A big infrastructure bet, a small regulatory footprint
On June 16, 2026, Israel's government approved a national AI program built under a newly formalized National Headquarters for Artificial Intelligence in the Prime Minister's Office, headed by Brig. Gen. (res.) Erez Askal. The resolution sets a target of 100,000 sovereign compute units — server-grade GPUs housed on Israeli soil rather than rented from foreign clouds — alongside a new National AI Institute meant to link government, academia, industry and investors, plus a national quantum computer built "as much as possible" on Israeli technology (Ynet; Jerusalem Post).
What the resolution does not do is create new AI law. Israel continues to govern AI the way it has since December 13, 2023, when the Ministry of Innovation, Science and Technology and the Ministry of Justice jointly endorsed a "Responsible Innovation" policy paper that explicitly rejected a horizontal, EU-style statute in favor of soft law: non-binding ethical principles, voluntary standards, sandboxes, and instructions to existing sectoral regulators — health, banking, consumer protection — to write AI-specific rules within their own existing authority (Israel's AI Policy 2023 policy paper).
The case for guardrails
The strongest objection to this pairing is straightforward: capacity and risk are scaling faster than accountability. A 100,000-GPU sovereign buildout is not a small commitment — Tel Aviv University AI researcher Nadav Cohen estimates the infrastructure alone could cost $20 billion to $30 billion once servers, power, cooling and multi-year hardware refresh cycles are counted, and that figure is an outside analyst's estimate, not a number the Prime Minister's Office resolution itself appropriates (Ynet). Spending at that scale, routed through a new coordinating body rather than a body with binding rulemaking power, is exactly the kind of program critics of soft law worry about: heavy public investment in capability, thin public accountability for how it is used. The plan also leans into Cyber AI and Physical AI — categories with obvious defense and dual-use applications — plus deepfake defense, three areas where the case for enforceable, not merely aspirational, standards is at its strongest. And the workforce side of the plan is not trivial: government estimates cited alongside the announcement put the number of Israeli workers needing partial or complete AI-driven reskilling at one million to as many as four million (Jerusalem Post) — a labor disruption on a scale that, in most jurisdictions, would come with statutory retraining entitlements, not just a labor-market "mechanism."
Why soft law is still the right call, for now
Even granting all of that, Israel's bet is defensible, and arguably better calibrated than the alternative. The EU's AI Act imposes horizontal, risk-tiered compliance obligations on developers before regulators have much empirical evidence about which specific harms materialize in which sectors; Israel's approach — a central coordinating hub with no independent enforcement power, paired with sector regulators writing fit-for-purpose rules as they observe real deployments — mirrors the model Israel already used successfully for cybersecurity, and lets rules track the technology instead of freezing it in a 2026 snapshot (IAPP). Where the sharpest risks already have a legal home, Israel has moved: the 2024 amendment to the 1981 Protection of Privacy Law substantially expanded enforcement powers over automated data processing, meaning the AI plan is not filling a total legal vacuum so much as leaving genuinely novel, fast-moving harms to catch up empirically rather than by statute drafted in advance of the evidence. For an economy where high-tech exports account for roughly half of all exports, betting compute infrastructure now and calibrating binding rules later is a coherent sequencing choice, not regulatory abdication.
What would make this credible
The honest gap is not that Israel chose soft law — it's that soft law without transparency invites exactly the accountability critique above. Three fixes would close most of it without requiring a single new AI statute. First, publish an actual appropriated budget for the compute buildout, distinct from the analyst range now doing the work an official figure should. Second, give the National AI Institute a public reporting mandate — who gets access to sovereign compute, on what terms — since a government-controlled GPU fleet is a novel form of industrial policy that taxpayers are entitled to see audited. Third, attach the reskilling commitment to something with legal teeth, not aspirational language, given the four-million-worker upper estimate now in circulation. None of that requires Israel to import the EU's horizontal AI Act. It does require the soft-law model to prove it can be transparent, not just fast.
The takeaway
Israel is making a deliberate, well-precedented choice: build capacity first, regulate empirically, keep enforcement inside existing sectoral regulators rather than a new AI authority. That is a more defensible sequencing than critics allow — but it only stays defensible if the accountability infrastructure around the spending and the labor transition keeps pace with the compute infrastructure itself.