Key Takeaways
- Anthropic’s newest large‑language model, Mythos, was initially limited to 50 U.S.–based organizations due to national‑security worries about malicious code generation.
- Canada has joined Anthropic’s preview program through the Canadian Centre for Cybersecurity, granting select Canadian firms access to Mythos.
- Reliance on U.S.–developed AI models exposes Canadian data to the U.S. CLOUD Act, undermining digital sovereignty and privacy.
- Canada can achieve AI independence by adopting and adapting open‑source LLMs and hosting the required infrastructure domestically.
- Recent open‑source releases (e.g., DeepSeek, Google’s Gemma 4 E4B) match the performance of frontier commercial models while being far cheaper and more hardware‑efficient.
- AI model capability density doubles every three to four months, enabling powerful models to run on modest devices such as smartphones.
- Canada’s highly educated workforce can sustain a home‑grown AI ecosystem if graduates are retained in domestic firms that develop or procure local AI services.
- Building Canadian‑owned AI infrastructure would secure personal, health, educational, and critical‑infrastructure data under Canadian law and political oversight.
- By pursuing open‑source models and domestic compute resources, Canada can attain a Mythos‑capable AI stack without the billions‑dollar investment required for proprietary U.S. models.
Introduction and Context
Anthropic, a leading Silicon Valley AI firm, unveiled a new iteration of its proprietary large‑language model called Mythos last month. Because of concerns that the model could be used to generate malicious code threatening mission‑critical systems, the initial release was restricted to 50 organizations, all based or operating in the United States and therefore subject to U.S. law. Recently, Canada’s Minister of Artificial Intelligence, Evan Solomon, announced that, via the Canadian Centre for Cybersecurity, Ottawa has become part of Anthropic’s preview campaign. This development will allow a limited number of Canadian companies to access Mythos, marking a step toward greater Canadian involvement in frontier AI technology.
Limitations of U.S.–Centric AI Models
Despite the welcome news of Canadian participation, the underlying reality remains that Mythos and comparable leading models are American technologies. The major commercial LLMs that power most AI applications in Canada today have been created by three large U.S. corporations. Under the U.S. CLOUD Act, data stored on these companies’ servers can be accessed by the U.S. government regardless of the physical location of the servers. This legal reach poses a clear national‑security risk: sensitive Canadian data—ranging from personal health records to critical‑infrastructure operational information—could be compelled to be disclosed to a foreign government with differing political and commercial priorities. Consequently, relying exclusively on U.S.–origin AI undermines Canada’s digital sovereignty and the expectation that personal and institutional data remain under Canadian jurisdiction.
Guiding Principles for Canadian AI Independence
To safeguard its data and processing capabilities, Canada should adopt two guiding principles. First, public‑ and private‑sector organizations ought to embrace and modify open‑source AI models for domestic AI applications. Open‑source code can be inspected, audited, and altered to meet Canadian legal and security requirements, eliminating the opaque black‑box nature of many proprietary models. Second, the physical infrastructure—servers, storage, and networking—used to run and store these models must be located in Canada and owned by Canadian entities. By controlling both the software stack and the hardware environment, Canada can ensure that data processing occurs under Canadian law, shielding it from extraterritorial assertions such as the CLOUD Act.
Evidence of Open‑Source Parity
The feasibility of this approach is bolstered by recent trends in the open‑source AI community. In January 2025, Chinese startup DeepSeek released a low‑cost, open‑source model whose performance rivals that of the latest commercial LLMs. Moreover, a pattern has emerged: roughly every three to four months after a major commercial LLM launch, an open‑source counterpart appears that closely matches its capabilities. If this cadence holds, an open‑source model comparable to Mythos should be expected within six months of its release. Such a model would provide frontier‑level language understanding and generation without the hundreds of millions—or even billions—of dollars required to train proprietary systems like ChatGPT or Claude.
Technical Advances Reducing Infrastructure Needs
Beyond software licensing, advances in model efficiency dramatically lower the hardware footprint needed for powerful AI. Researchers measure this efficiency with a metric called capability density—the amount of performance delivered per unit of compute. Capability density has been observed to double approximately every three to four months. Consequently, models that once demanded massive data‑center clusters can now run on far more modest hardware. For example, Google’s Gemma 4 E4B, an open‑source model, can be loaded and executed on an average smartphone while surpassing the capabilities of earlier iterations of OpenAI’s GPT‑4, which previously required a dedicated data centre. This trend means that Canada can deploy high‑performing AI using existing servers, edge devices, or modest regional compute farms, reducing both capital expenditure and energy consumption.
Leveraging Canada’s Educated Workforce
Canada possesses a competitive advantage in its talent pool: it leads the G7 in the proportion of working‑age residents holding college or university credentials. The authors, both directors of university programs—one in supply chain management and the other in management‑focused artificial intelligence—highlight that graduates from these programs can supply the skilled labor necessary to build, maintain, and innovate upon domestic AI systems. However, realizing this potential hinges on retaining these graduates within Canada. Policies and incentives should encourage them to join or create Canadian organizations that develop or procure AI services locally, rather than merely facilitating the purchase of foreign solutions. A vibrant local AI industry would keep expertise and intellectual property within national borders.
Vision for a Canadian‑Owned AI Stack
Combining open‑source models, domestically owned infrastructure, and a skilled workforce yields a clear path forward: Canada can develop its own Mythos‑capable AI stack without sacrificing performance or incurring prohibitive costs. Such a stack would enable Canadian institutions to train models on nationally sourced data, run inference on Canadian‑managed hardware, and derive insights that remain subject to Canadian law and democratic oversight. Applications could range from optimizing weather‑related supply chains and reducing internal trade barriers to securing health‑care records, protecting student data in educational platforms, and fortifying critical infrastructure like power grids and pipelines against cyber threats.
Conclusion
While Canada is unlikely to become a conventional AI powerhouse dominated by a few massive corporate players, it does not need to follow that trajectory to achieve strategic advantage. By adopting open‑source LLMs, ensuring that the compute layers reside on Canadian‑owned soil, and retaining its highly educated workforce, Canada can secure digital sovereignty, protect privacy, and maintain control over the AI‑driven knowledge that underpins its economy and safety. The recent inclusion in Anthropic’s Mythos preview program is a positive signal, but the true safeguard lies in building an autonomous, home‑grown AI ecosystem that reflects Canadian values and serves Canadian interests.