What Positron does

Positron is a US-based AI infrastructure company that designs and supplies purpose-built hardware and full-stack software for transformer-based model inference. The product line is vertically integrated: the shipping Atlas inference appliance combines 8x Positron Archer FPGA accelerators (built on Altera Agilex) with 256 GB HBM, AMD EPYC Genoa CPUs, and the Positron Inference Engine; the announced Asimov custom silicon (tape-out late 2026, production early 2027) introduces a memory-first LPDDR5x architecture with up to 2.3 TB per chip, a proprietary TransWarp systolic engine, and PCIe Gen 6 + CXL host connectivity; and the announced Titan system (2027) integrates 4x Asimov chips into an air-cooled 4U chassis scaling to 4,096 systems per cluster and supporting up to 16 trillion parameters per server. The software stack exposes an OpenAI API-compliant endpoint (api.positron.ai) and natively maps any trained HuggingFace Transformers model onto Positron hardware without a new compiler, with use cases spanning code generation, long-context reasoning, and next-generation multimodal workloads.

Positron generates revenue through sales of inference appliances and, increasingly, metered access to its OpenAI-compatible inference API. The company positions Atlas at 3.07x perf/dollar and 4.52x perf/watt versus the NVIDIA DGX H200 on Llama 3.1 8B, and Asimov at 5x tokens/dollar and 5x tokens/watt versus NVIDIA Rubin, targeting enterprise, sovereign, and hyperscaler buyers seeking to diversify away from a single-vendor inference stack. Go-to-market is anchored in a "made-in-America" supply chain narrative (Cadence/TSMC 3nm/2nm silicon path, US-based support) and a high-profile partner ecosystem spanning Arm (AGI CPU launch partner), Altera, Cadence, and TSMC, with the February 2026 $230M Series B at a $1B+ post-money valuation providing the capital base for Atlas scaling and Asimov bring-up. International expansion began with a DIFC-licensed Dubai office targeting the Middle East, Africa, and South Asia region.

The company is headquartered in Reno, Nevada, employs 11-50 people, and is led by CEO Mitesh Agrawal, with founder Thomas Sohmers now in a CTO-equivalent capacity. As of the February 2026 Series B, institutional backers include ARENA Private Wealth, Jump Trading, Unless (co-leads), Qatar Investment Authority, Arm Holdings, Helena, Atreides Management, DFJ Growth, Valor Equity Partners, 1517 Fund, Flume Ventures, Resilience Reserve, and earlier-stage investors Banyan Ventures and Oakseed Ventures.

Name

Positron

Legal name

Positron AI

Website

https://positron.ai

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

101–250 employees

Short description

Positron is a US-based AI infrastructure company that designs custom silicon (Asimov) and FPGA-based inference appliances (Atlas) purpose-built for transformer model inference, exposing an OpenAI-compatible API and HuggingFace-native software stack to compete with NVIDIA on cost and power efficiency.

Ownership category

akta.pro rank

Product category

AI Inference Hardware

NAICS

Computer Systems Design Services (541512), Computer Systems Design and Related Services (54151)

SIC

Services-Computer Integrated Systems Design (7373), Services-Prepackaged Software (7372)

akta.pro primary industry

AI Accelerators (GPUs/TPUs/NPUs/ASICs) (HDAAAAAA)

akta.pro secondary industry

Model Deployment, Serving & Inference Platforms (HDAAABAF)

Keywords

• AI inference hardware
• Transformer accelerators
• AI inference servers
• AI accelerator silicon
• Machine learning inference

Headquarters

HQ city

Reno

HQ country

United States

HQ region

North America

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

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

Core offering

Positron designs and supplies purpose-built AI inference hardware and software for transformer-based models. Its product stack includes the shipping Atlas FPGA-based inference appliance (8x Positron Archer accelerators), the announced Asimov custom AI inference accelerator silicon, the announced Titan 4U inference system, and the Positron Inference Engine software with an OpenAI API-compatible endpoint that runs any HuggingFace Transformers model directly on the hardware.

Product overview

Positron offers a vertically integrated AI inference platform combining custom silicon, purpose-built systems, and full-stack software, delivered as a platform-plus-modules architecture. The currently shipping Atlas inference appliance (8x Positron Archer FPGA accelerators, made-in-America) is the production tier and runs the Positron Inference Engine with HuggingFace Transformers support and an OpenAI-compatible API. The next-generation Asimov custom silicon (LPDDR5x memory-first architecture, 2027) powers the flagship Titan inference system (4x Asimov chips, 8+ TB memory, scales to 4,096 nodes per cluster) for multi-trillion-parameter and million-token-context workloads. The whole stack is designed to compete with NVIDIA on tokens-per-dollar and tokens-per-watt for generative AI inference.

Differentiator

Problem solved

Functional benefit

Products and services

• Atlas Transformer Inference Server Atlas is a production-ready 4U Transformer Inference Server for enterprise AI/ML teams running large language and other transformer models. It ships with 8x Positron Archer FPGA accelerators, 256 GB HBM plus 384 GB to 2 TB system DDR5, dual AMD EPYC Genoa CPUs, and the Positron Inference Engine, and is positioned as a domestic, energy-efficient alternative to NVIDIA Hopper-class GPU servers for inference workloads up to 500B parameters.
• Positron Inference Engine & OpenAI API-compatible Endpoint The Positron Inference Engine is a software platform that runs any trained HuggingFace Transformers model directly on Positron hardware, supported by the Positron Model Manager for upload/management of .pt and .safetensors model files. It exposes an OpenAI API-compliant chat completions endpoint at api.positron.ai, allowing customers to switch clients via the OpenAI Python SDK to run inference on Positron hardware with no model conversion or new compiler.

Ideal customer profiles1 record

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration2 records

AI capability7 records

Recent moves8 records

Expansion highlights6 records

Direct peers

• Groq: Groq builds LPU inference accelerators purpose-built for low-latency generative AI inference, the same niche Positron targets. Both compete on tokens-per-dollar and tokens-per-watt against NVIDIA GPUs for transformer model serving.
• SambaNova Systems: SambaNova develops custom AI accelerator chips (RDU) and integrated systems for enterprise generative AI inference and training. Like Positron, SambaNova pitches vertically integrated silicon-plus-software for transformer workloads with a focus on data-sovereign, on-prem deployment.
• d-Matrix: d-Matrix builds purpose-built inference accelerators (NPU) optimized for transformer and generative AI workloads, with a memory-centric architecture similar to Positron's design philosophy. Both target hyperscale LLM serving economics.

Broad incumbents

• Cerebras Systems: Cerebras is a more established AI silicon company whose wafer-scale systems serve both training and inference for large models. It shares Positron's positioning as an NVIDIA alternative with co-designed hardware and software, and is a fellow Arm AGI CPU launch partner.
• NVIDIA: NVIDIA is the dominant incumbent that Positron benchmarks against (Hopper, Rubin). It offers the broadest AI accelerator portfolio (H100, H200, B200, GB200, future Rubin) with the deeply entrenched CUDA software ecosystem that Positron's inference engine is explicitly designed to bypass.

Emerging players

• Tenstorrent: Tenstorrent develops RISC-V-based AI accelerators for training and inference, with an open architecture strategy. While broader than Positron's inference-only focus, both target NVIDIA displacement with custom silicon and developer-friendly software stacks.
• Rebellions: Rebellions designs AI inference accelerators (e.g., Rebel and ATOM) for data center and edge deployments. It is also a launch partner for Arm's AGI CPU alongside Positron, making it a directly comparable emerging Asian-headquartered inference silicon player.
• Etched: Etched develops a transformer-specific ASIC (Sohu) that hardcodes transformer attention into silicon for ultra-high inference throughput. While architecturally distinct from Positron's general-purpose inference chip, it competes directly in the 'inference-only accelerator' category Positron targets.
• FuriosaAI: FuriosaAI builds NPU-based accelerators (RNGD) for data center LLM inference. Like Positron, it pitches performance-per-watt advantages over NVIDIA and targets hyperscale and enterprise inference workloads with custom silicon.
• Lightmatter: Lightmatter develops photonic AI compute systems (Envise, Passage) for inference and interconnect. While using a fundamentally different compute substrate (photonic vs CMOS), it competes in the same inference accelerator market segment and pursues similar hyperscaler customers.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat5 records

Key risks5 records

Key highlights6 records

Customer concentration

• Xx.com/positron_ai
• LinkedInlinkedin.com/company/positron-ai

Revenue estimate

Valuation estimate

Number of profiles

Profiles2 records

Funding overview

Funding rounds5 records

Investors27 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 Positron