Tensordyne

What Tensordyne does

Tensordyne (formerly Recogni, renamed September 2025) is a venture-backed AI chip and rack-systems company headquartered in Sunnyvale, California with a European regional office in Munich, Germany. The company designs and manufactures the Napier (TDN) AI inference system, which combines a 3nm custom AI processor (TDN AIP) implementing the patented Pareto logarithmic number system (TDN Math), a sub-microsecond any-to-any scale-up interconnect (TDN Link) embedded into the die, and standardized air-cooled compute trays (TDN ACT) packaged into 72-chip TDN72 pods and full TDN racks. The platform targets hyperscalers, neoclouds, sovereign AI infrastructure operators, and enterprise on-prem customers running large generative AI inference workloads, including large language models, image and video generation, and agentic coding applications. The company pivoted in 2024 from autonomous-vehicle perception (the original Scorpio chip platform) to datacenter AI inference, on the thesis that its log-math compute architecture yields disproportionate efficiency gains on transformer-based models.

Name

Tensordyne

Legal name

Tensordyne, Inc.

Website

https://tensordyne.ai

Company type

Private

Founded year

2017

Operating status

Operating

Headcount range

101–250 employees

Short description

Tensordyne designs and manufactures rack-scale AI inference systems built on proprietary logarithmic math (Pareto/LNS), 3nm custom silicon, and sub-microsecond scale-up interconnect, targeting hyperscalers, neoclouds, sovereign AI operators, and enterprise on-prem customers.

Ownership category

akta.pro rank

Product category

AI Inference Systems

NAICS

Computer Systems Design Services (541512)

SIC

Services-Computer Integrated Systems Design (7373)

akta.pro industry

PRIMARYHDAAAAAC

AI Inference Hardware & Edge Accelerators

Hardware, Technology & Telecom/Artificial Intelligence & Machine Learning/AI Infrastructure & Compute (GPUs/ASICs, accelerators)

SECONDARYHDAAABAF

Model Deployment, Serving & Inference Platforms

Hardware, Technology & Telecom/Artificial Intelligence & Machine Learning/ML Platforms, MLOps & Model Lifecycle Management

SECONDARYHDAAAAAI

AI Compiler, Runtime & Kernel Optimization Software (CUDA/ROCm/XLA, graph compilers)

Hardware, Technology & Telecom/Artificial Intelligence & Machine Learning/AI Infrastructure & Compute (GPUs/ASICs, accelerators)

SECONDARYHDAAAJAB

On-Device Inference Runtimes & SDKs (mobile/embedded)

Hardware, Technology & Telecom/Artificial Intelligence & Machine Learning/Edge AI & On-Device Machine Learning

Keywords

• AI inference systems
• AI accelerator chips
• Datacenter AI hardware
• Logarithmic AI compute
• AI inference silicon

Headquarters

HQ city

Sunnyvale

HQ country

United States

HQ region

North America

Offices2 records

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Technology or R&D, Personnel, Supply Chain, Operations, Marketing or Sales, Infrastructure

Revenue model

1. Hardware Sales (Inference Systems / Chips): Pre-revenue commercial launch: Tensordyne has secured more than $200 million in Letters of Intent from hyperscaler, neocloud, and sovereign AI infrastructure operators for its TDN72 inference pod and full Napier rack. Anticipated to convert LOIs into bookings tied to beta evaluation, with Series D funding later in 2026 to scale manufacturing and shipments.
2. Early Access / Evaluation System Purchases: Strategic partners such as DataVolt have agreed to purchase inference systems for evaluation prior to production, representing a pre-production evaluation revenue stream ahead of broader commercial deployment.

Pricing tiers

Go-to-market motion1 record

Distribution channels3 records

Marketing channels8 records

Core offering

Tensordyne designs, manufactures, and sells the Napier (TDN) AI inference system — a rack-scale datacenter platform combining the proprietary Pareto logarithmic number system, a custom 3nm AI processor (TDN AIP), tightly integrated HBM3e plus SRAM memory, and an embedded sub-microsecond any-to-any scale-up interconnect (TDN Link). The system is delivered as air-cooled TDN72 inference pods and full TDN racks for hyperscalers, neoclouds, sovereign AI infrastructure operators, and large enterprise on-prem deployments.

Product overview

Tensordyne offers a single unified inference platform — Tensordyne Napier (TDN) — composed of the TDN AIP 3nm AI processor, TDN Math (Pareto logarithmic number system), TDN Link scale-up interconnect, TDN ACT compute tray, the TDN72 inference pod, and the full TDN Rack (four TDN72 pods). The platform targets hyperscalers, neoclouds, sovereign AI, and enterprise on-prem with air-cooled disaggregated inference and is supported by a Python-inspired SDK (Coming Soon), Hugging Face model zoo, and a Token Economics Calculator tool for cost and energy comparison. Legacy vision-cognition products (Recogni Scorpio) preceded the platform before the 2024-2025 pivot and 2025 rebrand to Tensordyne.

Differentiator

Problem solved

Functional benefit

Brands

• Tensordyne Napier (TDN): Flagship AI inference system product line, including the TDN72 inference pod and full TDN rack systems built on logarithmic math and the Napier 3nm AI processor (TDN AIP).

Products and services

• Tensordyne Napier (TDN) AI Inference System The flagship air-cooled, fully disaggregated rack-scale AI inference system combining the TDN72 inference pod (72 chips per pod), TDN Rack (four pods per rack), TDN AIP 3nm AI processor, TDN Math (Pareto logarithmic mathematics), and TDN Link sub-microsecond scale-up interconnect, delivering up to 608 PFLOPS of dense compute per rack for hyperscalers, neoclouds, sovereign AI, and enterprise on-prem customers.
• Token Economics Calculator A web-based calculator that compares AI inference systems across tokens per second, tokens per USD, and tokens per kWh, helping companies deploying LLMs evaluate cost and energy trade-offs across different hardware and models using publicly available MLPerf data.

Quantifiable outcome

• 17x more tokens per watt versus NVIDIA Blackwell systems
• +7 more outcomes

Named customers3 records

Segments5 records

Ideal customer profiles5 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

Integration6 records

AI capability11 records

Feature6 records

Partnerships

Five partnerships are on record, tiered minor, flagship and major.

• Nara CommunicationsminorGTM or Marketing Partner · 15 June 2026Media/PR agency handling press inquiries for Tensordyne; named contacts Frankie McGovern and Yeona Choi.
• TSMC (Taiwan Semiconductor Manufacturing Company)flagshipOEM/ Whitelabel/ Licensing Partner · 1 June 2026TSMC manufactures Tensordyne's Napier 3nm AI inference chips. Tensordyne has completed tape-out and is transitioning into High-Volume Manufacturing (HVM) at TSMC facilities to meet demand for datacenter AI infrastructure. Previously manufactured the first-generation 7nm Scorpio chip.
• DataVoltmajorChannel Partner/ Reseller/ Distributor · 28 May 2025Strategic partnership to build the most energy-efficient AI cloud. DataVolt has agreed to purchase Recogni/Tensordyne's inference systems for evaluation prior to production. Leverages DataVolt's low-cost energy and global reach (Saudi Arabia HQ with offices in USA, Uzbekistan, India, UAE, presence in South Korea and South Africa). DataVolt signed a $20B agreement at the U.S.–Saudi Investment Forum.
• BroadcomflagshipStrategic or Co-development Partner · 1 November 2024Broadcom is a development partner on the Napier 3nm AI Inference Accelerator. Frank Ostojic (SVP & GM, ASIC Products Division at Broadcom) publicly endorses the partnership: 'They are leveraging Broadcom's fundamental IP Platform, advanced Silicon and Packaging design technologies to bring their 3nm AI Inference Accelerator to market.' Broadcom also co-developed the first-generation 7nm Scorpio chip.
• HPE Juniper NetworksflagshipStrategic or Co-development Partner · 1 November 2024Juniper Networks (now HPE) participated in the $102M Series C funding round and entered a strategic technology collaboration to integrate industry-leading scale-up networking components directly with the Tensordyne platform, enabling massive fabric scalability required for modern LLM clusters.

Scale indicators8 records

Recent moves6 records

Expansion highlights6 records

Broad incumbents

• NVIDIA: Dominant GPU and AI accelerator incumbent whose Blackwell (B200/GB300) and NVL72 systems are the explicit benchmarks Tensordyne compares against on tokens/watt and tokens/sec. NVIDIA defines the software ecosystem (CUDA, Triton) and hyperscaler procurement patterns that Tensordyne must displace.
• Intel (Habana Labs): Intel's Habana unit ships Gaudi accelerators for AI training and inference, sold as part of Intel's broader data center portfolio. A broad incumbent competitor that overlaps with Tensordyne in hyperscaler and enterprise inference procurement.

Direct peers

• Groq: AI inference-focused chip company offering LPU-based systems targeting low-latency, high-throughput LLM serving for hyperscalers and enterprises. Closely aligned with Tensordyne's inference-first, low-latency, disaggregation-friendly thesis and customer base.
• Cerebras Systems: AI accelerator company building wafer-scale chips for training and inference, selling rack-scale systems to hyperscalers and enterprises. Competes head-to-head with Tensordyne in the high-end AI inference hardware market with a similar performance-per-watt and total-cost-of-ownership narrative.
• SambaNova Systems: AI chip startup offering reconfigurable dataflow units (RDUs) packaged as rack-scale systems for enterprise and government AI inference and training. Targets the same enterprise on-prem and sovereign AI buyer segments as Tensordyne's Napier.
• Graphcore (SoftBank): UK-based AI accelerator (IPU) company now owned by SoftBank, selling Bow-class systems for AI training and inference to enterprises and hyperscalers. Competes with Tensordyne on European AI infrastructure deals and inference-focused value propositions.

Emerging players

• Tenstorrent: AI accelerator startup (backed by Hyundai, Samsung, Bezos Expeditions) building RISC-V-based chiplet architectures for training and inference at hyperscaler and enterprise. Competes on open-architecture and per-watt economics for AI inference.
• Etched: AI inference chip startup building a transformer-specific ASIC (Sohu) that claims large inference throughput advantages over GPUs. Targets the same hyperscaler and frontier-model customer base as Tensordyne with an architectural specialization angle.
• Lightmatter: Photonics-based AI compute and interconnect startup whose products include inference accelerators (Envise) and photonic scale-up fabrics (Passage). Direct adjacency to Tensordyne on the silicon-plus-interconnect inference stack thesis.

Others

• Marvell Technology: Custom silicon (ASIC) and interconnect provider that builds AI training/inference accelerators for hyperscalers in partnership with customers like Amazon (Trainium/Inferentia). Comparable to Tensordyne in targeting custom-silicon, scale-up networking, and rack-level inference systems for hyperscaler buyers, often via Broadcom as the enabler.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat6 records

Key risks7 records

Key highlights7 records

Customer concentration

• Xx.com/TensordyneInc
• YouTubeyoutube.com/@TensordyneInc
• LinkedInlinkedin.com/company/tensordyne
• Instagraminstagram.com/tensordyne

Compliance3 records

Revenue estimate

Valuation estimate

Number of profiles

Profiles17 records

Subsidiaries1 record

Funding overview

Funding rounds6 records

Investors21 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Tensordyne