NeuReality brings on former Google AI product lead to push NR-NEXUS to market

NeuReality has appointed Shalini Agarwal as a strategic advisor to speed up the market launch of NR-NEXUS, its operating system for AI inference. For the Israeli startup, this is more than a staffing boost: the company is trying to claim a spot between expensive GPU clusters and enterprise customers who need a manageable layer on top of fragmented infrastructure.

Why an advisor

Agarwal is not taking an operational job; she joins as a senior strategic advisor. She previously led AI product initiatives at Google Cloud and Google Workspace, including the rollout of Gemini in Gmail, Docs, Slides and Sheets. For NeuReality it is an important signal to the market: the startup needs not only a strong engineering stack but also someone who can turn complex infrastructure technology into a clear offering for large customers, partners and hardware vendors.

"Enterprise AI is entering a new phase," says Agarwal.

The point of her appointment is go-to-market, not rewriting the architecture. NeuReality is already building the product with co-founder and CEO Moshe Tanach and president Hiren Majmudar, a former senior executive at GlobalFoundries and Intel Capital. Now the company has to prove that its orchestration layer is worth the integration effort, especially in a world where many customers are already deeply tied to the NVIDIA ecosystem and reluctant to add new infrastructure layers to their stack.

How NR-NEXUS works

The platform was introduced on March 12, 2026 as a hardware-neutral operating system for AI factories, or token factories. The idea is not to tie model deployment to a single type of hardware: NR-NEXUS runs on top of CPUs, GPUs and network infrastructure, and also supports more varied configurations with different accelerators. This approach is meant for companies that have already built clusters from heterogeneous hardware and do not want to rebuild everything for each new model or API.

According to NeuReality, the system handles orchestration of the full inference stack and helps distribute load across compute, memory and network. The company specifically highlights the prefill and decode stages, which can be spread more efficiently across different resources. In practice, this should stabilize performance, hold SLAs under load and squeeze more useful work out of accelerators that, in ordinary clusters, often sit idle part of the time in real-world operation.

• A single management layer for CPUs, GPUs and NICs
• Support for mixed infrastructure without a full architectural overhaul
• Routing of workloads between open-source and proprietary models and APIs
• Higher accelerator utilization and more predictable SLAs
• Lower cost of token generation as volumes grow

The software is already in use by beta customers, and a full commercial launch is expected later this year. The target audience for NR-NEXUS is fairly clear: neocloud providers, large companies building their own inference capacity, and chipmakers that need a ready-made software layer on top of their hardware. In all three scenarios NeuReality is selling neither a model nor a chip, but an infrastructure layer meant to simplify the launch of production AI services and shorten the time to deploy new models.

Why the timing fits

The bet targets the hottest part of the market. According to Deloitte, inference already accounted for roughly half of all AI compute in 2025, and in 2026 its share may grow to two-thirds. That also explains the surge in capital spending: Amazon is budgeting about $200 billion for 2026, and Google $175 billion to $185 billion. Even with budgets like these, enterprise customers still face an old problem: expensive hardware is often unevenly loaded, and the stack is assembled from too many incompatible components. NeuReality is trying to wedge itself into exactly that gap between hardware and operations.

The company has raised about $70 million, including a $35 million Series A round in late 2022 and another $20 million in March 2024 backed by the European Innovation Council Fund. Competition is already dense: Modal Labs, Baseten and Fireworks AI are all fighting for the inference-optimization market, each betting on who will become the main control layer after the model-training era.

What this means

NeuReality's story shows where the value in AI infrastructure is moving: from training models to running them day to day. If the startup can prove that NR-NEXUS genuinely raises cluster utilization and lowers the price per token without vendor lock-in, it has a chance to become a useful layer for enterprise customers who want to build AI services on hardware they have already bought. The winner here may not be whoever has the most GPUs, but whoever manages inference in production best.