The Israeli AI Silicon Stack: Annapurna, Habana, Mellanox, Run:ai, Hailo
Annapurna (Amazon), Habana (Intel), Mellanox and Run:ai (Nvidia), and Hailo together constitute the Israeli AI silicon-and-infrastructure cluster. The most consequential national silicon footprint outside the US-Taiwan-China triad.
The Israeli AI silicon stack is the most consequential national silicon-and-infrastructure cluster outside the US-Taiwan-China triad. Annapurna Labs (Amazon), Habana Labs (Intel), Mellanox (Nvidia), Run:ai (Nvidia), and Hailo together constitute an Israeli engineering footprint that designs and orchestrates a substantial share of the silicon and networking infrastructure underpinning global AI deployment.
The structural feature is that four of the five named companies have been acquired by major US semiconductor or hyperscaler platforms. The Israeli silicon-and-AI-infrastructure engineering bench is now distributed across Amazon, Intel, Nvidia, and — in the case of Hailo — the remaining independent edge-AI cohort. The institutional concentration matters: the AI hyperscalers run their chip and infrastructure roadmaps substantially through Israeli R&D.
Annapurna Labs (Amazon)
Annapurna Labs, founded in 2011 by Nafea Bshara and Hrvoye (Billy) Bilic in Israel, was acquired by Amazon in January 2015 for approximately $370 million. The acquisition has produced one of the most strategically consequential silicon outcomes in cloud computing history.
Annapurna's Tel Aviv and US engineering teams design the AWS Graviton CPU family (Graviton 1, 2, 3, 4 across the 2018-2024 product cycle) and the AWS Nitro System silicon underpinning the broader EC2 architecture. AWS's ability to operate the world's largest cloud-computing infrastructure at competitive cost economics is materially dependent on the Annapurna silicon. The Israeli engineering footprint inside Amazon now numbers in the multiple thousands.
Habana Labs (Intel)
Habana Labs, founded in 2016 by David Dahan and Ran Halutz, was acquired by Intel in December 2019 for approximately $2 billion. The company's Gaudi AI accelerator series — Gaudi 1, 2, and 3 — represents Intel's most strategically important AI hardware franchise.
Gaudi 3, announced in 2024 and ramping through 2025-2026, positions Intel as the credible third-vendor option in the AI training and inference accelerator market alongside Nvidia and AMD. The Habana engineering footprint in Caesarea is one of the largest Israeli AI-silicon design operations and the institutional anchor of Intel's broader Israeli engineering presence.
Mellanox (Nvidia)
Mellanox Technologies, founded in 1999 by Eyal Waldman and others, was acquired by Nvidia in 2020 for approximately $6.9 billion. The acquisition gave Nvidia ownership of the InfiniBand networking architecture and the Spectrum Ethernet portfolio — the networking infrastructure underpinning modern AI training cluster deployments.
Nvidia's competitive position in AI training infrastructure depends substantially on the Mellanox-derived networking stack. The Yokneam, Tel Aviv, and broader Israeli Mellanox engineering footprint operates inside Nvidia as the institutional anchor of Nvidia's broader Israeli presence — which extends into the 2025-2026 Nvidia Israeli expansion across Tel Aviv and the Kiryat Tivon planning.
Run:ai (Nvidia)
Run:ai, founded in 2018 by Omri Geller and Ronen Dar, was acquired by Nvidia in 2024 for a reported approximately $700 million. The company's GPU orchestration software has become a strategic component of Nvidia's broader AI infrastructure software stack.
The acquisition is significant beyond the deal value because it positions Nvidia's software stack to manage GPU orchestration across hyperscaler and enterprise customer environments — a category that historically had multiple credible competitors. Run:ai's Israeli engineering team integrates into the broader Nvidia Israel architecture, alongside the Mellanox-derived networking team and Nvidia's other Israeli engineering footprint.
Hailo: the independent edge-AI position
Hailo, founded in 2017 by Orr Danon and Hadar Zeitlin, operates the most institutionally substantial independent Israeli AI silicon company. The company's Hailo-8 and Hailo-15 edge AI processors target the embedded and edge AI inference market — automotive, industrial, security, and consumer-device applications where running AI inference at the edge rather than the cloud carries cost and latency advantages.
Hailo's strategic position is differentiated from the four acquired companies above: edge AI, embedded design wins, and a customer base concentrated in OEM partnerships rather than hyperscaler-direct relationships. The independent position has been preserved through 2025-2026, with capital and customer commitments supporting continued standalone operation.
What the silicon stack means institutionally
The Israeli AI silicon-and-infrastructure cluster operates a distinctive institutional pattern: deep technical capability concentrated in a small number of companies, the majority of which have been acquired by major US hyperscaler or semiconductor platforms, with the Israeli engineering footprint preserved post-acquisition as the institutional anchor of the acquirer's broader AI strategy.
The 2026 implication is that AI hyperscaler infrastructure strategy is materially dependent on Israeli engineering execution. AWS Graviton, Intel Gaudi, Nvidia's networking and orchestration stacks — all of which trace through Israeli design teams. The Hailo independent position provides an additional layer of optionality in edge AI specifically.
For Israeli technology export economics, the silicon stack is the highest-value-per-employee category. The combined revenue and product value generated by these five companies and their post-acquisition platforms operates on a scale that no other Israeli technology category matches. The institutional concentration in Israel — even as the corporate ownership shifts to US platforms — is one of the structural features that defines the global AI infrastructure architecture in 2026. The silicon layer sits below the Israeli foundation-model bench and the AI-agents layer covered in adjacent Olam coverage.
