Tensordyne Napier AI Inference System

Tensordyne sells AI inference systems for data centres, with offices in Sunnyvale, California and Munich, Germany. It targets operators that want more tokens or revenue from a fixed power envelope, and its core technology is logarithmic math applied to transformer models.

Napier combines the TDN AIP 3nm processor, 72-chip air-cooled TDN72 pods, four of which form a TDN Rack, and the NapierLink interconnect with under 1,000 ns latency and 1 TB/s any-to-any bandwidth. A Python-style eDSL, a Hugging Face model zoo and logarithmic quantization make up the software, with a beta program open by application.

• Napier AI Processor: 3nm TDN AIP with a Universal Inference Engine, HBM and SRAM.
• Pod and rack design: 72-chip air-cooled TDN72 pods, four per TDN Rack.
• NapierLink: under 1,000 ns latency and 1 TB/s any-to-any bandwidth.
• Logarithmic math: replaces multiplication with addition to cut power and die area.
• Software stack: Torch-like eDSL, curated Hugging Face model zoo, automatic workload balancing and scaling.
• Inference modes: standard and disaggregated, toggled in software without extra hardware.

A cloud or neocloud operator in Southeast Asia can host large language model serving for enterprise customers in an air-cooled hall. An enterprise running on-premises inference for MoE models can reduce footprint and power. A model provider can serve long-context agentic workloads at high per-user speed.

Data centre operators can raise token output per megawatt within existing air-cooled facilities. The company claims 13x tokens per second per rack and 17x tokens per MW against an Nvidia NVL72 GB300 on DeepSeek-R1, based on internal simulation. Standard and disaggregated inference modes switch in software, and Mixture of Experts models up to 10 trillion parameters are supported.

Inference cost and power now dominate AI infrastructure spending as agentic and Mixture of Experts workloads grow. Power-limited data centres look for higher tokens per watt without liquid cooling. Southeast Asian cloud and sovereign AI operators are expanding capacity and may evaluate alternatives to GPU racks.

Offices in Sunnyvale, California and Munich, Germany. Technology partners named on the site are Broadcom, TSMC and Juniper Networks. First-generation Scorpio chip taped out in 2021, and Napier is now moving into high-volume manufacturing at TSMC. The SDK is listed as coming soon, and no pricing is disclosed.