Tarrasque Silicon
Own your intelligence. Don't rent it.
Cheap, cluster-scale AI-inference silicon built on commodity memory and a lean NRZ interconnect — no HBM, no CoWoS, no leading-edge node.
Inference isn't compute-bound — it's memory-bound and capacity-bound. The industry pays a fortune for HBM and advanced packaging it doesn't need to serve models. We don't.
Commodity memory, not HBM
LPDDR/GDDR on a mature node delivers the capacity and bandwidth serving actually needs — at a fraction of the cost and with a multi-foundry supply chain.
Loom-Link — a lean NRZ fabric
Only small activations cross chips; the KV cache stays local. The interconnect runs under 10% utilized, so cheap 28G NRZ is plenty — no PAM4, no switched fabric, no leading-edge SerDes.
Scale by adding chips
Cluster-scale memory is the aggregate of local memories. Add a chip, grow the model — the cost curve stays flat as models grow.
Verified from day one
A golden-verified RTL stack today; a two-board NRZ hardware benchmark next; an MPW test chip after that. Rigor before mask.
The full technical thesis — memory, interconnect, node choice and unit economics — is in the whitepaper.