Skip to main content

Tensordyne

Category: AI Chips / Semiconductors

AI inference chip and rack-scale system company building the Napier platform with proprietary logarithmic math to deliver high-speed, energy-efficient AI inference as an alternative to Nvidia. Tensordyne was founded in 2017. The company is led by Marc Bolitho. Based in Sunnyvale, California, United States. Team size: 101-500. Total funding raised: $176M. Latest round: Series C. Key investors include Celesta Capital, GreatPoint Ventures, Juniper Networks, Mayfield, WRVI Capital, DNS Capital, BMW i Ventures, HSBC Innovation Banking, Pledge Ventures, Tasaru Mobility Investments (PIF).

Founded
2017
Headquarters
Sunnyvale, California, United States
Team size
101-500
Total funding
$176M

Value proposition

Tensordyne's Napier platform uses a proprietary logarithmic number system (Pareto) that replaces multiplication with addition, enabling 17x more tokens per watt and 13x higher throughput than Nvidia Blackwell systems, while being fully air-cooled and fitting in standard data center racks.

Products and solutions

Tensordyne Napier (TDN) AI inference system, TDN AIP (Artificial Intelligence Processor) — 3nm chip with 144GB HBM, TDN Math — proprietary logarithmic number system, TDN Link — any-to-any scale-up interconnect with sub-microsecond latency, TDN72 — 72-chip inference pod, Full rack system combining 4 TDN72 pods delivering 608 PFLOPS FP8 dense compute

Unique value

First company to successfully commercialize logarithmic math for AI inference at scale — turning multipliers into adders on silicon to dramatically reduce power consumption and cost while increasing inference speed. The TDN72 pod outperforms a full Nvidia NVL72 rack while consuming substantially less power.

Target customer

Hyperscalers (large-scale AI inference factories), neoclouds (AI cloud service providers), and enterprise AI operators (on-premise AI infrastructure)

Industries served

AI Infrastructure / Data Center, Cloud Computing, Enterprise AI, Sovereign AI

Technology advantage

Logarithmic math architecture (Pareto number system) patented since 2019; 3nm TSMC process with Broadcom partnership; HPE Juniper Networks scale-up networking integration; 17x tokens/watt vs Nvidia GB300; 13x throughput vs Nvidia Blackwell; 608 PFLOPS FP8 per rack; sub-microsecond scale-up interconnect

How they differentiate

Proprietary logarithmic math (Pareto number system) that replaces multiplication with addition, enabling dramatically better power efficiency. Full-stack co-design from math through silicon, networking, and system software. Single-rack solution that handles both prefill and decode stages of LLM inference, eliminating the need for multi-vendor, multi-rack setups. 100% air-cooled at 30kW per pod — no liquid cooling required.

Main competitors

Nvidia (GB300/B300, Blackwell), Groq (LPX inference systems), Cerebras (wafer-scale inference)

Key partnerships

Broadcom (chip design partnership, 3nm tapeout), HPE Juniper Networks (scale-up networking), TSMC (3nm manufacturing), Cirrascale Cloud Services (infrastructure partner), BlueSky Compute (infrastructure partner)

Notable customers

Cirrascale Cloud Services, BlueSky Compute, over a dozen letters of intent from hyperscalers, neoclouds, and sovereign AI operators

Major milestones

2017: Founded as Recogni, 2019: Filed first 5 foundational patents for log-math architecture, 2019: $25M Series A, 2021: $48.9M Series B, 2022: First-gen Scorpio 7nm chip taped out with Broadcom/TSMC, 2024: $102M Series C, 2024: Strategic pivot from autonomous vehicles to datacenter AI inference, 2025: Rebranded from Recogni to Tensordyne, 2026: Napier 3nm chip tapeout with Broadcom/TSMC, 2026: $200M+ forecasted demand for Napier systems, 2026: Preparing Series D

Growth metrics

Forecasted $200M+ in orders for Napier inference system; over a dozen LOIs from evaluation partners; preparing Series D funding round

Market positioning

Direct challenger to Nvidia in the AI inference market, targeting the growing demand for power-efficient alternatives. Positioned as a single-rack solution that can replace multi-rack Nvidia+Groq or AWS Trainium+Cerebras setups for large model inference.

Geographic focus

North America (Sunnyvale, CA HQ) and Europe (Munich, Germany office)

Patents and IP

Filed first five foundational patents for log-math-based architectures in 2019. Proprietary logarithmic number system (Pareto) for AI compute.

About Marc Bolitho

Former Senior Vice President and General Manager, Electronics and ADAS business unit at ZF Group (nearly 30 years automotive engineering experience). Appointed CEO of Recogni (now Tensordyne) in July 2022.

Official website: