2 View gallery <\/p>\n
![\"Israeli](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/02\/H1tVOlLv11g_0_0_3000_1568_0_x-large.jpg\" "\\"Israeli")
<\/a><\/p>\n
Israeli infrastructure AI map. <\/p>\n
(TLV Partners)<\/p>\n
Mandelbaum describes the shift as a series of abrupt technological eras. \u201cThe AI infrastructure opportunity has shifted dramatically three times in less than a decade. First it was about training. Then it was about inference. Now it’s about agents.\u201d<\/p>\n
During what he calls the Training Era, from 2015 to 2022, the central problem was how to build models that could predict binary outcomes. Companies struggled with data collection, labeling, and the mechanics of turning research into repeatable engineering. <\/p>\n
The arrival of large language models overturned that logic. \u201cThe birth of LLMs saw the emergence of the AI hyperscalers, which took upon themselves the massive costs and burden of training \u2018super\u2019 models,\u201d Mandelbaum notes. These systems were no longer predicting a single outcome but generating sequences of possibilities. The constraint shifted to inference, how to serve enormous models quickly and cheaply. Token costs pushed many firms toward open-source alternatives such as Llama, Kimi, and Qwen, which could be fine-tuned and run on private infrastructure.<\/p>\n
By late 2024 another inflection arrived. Models were finally good enough; the new challenge was usefulness. \u201cThe bottleneck is how to get them to do useful work reliably in the real world. Which is to say: the bottleneck is now agents,\u201d Mandelbaum explains. The questions confronting engineers are less about model size and more about behavior: retrieving the right context, planning actions, coordinating with other agents, and proving that the system did what it was asked.<\/p>\n
\u201cFor the most part, these are engineering challenges. Exactly the type of problems Israeli founders are uniquely qualified to solve,\u201d he says.<\/p>\n
That confidence is backed by a wave of infrastructure investment. Nebius has deployed 4,000 Nvidia HGX B200 GPUs in Israel, one of the country\u2019s first publicly available Blackwell installations, and signed agreements for 80 megawatts of data-center capacity with almost $900 million in investment. Crusoe\u2019s acquisition of Atero signals a similar bet on local talent. Nvidia, meanwhile, has effectively made Israel its second global headquarters following the Run:AI and Deci acquisitions and plans for a massive campus in Kiryat Tivon.<\/p>\n
These moves are beginning to anchor what Mandelbaum calls \u201cagentic sovereignty.\u201d \u201cThe dependence on American or Chinese model providers is a strategic vulnerability,\u201d he warns. \u201cWhen agents control infrastructure, manage sensitive data, and make decisions, you can’t outsource that capability entirely.\u201d Companies such as AI21, AAI, and Decart have started local model building, but he argues that \u201cmore attempts are needed.\u201d<\/p>\n
![\"Yonatan](\"https:\/\/www.newsbeep.com\/nz\/wp-content\/uploads\/2026\/02\/By5mjprwWg_0_37_747_561_0_x-large.jpg\" "\\"Yonatan")
<\/a>