Fractile

AI inference chips, memory-compute fusion

GB$1B privateest. 2022fractile.ai (opens fractile.ai)Checked 11 Oct

The brief

Analyst view
  • Anthropic signed an initial ~$250M chip purchase deal in 2026, for chips not ready until 2027.
  • Raised $220M at about $1B in May 2026; in August was in talks to raise ~$600M at $6.5B pre-money.
  • Claims up to 25x faster model serving at a tenth of the cost; no product has shipped yet.

Technical approach

As reported
Memory

Compute and memory interleaved

Serving AI models is limited by moving weights from memory to compute. Fractile puts fast on-chip SRAM right next to the arithmetic, so data barely travels.

Target

Fast answers for many users at once

It aims to serve thousands of tokens per second to thousands of users at the same time, where GPUs usually trade speed per user against total throughput.

Stack

From circuits to servers

The team designs everything from transistor-level circuits to the software that runs inference servers, so the chip is built around real model workloads.

Primer