Jim Keller: ‘AI Still Obeys the Old Laws of Compute’

At Tenstorrent's TT-Deploy event, the company demonstrated the performance of its BlackHole Galaxy servers in large-scale deployments, showing that they can outperform traditional GPUs and specialized... Read original →

Tenstorrent’s inference-centric architecture—prioritizing memory bandwidth and network density over raw FLOPs—is triggering a stack-level redesign across AI hardware. With 56 Ethernet ports per server, BlackHole Galaxy undermines Nvidia’s dominance in distributed inference and pressures Cerebras to re-evaluate on-chip cache economics. This shift redirects advanced-node (3nm and below) EUV allocation away from monolithic GPU-like dies toward disaggregated, energy-efficient tensor engines. Geopolitically, Tenstorrent’s push for a non-U.S.-restricted supply chain positions its RISC-V CPU IP favorably with hyperscalers in Taiwan, China, South Korea, and Europe. Rather than accept acquisition by Intel or Qualcomm, the firm is likely to pursue an IPO to retain control. If major cloud providers adopt its fully on-chip KV cache approach within 12–18 months, GPU-free inference clusters could become the new standard, collapsing traditional AI server BOM assumptions.