Nvidia plans to expand R&D in Israel’s south, hire hundreds by 2026

Nvidia will expand its R&D presence in Be’er Sheva, Israel, by relocating to a new 3,000-square-meter site and planning to hire hundreds of engineers and developers by 2026.

Nvidia’s new site is located in Be’er Sheva’s high-tech park, which is home to companies like Microsoft and Elbit Systems, as well as Israeli defense units.

The move will triple the US chipmaker’s footprint in the southern city, where it first established operations after acquiring Mellanox for US$7 billion in 2020.

The company employs over 5,000 people in Israel, making it one of the country’s largest private tech employers.

Nvidia is also reviewing proposals for a new mega-campus in northern Israel and recently enlarged its Tel Aviv offices.

Israel is Nvidia’s second-largest R&D hub outside the US, with several key products developed locally.

🔗 Source: Calcalist

🧠 Food for thought

Implications, context, and why it matters.

Israel’s role in NVIDIA networking and infrastructure

NVIDIA’s AI networking spans BlueField-3 SuperNICs (high-performance network interface cards that offload data processing) and Spectrum-4 Ethernet switches (switches designed for AI data centers) that power Spectrum-X (an Ethernet-based network design optimized for AI) 1.
Israel-1, an NVIDIA AI system with 2,048 H100 GPUs and 2,560 BlueField-3 SuperNICs tied through 80 Spectrum-4 switches, proved Ethernet can run large AI jobs 1 2. Network throughput rose by 1.6x in that setup.
A $500 million, 30-megawatt R&D facility is under construction in northern Israel with Blackwell GPUs (NVIDIA’s next-generation architecture) plus next-generation networking gear 3. The project comes from NVIDIA and cements the country’s role.

Enterprise networking vendors can partner on Spectrum-X storage acceleration

DDN (DataDirect Networks, a high-performance storage vendor), VAST Data (an AI data infrastructure company) and WEKA (a parallel file system provider) integrated Spectrum-X 4. Read bandwidth rose up to 48% and write bandwidth up to 41% versus RoCE v2 (Remote Direct Memory Access over Converged Ethernet, version 2). Checkpoint saving (periodically saving model state during training) and RAG inference (Retrieval-Augmented Generation) improved.
Vendors can stand out with Spectrum-X adaptive routing plus congestion control to move multi-terabyte elephant flows (very large, long-lived data transfers) without crushing links 4.
Solution providers that target AI data centers should track NVIDIA DOCA (Data Center on a Chip Architecture) progress 5. The framework now supports up to 2,000 VirtIO-net devices (a standard virtual network interface used by hypervisors) per BlueField-3 DPU (Data Processing Unit) and offers encryption offloads that cut CPU overhead in multi-tenant environments.