On 17 May 2026, Israel's cabinet unanimously approved Prime Minister Benjamin Netanyahu's preliminary work plan for the National AI Directorate — a body established inside the Prime Minister's Office on 12 October 2025 and led by Brig.-Gen. (res.) Erez Askal, formerly head of the IDF's Digital Transformation Directorate. The plan rests on three pillars: deepen the AI talent pool (including reverse-brain-drain incentives for Israeli expatriates), guarantee Israeli researchers and start-ups access to 5,000 frontier-class GPUs per year from 2027 through 2032, and stand up fast-track 'applied AI' acceleration centres in priority verticals.
What is striking is what the plan is not. It is not a horizontal AI Act. It is not a licensing regime. It is not a new general-purpose risk classification. It is essentially a state-capacity bet — compute, people, and translation labs — layered on top of the sectoral 'Responsible Innovation' policy that the Ministry of Innovation, Science and Technology and the Ministry of Justice published in December 2023, which leaves AI oversight in the hands of incumbent regulators (banking supervision, health, transport, the Privacy Protection Authority) rather than a horizontal AI regulator.
The case for a heavier hand — fairly stated
There is a serious argument that Israel is moving in the wrong direction. The directorate was set up by government decision in September 2025; Askal was appointed without a competitive tender, with the Civil Service Commission expressly invoking an urgency exemption; and the policy unit was pulled out of the Ministry of Innovation and into the Prime Minister's Office, where parliamentary oversight is weakest. Critics point out that the Nagel Committee's own diagnosis — that 'Israel today is not in a good position to accelerate the field' while peer states 'are investing hundreds of billions' — is being used to justify both the speed and the concentration of authority. If AI is genuinely an inflection-point technology with safety, civil-rights, and military-export dimensions, the case for an independent statutory regulator with audit powers, a parliamentary reporting line, and rule-making authority — closer to the European AI Office, or even to Israel's own (still legally underpowered) National Cyber Directorate — is not frivolous.
Why the lighter approach is still the right one
Take that critique seriously, and the work plan still looks more proportionate than its loudest detractors allow. Three reasons.
First, Israel does not actually have a regulation gap; it has a compute gap and a talent gap. The 2023 Responsible Innovation framework, the 2024 financial-sector AI interim report, the 2024 Protection of Privacy Law amendment, and the regulatory-sandbox frameworks documented by Cahane and Sierra in the Cambridge Forum on AI: Law and Governance (December 2025) already cover most live AI risks — credit decisioning, medical devices, autonomous vehicles, personal-data processing — through agencies that have enforcement teeth and know their sectors. A new horizontal AI Act would duplicate this scaffolding, not strengthen it.
Second, the EU AI Act's first eighteen months have made clear that horizontal, rules-based AI regulation is heavy. Foundation-model providers serving European users now face documentation, evaluation, and downstream-deployment obligations that small ecosystems struggle to absorb. Israel — population under 10 million, but with one of the densest AI start-up bases per capita anywhere — cannot realistically run a duplicate regime and still attract frontier model training. Picking compute and talent first, with rules layered in sectorally, is what a small open economy with a credible export sector should do.
Third, the bottleneck is real. The Israel Innovation Authority confirmed on 19 January 2026 that the country's first AI training supercomputer — 1,000 Nvidia B200 accelerators, 70 percent allocated to industry and 30 percent to academia, operated by Nebius under the Telem programme — is now live. The 5,000-GPU-per-year guarantee through 2032 is, in essence, a sovereign hedge against an export-control or pricing shock from the United States. State action here is market-completing rather than market-distorting.
What the work plan still needs
A pro-innovation reading is not an uncritical one. The directorate should clear three accountability questions before it spends a shekel.
- Allocation transparency. Who gets the 5,000-GPU allocations each year, on what published criteria, and with what appeal mechanism? An opaque grant-of-favour from the PMO to favoured contractors would convert a market-completing programme into a patronage one.
- Statutory footing. The National Cyber Directorate spent more than a decade operating without a constitutive statute; that ambiguity has cost it in court and in mandate fights. The AI Directorate should be put on a Knesset-passed legal basis within twelve months, not left to govern by cabinet decision and budget ordinance.
- Independent oversight. Concentrating both promotion and oversight of AI inside the Prime Minister's Office creates a structural conflict. A parliamentary subcommittee, an external advisory council with civil-society and academic seats, or both, would be cheap insurance.
The bigger picture
Israel is, in effect, testing a third lane between Brussels's rules-first model and Washington's market-first one: build the inputs, leave the rules sectoral, and accept that some risks will be governed retrospectively rather than ex ante. That is a defensible bet for an economy whose comparative advantage is exporting AI capability rather than importing AI services. The accountability gaps in the 17 May work plan are real, but they are fixable without abandoning its core architecture — and abandoning that architecture for an EU-style horizontal regime would, for Israel specifically, trade a clear competitive position for a paper one.