NeuReality

NeuReality is an AI inference infrastructure company founded in 2019 in Israel that develops purpose-built silicon, systems and software intended to raise the utilization of AI accelerators in data centers and on-premises deployments. Its first-generation product line centers on the NR1 chip, described as a 7nm "AI-CPU" and Network Addressable Processing Unit (NAPU) that replaces conventional host CPU and NIC architecture, combining an ARM-based CPU, media/DSP processors, integrated network functions, a neural network engine and hardware-based AI-Hypervisor IP on a single die. The chip is packaged as the NR1-M module, a full-height double-wide PCIe card, and inside the NR1-S appliance, which pairs several NR1 modules with third-party AI accelerators in 1:1, 1:2 or 1:4 configurations and has been demonstrated with Qualcomm, AMD and IBM accelerators.

Alongside the hardware, the company offers a full-stack software layer organized into model, pipeline and service layers, with an SDK exposing three APIs (toolchain, provisioning and inference) covering development, deployment and serving of AI pipelines. The stack targets generative AI, agentic AI, RAG, computer vision, natural language processing and speech recognition workloads, runs on ARM server-class standard Linux, integrates with Kubernetes orchestration, NFS and S3 storage, and connects to common AI frameworks and MLOps backends. More recently the company has positioned its portfolio around "token factories": NR-NEXUS, an inference operating system providing orchestration, routing, observability and policy controls across heterogeneous GPU and XPU environments, and the NR2 AI-SuperNIC, scale-out networking silicon specified at 1.6 Tbps throughput with Ultra Ethernet Consortium support and in-network compute for low-latency east-west data movement in distributed clusters.

Founding story

NeuReality was founded in 2019 by Moshe Tanach, Tzvika Shmueli and Yossi Kasus. Tanach had been a director of engineering at Marvell and Intel and AVP R&D at DesignArt-Networks (acquired by Qualcomm); Shmueli was VP of back-end infrastructure at Mellanox Technologies and VP of engineering at Habana Labs; Kasus was senior director of engineering at Mellanox and head of VLSI at EZChip. CTO Lior Khermosh, previously co-founder and chief scientist of ParallelM and a fellow at PMC Sierra, is also part of the leadership team.

Business model

NeuReality sells inference infrastructure as a combination of silicon (NR1 AI-CPU, NR2 AI-SuperNIC), integrated modules and appliances, and an accompanying software stack and APIs. For NR-NEXUS it offers three commercial paths: a serverless API billed per token consumed, a fully managed dedicated deployment with enterprise SLAs, and an annual license for customers running managed or owned infrastructure on private capacity.

Stated deployment paths include pay-per-token consumption on a serverless API, managed dedicated capacity with enterprise SLAs, and annual software licensing for self-owned or managed infrastructure, in addition to sales of NR1 modules and appliances.

Traction

First cloud computing and financial services customers were reported to be running NR1 appliances on site, showing up to 6x better total cost per AI mega-token versus traditional CPU/GPU setups. The NR1-S appliance has been publicly demonstrated with Qualcomm, AMD and IBM accelerators, and the company announced collaborations with IBM, AMD and Lenovo. Cumulative disclosed funding reached $70 million as of March 2024.

Latest developments

The company has shifted its public positioning toward "token factories," unveiling the NR-NEXUS inference operating system and the NR2 AI-SuperNIC scale-out networking chip, publishing a white paper on AI networking bottlenecks, and appointing a former Google AI director to take the inference operating system to market. An October 2025 technical interview with CEO Moshe Tanach outlined the move from the first-generation NR1 AI-CPU to the NR2 AI-SuperNIC for east-west scale-out communication in large disaggregated GPU clusters.