NeuReality

The problem: expensive accelerators left waiting

Inference is the stage in which a model that has already been trained actually runs and answers users. NeuReality started from the assumption that at this stage the bottleneck is not only the accelerator (the GPU or another dedicated chip) but the system around it: the host processor that manages the data, the network and the software. The company's CEO, Moshe Tanach, argued in 2024 that AI accelerators actually run at only about 30 to 40 percent utilization, and that adding more accelerators does not solve the system's inefficiency.

The founders

According to The Next Platform, the company was founded in September 2018 by Moshe Tanach, formerly an engineering director at Marvell and Intel; Tzvika Shmueli, formerly a vice president at Mellanox and Habana Labs; and Yossi Kasus, formerly a senior director at Mellanox and head of chip development at EZChip. Their shared background in communication chips and data centers explains the approach: to treat the AI chip as a component in the network, not as a separate computer.

NR1: a processor connected directly to the network

The first product, NR1, was defined by the company as a "network addressable processing unit" (NAPU). According to Calcalist, the chip was made on TSMC's 7-nanometer process and delivered in 2023, and according to The Next Platform it included eight Arm cores, signal processors and video engines, along with a hardware component the company calls the AI-Hypervisor, which manages the flow of data to the accelerators. The chip is sold as a card for an existing server and as a complete server. According to the company, the server cuts costs and power consumption by a factor of 50 compared with competing hardware, a claim that has not been independently tested.

Funding and partnerships

The company raised its first round in early 2020. In December 2022 it raised a $35 million Series A led by Samsung Ventures, Cardumen Capital, Varana Capital, OurCrowd and XT Hitech, with participation from memory maker SK Hynix, bringing total funding to $48 million. In March 2024 it raised another $20 million, partly from the investment fund of the European Innovation Council (EIC), and total funding rose to $70 million. In 2022 the company cited collaborations with IBM, AMD and Lenovo.

NR2 and NR-NEXUS: from the chip to the operating system

In 2025 the company announced NR2, a more powerful host processor based on advanced Arm cores. According to its plan, it will launch in 2026 and ship in volume in 2027. In March 2026 it introduced NR-NEXUS, which it describes as an operating system for "token factories": a software layer that manages inference workloads across processors, accelerators and network cards from any manufacturer. The product is aimed at new cloud providers, enterprises and chipmakers.

What the story shows

NeuReality started as a chip company and by 2026 had reached a product declared to be hardware-independent. The shift reflects a familiar pressure in the AI chip market: it is hard to compete with the giants' hardware, and easier to become the layer that makes better use of it. Whether the bet succeeds depends on its ability to prove its efficiency claims with customers, beyond the company's own announcements.