On May 17, 2026, Israel's cabinet unanimously approved Prime Minister Benjamin Netanyahu's preliminary work plan for the National AI Directorate — the body established within the Prime Minister's Office on October 12, 2025 and headed by Brig.-Gen. (res.) Erez Askal. What is striking is not what the plan regulates, but what it declines to. Where Brussels spent four years drafting a horizontal risk law, Jerusalem approved a capability-building program and called it "a strategic move designed to ensure Israel's technological superiority" and keep the country "in the first line of world powers."
This is industrial strategy, not rulemaking — and for a country whose comparative advantage is speed and depth rather than scale, it is largely the correct call.
What the cabinet actually approved
The work plan rests on three pillars. First, human capital: bringing back expatriate researchers and funding retraining and computer-science degrees. Second, compute: making roughly 5,000 of the most advanced GPUs available each year for six years (2027–2032) to academia, the public sector, and startups. Third, applied-AI acceleration centers to push adoption into real products and public services.
The directorate's own footprint is modest. According to Calcalist's Ctech, it received about NIS 120 million — most of it for 2026, drawn largely from cross-cutting budget cuts — plus 20 staff positions, with roughly NIS 13 million a year for salaries and operations from 2027. The compute pillar builds on infrastructure already live: in January 2026 the Israel Innovation Authority announced a national AI supercomputer that will distribute 1,000 Nvidia B200 accelerators, 70% to high-tech firms and 30% to academic researchers.
The road not taken: no AI Act
The most consequential feature of Israel's approach is the regulation it isn't writing. Israel's 2023 "Responsible Innovation" AI policy — issued jointly by the Ministry of Innovation, Science and Technology and the Ministry of Justice — deliberately chose sector-specific, principles-based soft law, regulatory sandboxes, and existing statutory powers over a single horizontal act. The 2026 directorate does nothing to reverse that; it layers an investment program on top of a light-touch regulatory base.
The case for the opposite path deserves a fair hearing. A horizontal regime gives firms one rulebook instead of a patchwork, signals trustworthiness to export markets, and forces high-risk uses — biometric surveillance, automated hiring, credit scoring — to clear a bar before deployment rather than after harm. Those are real benefits, and the EU AI Act exists because diffuse sectoral enforcement can leave genuine gaps.
But the evidence so far favors Israel's restraint. Compliance cost falls hardest on the small firms and academic spinouts that are the engine of Israeli tech, and a pre-market conformity regime taxes experimentation precisely where Israel wins. Regulating concrete harms through the sectoral regulators that already understand health, finance, and defense — while keeping the national strategy focused on enabling capability — is the more proportionate division of labor. Notably, Israel still aligns its soft framework with the OECD AI Principles, preserving "international interoperability" for exporters without importing the AI Act's fixed-cost burden.
The honest case for state-led compute
Subsidizing GPUs and repatriating talent is industrial policy, and skeptics of industrial policy — this publication usually among them — should state the strongest version of the other side. The government-appointed Nagel Committee found that Israel "is not in a good position to accelerate the field," lagging peers who are spending hundreds of billions, and recommended roughly NIS 25 billion (about $6.6 billion) over five years. Sovereign-scale compute and the energy to run it are now strategic chokepoints that no single Israeli firm will build alone; coordinating ministries, universities, and industry plausibly needs a center of gravity. On those terms, a focused state push to remove a capability bottleneck — rather than to direct what gets built — is defensible.
Where the strategy risks tripping
The danger is not the ambition; it is the execution and the governance. Three concerns stand out.
- Concentration of power. Authority over AI policy moved from the Innovation Ministry to the Prime Minister's Office, and Askal — a capable defense-tech figure — was appointed without a competitive tender, with the Civil Service Commission citing urgency and security. Calcalist reports critics questioning the bypass and flagging the role of business figures close to Netanyahu in advising on policy. A national champion built through opaque appointments invites exactly the cronyism that decentralized, bottom-up Israeli tech has thrived by avoiding.
- Scale mismatch. The flagship supercomputer costs roughly $140 million and the directorate runs on NIS 120 million. Against a $40 billion Saudi AI fund or the U.S.–UAE mega-campus in Abu Dhabi, Israel cannot win a spending race — and it should not try. Money is the wrong metric; the country's edge has always been talent density and regulatory restraint, not capital intensity.
- Picking winners. State-allocated compute, split by quota between industry and academia, risks crowding out private provisioning and steering research toward politically legible projects. The 70/30 split should be a floor for market access, not a substitute for it.
The verdict
Israel's instinct — invest in enablers, regulate harms sectorally, avoid a horizontal AI Act — is the right one, and a welcome counter-model to the compliance-first reflex spreading across democracies. The plan will succeed to the extent it stays an enabling program: open compute access, talent magnets, and standards interoperability that let founders move fast. It will fail to the extent the Prime Minister's Office treats it as a patronage instrument or a procurement monopoly. The next test is the promised decade-long national plan. If it is written by tender, audited transparently, and keeps the regulatory hand light, Israel's bet on capability over control will look prescient.