Shalev Hulio built a company that turned smartphones into glass houses. Now the former NSO Group chief executive is asking governments to believe he can do the opposite—construct digital fortresses impervious to foreign snooping, powered by artificial intelligence that answers only to the state that buys it.
The pitch, which Hulio has been making to defense ministries and intelligence agencies across Europe and the Middle East, is elegantly opportunistic. Western democracies are increasingly nervous about their dependence on American AI infrastructure. OpenAI's models run on Microsoft's Azure; Anthropic leans on Amazon and Google. Every prompt, every fine-tuned dataset, every strategic query passes through servers ultimately governed by US law and, critics argue, US interests. Hulio's proposition: let us build you something that stays home.
The trust deficit problem
The obvious difficulty is that Hulio is perhaps the least trusted man in cybersecurity. Pegasus, NSO's flagship spyware, has been found on the phones of journalists, human rights activists, and opposition politicians from Mexico City to Budapest. The company was blacklisted by the US Commerce Department in 2021 and has faced lawsuits from Apple, WhatsApp, and multiple governments. Hulio stepped down as CEO in 2022 amid a debt restructuring that valued the company at a fraction of its peak.
Yet the sovereign-AI sales tour suggests he believes the geopolitical mood has shifted in his favor. Governments that once condemned NSO's client list are now confronting a starker question: whose surveillance apparatus do they prefer to be subject to? American hyperscalers hoover up data with legal impunity under the CLOUD Act. Chinese alternatives are off the table for NATO members. A boutique Israeli firm promising on-premises AI that never phones home starts to look less like a pariah and more like a pragmatic compromise.
What sovereign AI actually means
The term "sovereign AI" has become fashionable shorthand for nationally controlled compute and models, but its practical meaning varies wildly. France's Mistral AI offers open-weight models that can be deployed locally. Saudi Arabia's nascent AI initiatives emphasize data residency. Hulio's version, according to people briefed on his presentations, emphasizes hardware-software integration—custom inference chips, air-gapped training environments, and models fine-tuned on classified government corpora that never touch the public internet.
Whether his new venture can deliver remains speculative. Building competitive large language models requires billions in capital, scarce GPU allocations, and engineering talent that overwhelmingly clusters in San Francisco and London. Hulio's Rolodex of intelligence contacts is formidable; his access to Nvidia H100s is less certain.
Our take
The Hulio rehabilitation tour is less about one man's second act than about a structural opening in the AI market. Governments want intelligence capabilities they control, and the incumbents—American giants, Chinese state champions—come with strings attached. That vacuum will be filled by someone. The uncomfortable question is whether the someone who pioneered phone-hacking as a service is the right vendor for the job, or simply the most brazen applicant.
