Tensordyne

Market position

Tensordyne competes in datacenter AI inference silicon and systems, positioning its log-math architecture against GPU-based inference infrastructure on cost, energy efficiency and latency.

The company differentiates on logarithmic math-based compute and its scale-up interconnect, claiming higher energy efficiency than GPU architectures, air cooling instead of liquid cooling, and cited figures including 608 PFLOPS of dense rack compute, over 1,000 tokens per second per user, and 2x the speed of leading solutions on large MoE models, while retaining full 16-bit precision.

Technology

The architecture is based on logarithmic mathematics for AI compute, protected by patents first filed in 2019, combined with a low-latency, any-to-any cell-based scale-up interconnect (TDN LINK) enabling linear scaling. Napier chips are fabricated on a 3nm process with Broadcom as design partner and TSMC as foundry; the earlier Scorpio generation used a 7nm process. The system is disaggregated to reduce bottlenecks in multi-trillion-parameter model serving, air-cooled at 30 kW per pod, supports 16-bit precision inference, and integrates with PyTorch, Triton and vLLM.

Go-to-market

Direct engagement via the company website, with named product and sales contacts offering booked calls for different customer segments, and a planned technical whitepaper on the Napier system.

Hyperscalers building large-scale inference deployments; "neo cloud" providers serving high-margin, low-latency inference; enterprises running frontier models on-premises; and application-layer companies and foundation model builders.

Geography

Dual presence in North America and Europe, with a US office at 1195 Bordeaux Dr, Sunnyvale, CA 94089 and a German office at Brienner Strasse 59, Munich 80333.

History

The company filed its first five foundational patents on log-math-based architectures in 2019 and began designing Scorpio, its first-generation chip. Scorpio was taped out in 2021 with Broadcom, with functional 7nm silicon received from TSMC, validating the log-math architecture in hardware. In 2022 functional Scorpio samples shipped to initial validation customers, including a development system performing surround-view object detection. During 2023 the company extended support to transformer-based models and identified opportunities in generative AI workloads such as GPT and Stable Diffusion. In 2024 it halted its legacy vision tracks and refocused entirely on Napier, a datacenter AI inference chip, and entered a technology partnership with Juniper Networks to integrate scale-up networking components. Through 2025 the company optimized silicon, systems and software for Mixture-of-Experts models and agentic workloads, citing industry shifts such as DeepSeek-R1. In 2026 it completed tapeout of a 3nm Napier chip with Broadcom and TSMC and stated that Napier silicon was transitioning into high-volume manufacturing at TSMC.