Sovereign Safety: Canada and Germany Bet CAD $300M on Bengio's LawZero

On Wednesday afternoon at the ALL IN conference in Montréal, something unusual happened in the funding history of artificial intelligence: two national governments jointly wrote a very large check not for compute, and not for a frontier lab, but for AI safety. Canada and Germany announced a commitment of up to CAD $300 million for LawZero, the non-profit founded by Turing Award winner Yoshua Bengio — the largest public commitment to date for “safe-by-design” AI, and a deliberate bet that there is a technological alternative to racing ahead and patching later.

The announcement was timed for symbolism. Germany was the “country of honour” at ALL IN, and the two countries had already co-signed a sovereign technology alliance in February. The LawZero grant deepens that axis — and on the same day, Toronto’s Cohere finalized its merger with Germany’s Aleph Alpha. Ottawa and Berlin are visibly constructing a transatlantic AI corridor that is neither Silicon Valley’s nor Beijing’s.

What the money actually buys

The commitment, split between the two countries (Canada’s share flows through the Strategic Response Fund, per reporting from The Globe and Mail), is earmarked for three things:

• People. Growing LawZero’s international scientific research team — the organization has said the funding will create “hundreds and hundreds” of full-time jobs, in Evan Solomon’s words at the announcement press conference. Solomon is Canada’s Minister of Artificial Intelligence and Digital Innovation, a portfolio that barely existed two years ago.
• Sovereign compute. A dedicated compute infrastructure in Canada, built with Montréal-area partners Hypertec and 5C. This is not cloud rental; it is anchored infrastructure on Canadian soil, explicitly framed as “sovereign.”
• A European foothold. A new LawZero office in Berlin, plugged into Germany’s innovation ecosystem — and, in the words of German digitalization minister Dr. Karsten Wildberger, “strengthening our technological sovereignty.”

The technology being funded

LawZero’s flagship project is Scientist AI — and its design philosophy is a direct rejection of the agentic paradigm the rest of the industry is scaling. Current frontier systems are trained to act autonomously and pursue goals. Scientist AI is being built to do the opposite: reason transparently, produce reliable evidence-based outputs, and remain “unbiased by goals of its own.”

The intended deployment is the interesting part. Scientist AI is positioned as a monitor — a guardrail system watching the agentic models everyone else ships. In practice: an AI whose job is to understand what other AIs are doing, flag misaligned or deceptive behaviour, and intercede — a kind of independent safety layer that doesn’t share the incentives of the system it supervises. Bengio has argued that a monitor built without reinforcement learning on goal-seeking behaviour can evaluate a goal-seeking agent without inheriting its blind spots.

That framing has become newly concrete. The same week as the funding announcement, Spain’s data protection agency disclosed what it calls the first personal-data breach executed end-to-end by an autonomous AI agent — reconnaissance, login, application probing, data modification, invoice access, all without human steering. OpenAI, meanwhile, published a misalignment reporting framework and disclosed six previously unreported safety incidents. “Recent events demonstrate that safety is itself a core capability: scientific innovation and public safety can and must advance together,” Bengio said in his statement.

Why a non-profit, and why governments

LawZero is structured as a non-profit “to ensure it is insulated from market and government pressures, which risk compromising AI safety,” incubated at Mila, the Montréal institute Bengio founded. Its board and advisory council are deliberately non-technical in part: Nike Foundation founder Maria Eitel chairs the board; former prime ministers Jacinda Ardern and Stefan Löfven, author Yuval Noah Harari, and Mila CEO Valerie Pisano sit alongside. Initial backing came from philanthropy — Gates Foundation, Schmidt Sciences, the Future of Life Institute, NVIDIA, Jaan Tallinn — on the order of $30M. Wednesday’s commitment is an order of magnitude beyond that, and it comes from treasuries rather than donors.

That changes the character of the organization. Government funding at this scale makes LawZero less an advocacy project and more infrastructure — closer in spirit to a national lab, with two of them. It also answers a question that has hovered over AI safety since the field professionalized: who pays for the unprofitable work of making the profitable systems safe? Canada and Germany have now offered one answer, and in doing so given themselves a stake in the outcome they want to see: AI that reflects, in Solomon’s phrasing, “our values and our interests,” built “on Canadian terms.”

The bet is not without tension. A monitor deployed as a guardrail for frontier models is only as influential as its adoption; OpenAI, Anthropic and Google have been quietly discussing their own industry standards body for testing and auditing, and there is no guarantee the big labs’ internal safety teams defer to an external, government-backed watcher. And “safe-by-design” remains a research hypothesis as much as a product — Scientist AI’s reliability claims will face the same adversarial pressure every other safeguard eventually meets.

But the structural signal is the story. At ALL IN, Bengio was asked — in French — whether researchers’ warnings that AI could threaten humanity within a decade were too outlandish. His answer was a single word: no. “I’m doing this with an urgency I feel in my bones,” he said. As of this week, two governments are feeling it too, and putting roughly CAD $300 million behind the feeling. The frontier labs have their capitals. Safety now has one as well.