October 2026):

Full-stack by design. Fractile employs about 150 people across front-end design, physical design, back-end implementation and advanced packaging — deliberately "skinny in every sector" but owning the whole silicon stack so it can run a fast closed loop rather than handing off to an ASIC house.

The hand-off it avoids. Most accelerators are developed with houses like Broadcom — the ~$2T company that owns analog IP for chip-to-chip connectivity and the physical layout that becomes the GDS2 bitmap shipped to TSMC.

A zoo of similar chips. From Google's TPU to OpenAI's Jalapeño, hyperscaler chips share the same bets: HBM DRAM, tensor cores and TSMC advanced packaging — with Meta (MTIA)and Microsoft (Maya) following the same pattern.

From SRAM to DRAM. Fractile's first two years were an SRAM-based chip, like Groq or Cerebras — extreme bandwidth but low capacity. By late 2023 it concluded growing model parameters and context lengths made SRAM unscalable, and pivoted to skunk-works work with memory vendors to push high bandwidth out of cheap DRAM.

Speed is for agents, not chatbots. Goodwin likens the snappier chatbot to Henry Ford's faster horses: the real prize is running multi-trillion-parameter models at many thousands of tokens per second to make long-running agents radically faster.

The economics collapse to memory cost. At data-centre scale, inference economics boil down to cost per gigabyte of the memory employed — today's fast chips pair high bandwidth wi a low-capacity memory, so long-context attention still flips back to a GPU.

Workloads churn, imperatives stay. A new LLM arrives about every two weeks, and attention design in frontier open-source Chinese models changes every few weeks — but every new model still wants orders-of-magnitude more memory bandwidth, the firm bets on that constant.

Carving a 3–6 month edge. Whale an Nvidia system already packs 6–9 custom Nvidia chips, strutturing out a 3-to-6-month vantage through a rolling frontier of aligned bets would, Goodwin argues, win the deployment race.

Read more: No Priors on YouTube