By: Alex Mercer – SeaPRwire – Enterprises chasing AI performance hit the same wall fast. Models keep evolving. Hardware bets lock in. Switching costs climb. Tenstorrent claims a different path. One architecture that handles language, video, and agentic workloads faster than GPUs while scaling from a licensable core to massive superclusters over plain Ethernet. At TT-Deploy JP, the company backed those claims with fresh records, a new CPU IP, and its biggest deployment yet in Japan.
The numbers stand out. On Kimi K2.6, Tenstorrent Galaxy Blackhole superclusters deliver 900 tokens per second per user, three times faster than GPUs. DeepSeek-R1-0528 671B reaches over 400 tokens per second per user, improved from earlier benchmarks. For video, LTX 2.3 Fast generates roughly six-second clips at 144 frames in 1080p with audio and lip-sync, four times quicker than GPU setups. These gains span different model families on the same foundation. Capacity grows near-linearly when adding more Galaxies. That efficiency matters for companies running premium inference at scale without constant hardware refreshes.
TT-Ascalon S expands the portfolio for agentic AI. This RISC-V CPU targets orchestration, I/O, and latency demands rather than raw compute. It packs density at about 50 percent the footprint of TT-Ascalon X while delivering roughly 140 percent performance per square millimeter. The design prioritizes power efficiency and handles branch-heavy, tool-connected patterns common in agent runtimes. Beyond agents, it fits high-efficiency servers, networking, storage SoCs, and edge deployments. As licensable IP, customers can integrate it into custom silicon.
Networked AI ties it together. Accelerators and CPUs connect over standard Ethernet with an open-source software stack. Galaxies and superclusters work standalone or drop into existing GPU fleets. No full rip-and-replace. Customers add capacity without locking into one model or vendor. Systems adapt as models change. Control stays internal. Jim Keller, CEO, put it plainly: the architecture runs everything, integrates with what you own, and scales from core to supercluster. This lets companies and countries own their AI amid constant shifts.
In Japan the approach lands at national scale. The largest deployment runs with ai&, a vertically integrated frontier AI platform. Over 120 Tenstorrent Galaxy systems support chat, RAG, vision, and post-training workloads entirely within the country. This marks the biggest sovereign AI compute footprint in the region. David Bennett, CEO and co-founder of ai&, noted that routing workloads to the best silicon proves itself on real enterprise needs.
Partnerships run deeper. Turing tested Blackhole inside an autonomous vehicle. Through the national 2nm program with Rapidus, adopted by NEDO and led by LSTC, Tenstorrent contributes the RISC-V CPU chiplet. The company has operated in Tokyo since 2023, runs an AI data center in Osaka, and brings up to 200 Japanese silicon engineers into design teams. TT-Deploy JP featured live demos and partners including ai&, Rapidus, Preferred Networks, Socionext, and Turing.
The bet is clear. AI infrastructure no longer needs single-vendor lock-in or constant forklift upgrades. A unified, open architecture reduces risk. Enterprises gain flexibility. Nations secure sovereign capabilities. For operators weighing next hardware cycles, the practical move starts with testing licensable IP or small Galaxy clusters against current workloads. Measure real tokens per dollar and latency under mixed agent flows before scaling. That data will decide if the single-architecture promise holds in production.
Author bio: Alex Mercer, long-term international tech journal commentator with over two decades covering semiconductor shifts and AI infrastructure deployments from Silicon Valley to Asia.
source https://newsroom.seaprwire.com/press-releases/technologies/tenstorrents-japan-push-one-architecture-to-challenge-gpu-dominance-in-sovereign-ai/
