Tenstorrent Unveils Next-Gen Servers for Fast Tokens, No Disaggregation Needed

Tenstorrent is set to unveil its next-generation Galaxy Blackhole servers and cluster systems, designed to deliver fast token generation without relying on the industry trend toward disaggregated infe... Read original →

By rejecting disaggregated inference and unifying prefill/decode on a single platform, Tenstorrent is forcing a paradigm shift back toward integrated AI server design. Its 32-chip Galaxy Blackhole system leverages 3nm EUV and massive on-chip SRAM to reduce DRAM bottlenecks—but intensifies co-design demands on EDA (e.g., Siemens) and advanced packaging. While sidestepping some U.S. export controls, reliance on TSMC’s 3nm node introduces supply chain fragility. Against NVIDIA’s CUDA moat, Tenstorrent’s bet on TTLang and general-purpose compatibility aims to lure enterprises frustrated by vendor lock-in—potentially pressuring NVIDIA to open its inference stack faster. If Equinix’s Distributed AI Hub proves deployment efficiency within 18 months, the industry may pivot from raw FLOPS to TCO-optimized edge clusters, undermining the NVLink-centric HPC orthodoxy.