Canada and Germany commit up to C$300M to Bengio's LawZero

Why it matters

Bengio is trying to fund advanced AI research without the incentives of a commercial lab. A C$300M ceiling could make that credible, while its conditions will shape LawZero's independence.

LawZero, the Montreal AI safety nonprofit founded by Yoshua Bengio (@Yoshua_Bengio), said in a September 16th release that Canada and Germany have committed up to C$300 million to expand its research team, build dedicated computing infrastructure in Canada and open an office in Berlin.

The phrase "up to" carries much of the weight. LawZero's release gives a ceiling for the joint commitment, without specifying the two countries' respective contributions, the disbursement schedule or the conditions attached to the money. This is government funding rather than a priced venture round, and LawZero disclosed no valuation.

For Bengio, LawZero's founder, co-president and scientific director, the commitment gives unusual scale to a project he began after changing the direction of his career in 2023. The 2018 A.M. Turing Award recipient founded Mila, the Quebec AI Institute where LawZero was incubated, and co-founded Element AI, which ServiceNow acquired in January 2021.

LawZero says Bengio redirected his work after concluding that private labs were moving too quickly toward increasingly capable and autonomous systems. When he launched LawZero on June 3rd, 2025, it had over 15 researchers. Independent coverage put its initial philanthropic backing at roughly $30 million, including support associated with Jaan Tallinn, Schmidt Sciences, Open Philanthropy and the Future of Life Institute.

That launch framed the nonprofit structure as protection from commercial and government pressures that could compromise safety. The structure remains. A prospective C$300 million from two governments will still make the funding agreement, governance boundaries and research milestones central to LawZero's claim of scientific independence.

A public-sector scale-up

The commitment follows months of policy work between Ottawa and Berlin. On February 14th, Canada and Germany signed an AI declaration and launched a Sovereign Technology Alliance focused on compute infrastructure, research, commercialization and reduced dependence on foreign technology.

That declaration specifically identified LawZero as a possible area for future cooperation. It also said Germany would be the country of the year at the ALL IN conference in Montreal, where the funding commitment was announced seven months later.

The timing puts Bengio's safety project inside a wider industrial policy push. Canada gets a large advanced-AI research program and dedicated compute on Canadian soil. Germany gets LawZero's first stated European office and a direct connection to a research program outside the dominant US frontier labs. LawZero gets public financing for the research team and computing capacity its roadmap calls for.

LawZero says Canadian infrastructure providers Hypertec and 5C will build and provide the dedicated compute capacity. LawZero has not given a planned opening date or technical specifications for that infrastructure in the funding release.

The Berlin office would expand LawZero's presence in Europe and put it closer to European researchers and institutions working on AI safety and governance. LawZero describes both projects as planned expansions, rather than facilities already operating.

Bengio's alternative to the agent race

The technical bet behind the funding is Scientist AI, Bengio's proposed alternative to systems trained to act as autonomous agents. LawZero wants to build a highly capable predictor that forms probabilistic explanations of the world without developing preferences about which outcomes should occur.

In the design LawZero published, the predictor would learn to distinguish a factual claim from the act of somebody making that claim. Its training process would avoid rewarding the model based on what happens after its answers are deployed, a property LawZero calls "consequence invariance." Tasks requiring action would be handled through separate, auditable software guarded by the predictor.

LawZero's July 2nd safety paper presents a formal argument for why that design should reduce the risk of the predictor developing hidden goals. It is an early research case, rather than evidence from a deployed product or an independent evaluation. LawZero's own account limits the argument to the predictor's internal agency and honesty. Human misuse, mistakes and the safety of larger agentic systems built around the predictor require separate safeguards.

That distinction matters because the government commitment is funding a research program whose central safety properties still need to be tested experimentally. LawZero says Scientist AI could eventually monitor other models, block dangerous actions and help researchers generate hypotheses in fields such as medicine, climate science, cybersecurity and AI safety. Those uses remain goals for the roadmap.

Building an institution around the research

Bengio has paired the scientific thesis with operators who have built and financed technology organizations. Sam Ramadori, LawZero's co-president and executive director, previously ran building-automation startup BrainBox AI and spent 15 years in private equity. Senior Research Director Joumana Ghosn previously led applied machine-learning research at Mila and completed her PhD under Bengio.

LawZero also has an independent board and advisory council that includes former New Zealand Prime Minister Jacinda Ardern, former Swedish Prime Minister Stefan Lofven, author Yuval Noah Harari, former Rolls-Royce CEO John Rose, Mila CEO Valerie Pisano and Maria Eitel, founder of the Nike Foundation and Girl Effect.

The C$300 million ceiling gives that institution a chance to move beyond a small philanthropic research lab. The decisive questions sit in execution: how much money arrives, which milestones release it, what compute is built and whether Scientist AI's safety argument survives empirical testing at scale.

Bengio's wager is that a nonprofit can pursue advanced AI with a training objective centered on accurate prediction instead of autonomous action. Canada and Germany have attached a large prospective budget to that wager. The funding terms and the research results will determine how much of the headline number becomes scientific capacity.