A compute bet, not a rulebook
On June 16, 2026, Israel's government approved the National AI Program, a sweeping resolution that commits the state to building roughly 100,000 processing units for sovereign AI compute, a new National Artificial Intelligence Institute linking government, academia, industry and investors, and a National AI Directorate inside the Prime Minister's Office to run all of it (Prime Minister's Office). Brig. Gen. (res.) Erez Askal, who has led the Directorate since his appointment as Israel's first AI chief on October 12, 2025, called it "a historic window of opportunity" in which "it will become clear which countries will succeed" (Let's Data Science). Analysts estimate the compute buildout alone could cost $20–30 billion or more, with the most advanced 5,000 GPUs reserved for research and industry access on a rolling basis from 2027 through 2032 (Let's Data Science). Three weeks later, on July 5, the Directorate held its first industry consultation, with eleven major tech firms asking to formalize the forum as a standing voice into the new bureaucracy (Prime Minister's Office).
What the resolution does not do is create any new binding rule about how that compute, or the AI systems it will power, may be used. Israel still has no horizontal AI Act. Its governing document remains "Responsible Innovation: Israel's Policy on Artificial Intelligence Regulation and Ethics," published in December 2023 by the Ministry of Innovation, Science and Technology with the Ministry of Justice, which explicitly states its six ethical principles "should not be construed as legally binding to regulators or organizations." The policy instead assigns AI oversight to existing sectoral regulators in health, banking, education and consumer protection, backed by voluntary standards and sandboxes rather than statute (IAPP).
The case for waiting
There is a real argument for that sequencing. Binding AI statutes age badly — the EU's own AI Act has already required patch amendments as general-purpose models outpaced its risk categories. A government that locks in compliance obligations before it knows what a state-scale compute buildout will actually be used for risks writing rules for the wrong problem, and Israel's sectoral regulators retain full authority to act under existing health, finance and privacy law regardless of whether a horizontal AI statute exists. Keeping the Program's launch decoupled from a legislative fight also means it clears cabinet in a single resolution rather than getting stalled in the Knesset, where competing bills on unrelated matters were already consuming floor time this summer.
But the same soft-law policy that permits this speed also concedes its own limits. Its authors listed discrimination, human oversight, explainability, disclosure of AI interactions, safety, accountability and privacy as the seven challenges sectoral regulators must now handle — without new legal tools to do it. Independent analysis of the framework found regulators facing budget constraints, thin AI-specific expertise, and reluctance to enforce guidance they have no explicit statutory mandate to enforce, meaning market pressure from international clients — not government oversight — has been the actual driver of compliance so far (IAPP). That gap was tolerable when Israeli AI activity was scattered private-sector deployment. It is a different problem when the state itself becomes the largest AI compute operator in the country, running 100,000 processing units through a directorate answering directly to the Prime Minister's Office, with an explicit mandate to fold Cyber AI and "Physical AI" into public services.
Infrastructure first still needs guardrails now
The Ministry of Justice has floated a cross-cutting "Framework Law" that would give statutory footing to exactly the accountability questions the 2023 policy left to voluntary principles — algorithmic discrimination, liability for autonomous systems, and the legal status of AI-generated content. That law has not been introduced as Knesset text. Nothing in the June 16 resolution or the July 5 industry consultation ties its timeline to the GPU rollout that starts delivering research access in 2027.
That sequencing gap is the actual risk, not the compute spend itself. Building sovereign infrastructure is a legitimate and, given the AI compute race among mid-sized economies, arguably necessary move — Israel is not wrong to want processing power it doesn't have to rent from a foreign hyperscaler for systems touching cyber defense and public services. The mistake would be treating the Directorate's procurement leverage as a placeholder for governance rather than a tool for it. A National AI Institute that requires transparency logs, audit access and human-oversight commitments as conditions of the compute allocations it hands out to government agencies and research partners could embed real accountability into the buildout starting now, without waiting on a Knesset floor vote. Absent that, Israel will have built one of the world's more ambitious sovereign-AI compute bases on top of the same non-binding ethics document it published before ChatGPT-era foundation models existed.