JapanFold Claims One Tenstorrent Server Beats an Nvidia DGX B200 on Bio Models at a Third of the Price

DavidBennett__ · x · 2026-10-16

JapanFold launches as an agent-native bio-model inference platform

Built by @aiand, JapanFold pitches "high throughput, low cost, agent native" access to leading bio models: structure prediction, binding affinity, de-novo design and protein embeddings.
Speed: TT-Bio got up to 4x faster in two weeks. Protenix-v2, OpenFold3 and OpenDDE now fold large complexes 2.4x–4.2x faster via fused kernels, less memory traffic and more work on Tenstorrent chips.
Hardware claim: one Tenstorrent Blackhole Galaxy server outruns an Nvidia DGX B200 on OpenFold3 and Boltz-2 (exact mode both sides, Anthropic's kit on the B200) at under a third of the price.
Models supported: structure — Boltz-2, ESMFold-2, OpenFold3, OpenBind-0, RoseTTAFold3, OpenDDE, AF2-IG; design — BoltzGen, RFdiffusion3, PXDesign; embeddings — ESMC, SaProt.
Access: browser Workbench, a single HTTP endpoint (api.japanfold.aiand.com/v1/predictions), and an agent skill (npx skills add japanfold/japanfold) for Claude Code, Cursor, Codex and more.
Pricing: $0.26 per processor hour, one rate for every model, billed on chip time actually held; queueing is free. Self-reported Boltz-2 fold costs: $0.0022 for 256 aa, $0.0062 for 512 aa, $0.022 for 1024 aa.
Extras: $100 free credit, no card required, 100% Japan-hosted, claimed 1:1 accuracy parity with reference implementations.

More from Infra