Tenstorrent BlackHole Galaxy reaches production; 10 customers, 96-unit order

Tenstorrent, led by Jim Keller, has moved Tenstorrent Galaxy Blackhole AI servers into production and customer deployment. The company reports ten customers with Galaxy systems on-site, having passed proof-of-concept, and received the largest purchase order of 96 Galaxy clusters (3,072 Blackhole chips total) to ship outside the U.S. The company is building 1,000 Galaxy servers, with at least half already sold, backed by over $1 billion raised to date.

The Galaxy Blackhole server delivers 23 PFLOPS of Block FP8 compute from 32 Blackhole chips, with 6.2 GB on-chip SRAM, 1 TB DRAM, and up to 56 × 800G Ethernet ports. In early tests, 16 Galaxy servers running DeepSeek-671B delivered 350+ tokens per second per user at batch 32, demonstrating unified compute, memory, and networking on a single chip—eliminating the KV cache fragmentation problems that plague disaggregated systems. A single Galaxy server starts at $110,000; a four-system supercluster at $440,000.

Keller told EE Times that Tenstorrent can match or exceed GPU and specialized accelerator performance at substantially lower hardware cost. He noted that some customers with large Nvidia orders facing year-plus delivery delays have opted for cheaper Tenstorrent capacity, positioning Blackhole as a hedge against supply bottlenecks. Tenstorrent is aiming for an IPO, with investors highly interested; Keller views a strategic JV or partnership with sovereign infrastructure buyers as more likely than acquisition.