NeuReality and the EIC Accelerator: A Chip for AI Inference

NeuReality, founded in 2019 in Caesarea, Israel, by veterans of Mellanox, Habana Labs, Intel and Marvell, designed NR1, a "server on chip" that takes the conventional processor out of the data path of AI inference servers so that the accelerators doing the work stay busy. It was selected in the EIC Accelerator's March 2023 cut-off with blended finance and received an EU grant of €2.50 million for a €7.25 million project that ran from July 2023 to June 2025, according to CORDIS. The goal was to finalise NR1 and "validate the optimized chip in operational settings for further commercialisation".

The final report says NR1 was "fully integrated and validated", booted Linux and ran inference at target performance, and that NR1 was repositioned as a turnkey "AI Inference Appliance" for large language models, according to the CORDIS reporting page. It reports 22,000 images per second on the ResNet50 benchmark per module, scaling to 220,000 on ten, integration with Qualcomm accelerators, and proof-of-concept deployments at Cirrascale, Fidelity and Rebellions. No commercial orders for NR1 have been announced.

Where NeuReality Stands Now

NeuReality's website in September 2026 leads with NR-NEXUS, an "inference operating system for token factories" launched in March 2026, and the NR2 AI-SuperNIC, a networking chip for GPU clusters unveiled in September 2025, according to its press page. NR-NEXUS is described as hardware-agnostic across CPUs, accelerators and network cards, which means it does not need NR1. In August 2025 the company wrote that it would build on NR1 "as we now begin to introduce our next generation NR2 chip". Revenue and customer numbers are not public.

The Technology: Taking the CPU Out of the Path

In a standard AI server, a general-purpose processor handles networking, data preparation and scheduling before work reaches the accelerator, and it can become the bottleneck. NR1 moves those tasks into dedicated hardware engines, with its own network interface and Arm cores. NeuReality claims this raises accelerator utilisation from about 30% to 100% and cuts cost and energy per query. The figures come from the company; no independent benchmark has been published.

What the EIC Award Promised, and What the Record Shows

The award promised to finalise NR1 and validate it in operational settings. The report shows a validated chip, a software stack for large language models, and pilots with a cloud provider, a financial company and a chip maker. That meets the project's stated goal. Commercial traction beyond pilots is not documented, and within a year the company's public focus moved to other products.

1. NR1 chip validationDeliveredFull-speed tests, per report.
2. Software stackDeliveredLLM, PyTorch, Kubernetes.
3. Operational pilotsDeliveredCirrascale, Fidelity, Rebellions.
4. Commercial salesUnverifiedNone announced.

The Money

The EU paid only 34% of this €7.25 million project. In March 2024 NeuReality raised $20 million from investors including the EIC Fund, bringing its total to $70 million, according to DCD. No later round has been announced. Designing a second chip is expensive, and how NR2 is being financed is not public.

What Cannot Be Checked

Why Europe Should Care

Running AI models at scale is expensive, and NeuReality's pitch was to make it cheaper and less power-hungry. The EIC backed it through grant and equity. The final report says the chip was validated, and the company's public focus has since moved towards software and networking.

The Verdict

NeuReality and the EIC have not been asked for comment for this article. It is based entirely on public records and published statements, linked throughout.