by FractileGB
Serve large AI models much faster and cheaper than GPUs by removing the memory-bandwidth bottleneck that limits how fast each answer is generated.
In development; first chips to customers expected in 2027, with Anthropic as anchor buyer.
Updated 19 Aug 2026Checked 11 Oct0 updates this week
To produce each word, a model reads all its weights. On GPUs those sit in separate memory chips, and the trip, not the maths, sets the speed.
Fractile interleaves SRAM memory cells and compute units on one die, so weights are used where they are stored instead of being shipped across.
Because on-chip memory is small, a large model is split across many chips working in step, linked into inference servers.
| Spec | Fractile inference chip |
|---|---|
| Claimed inference speed | Up to 25x faster than current hardwareR (reported) |
| Claimed inference cost | One tenth of current alternativesR (reported) |
| Memory type | On-chip SRAMR (reported) |
| Initial Anthropic order | About $250MR (reported) |
R reported by the company
Fractile
AI inference chips, memory-compute fusion
Fractile is a London startup designing chips that run trained AI models, with memory and compute woven together on the same silicon. Anthropic signed an initial deal to buy about $250M of its chips, due in 2027.
