Governments are buying equity in frontier AI labs rather than funding them with grants. Canada committed up to C$1.6 billion for a stake in Cohere, Chinese state investors supplied about 27% of DeepSeek’s US$15.8 billion round, and Washington has debated taking stakes in American labs. Frontier model work is being treated as national infrastructure, which splits AI into capital-intensive foundation labs backed by sovereign money and a much larger layer of product companies that buy intelligence instead of training it.
For decades the startup playbook was simple. Build something better, raise venture capital, find customers, grow.
Frontier AI breaks that model. Training advanced models costs billions in compute, talent and electricity. And the price of intelligence keeps falling while the cost of staying competitive keeps rising.
Now governments are showing up. Not with grants and research funding. As shareholders.
Canada bought a stake in Cohere
On October 29, Prime Minister Mark Carney announced Canada would invest up to C$1.6 billion for an equity stake in Toronto-based Cohere. It’s part of a round targeting C$7 billion, valuing the company around C$36 billion. Cohere has committed to 1,000 Canadian technical jobs, C$1.7 billion in domestic compute, and keeping its headquarters in Canada.
This isn’t a subsidy. Canada is becoming an owner.
The calculation is bigger than financial return. Canada wants the capability, the IP, the jobs, and some independence from foreign providers. It is treating Cohere as a strategic asset, the way it treats energy or telecom. It is also a sharper instrument than the data centre strategy I wrote about in July, which had the ingredients and not the recipe.
China is doing it at scale
The same day, The Information reported DeepSeek raised US$15.8 billion, with about 27% from Chinese state-owned investors. Roughly US$4.3 billion of state-backed capital.
DeepSeek reportedly doubled its target partly because American export controls made the NVIDIA hardware it needs more expensive. The company that got famous for building capable models efficiently still needs nearly US$16 billion to stay at the frontier.
Its open-weight strategy does something else too. Making strong models widely available spreads Chinese technological influence even when the financial return isn’t obvious. For a state, that value can sit right alongside profit.
Washington wants a piece too
During 2026, Trump administration officials explored taking government equity in leading American AI companies, including discussions involving OpenAI. The President also raised the idea publicly for OpenAI and Anthropic.
No federal stake in Anthropic has been confirmed. Reuters reported in July that Anthropic and the administration had not discussed an equity transaction.
The discussion is still revealing. The country with the deepest venture capital markets in the world is asking whether private ownership alone is the right model for companies building strategically important infrastructure.
The math changed
A software business can grow revenue faster than its infrastructure costs. Frontier AI works the other way.
Every generation of models demands enormous investment. Inference costs money forever. Hardware depreciates fast. And open-weight competition and falling API prices keep pushing revenue down. The cost of competing goes up while the price of intelligence goes down.
Governments do a different sum. They can value national security, technological independence, jobs, research capacity and economic competitiveness alongside financial return.
That doesn’t make government investment wise. Governments overpay, protect incumbents, make bad calls.
But it means a frontier lab backed by sovereign capital can compete under financial assumptions a normal startup can’t.
What happens to everyone else
I think AI companies are splitting in two.
The first builds foundation models and infrastructure. These look more like utilities over time, and they need access to sovereign capital, hyperscaler balance sheets, or both.
The second builds products on top of those models. This is where most startups will live.
A company solving a problem in healthcare, legal, manufacturing or enterprise operations doesn’t need to train a frontier model. It can buy intelligence from competing providers and put its work into customers, proprietary data, integration and outcomes.
The irony is that government money poured into foundation models may make it cheaper for everyone else to build useful businesses on top.
From venture capital to industrial policy
Governments have long treated energy, telecom, aerospace and semiconductors as strategically important. AI is entering that same category.
Canada buying into Cohere, China backing DeepSeek, Washington debating ownership: different approaches, one shared recognition. Advanced AI may matter too much to a country’s economy and security to leave entirely to private capital.
Ownership doesn’t guarantee sovereignty. A Canadian AI company still depends on American chips, foreign cloud and global supply chains. Public money doesn’t guarantee commercial success either.
But something shifted. Some AI companies are building software businesses. Others are building capabilities governments now treat as essential national infrastructure.
They still look like startups, with founders, rounds and ambitious valuations. The economics are starting to look like strategic industry.
The race isn’t just to build the best model anymore. It’s to decide who owns, funds and controls the infrastructure of intelligence. And that changes the startup math for everyone.
Frequently Asked Questions
How much is Canada investing in Cohere?
Up to C$1.6 billion, as an equity investment made in stages rather than a grant. It was announced on October 29, 2026, alongside private investors including NVIDIA, Bell and Germany’s Schwarz Group. In return Cohere committed to creating at least 1,000 Canadian technical jobs by 2030, spending at least C$1.7 billion on computing infrastructure in Canada by 2030, and keeping its Canadian incorporation and headquarters.
What is the difference between a government grant and a government equity stake?
A grant is money spent. An equity stake makes the government an owner, so it shares in the company’s gains and losses and gets a seat at the table on things like where the company is headquartered and who can buy it. Canada has used grants for AI before, including an earlier C$240 million commitment to Cohere’s compute capacity. This one is different in kind, not just in size.
How much state money went into DeepSeek’s funding round?
About 27% of a US$15.8 billion round, so roughly US$4.3 billion, according to The Information. DeepSeek reportedly doubled its original target, partly because American export controls pushed up the cost of the NVIDIA hardware it needs.
Does the US government own a stake in OpenAI or Anthropic?
No stake has been confirmed in either. Trump administration officials explored how such stakes might be structured during 2026, and the President raised the idea publicly. Reuters reported in July 2026 that Anthropic and the administration had not discussed an equity transaction.
Does an AI startup need to train its own foundation model?
Most don’t. Training at the frontier now takes the kind of capital that comes from governments or hyperscalers. A company solving a problem in healthcare, legal, manufacturing or enterprise operations can buy intelligence from competing providers and compete on customers, proprietary data, integration and outcomes instead. Sovereign money flowing into foundation models may make that cheaper.
Does government ownership give a country AI sovereignty?
Not on its own. A Canadian AI company still depends on American chips, foreign cloud capacity and global supply chains, and public money is no guarantee of commercial success. Ownership changes who holds the asset and who has a say in its future. It doesn’t by itself change what the asset depends on.
Sources
- Government of Canada — Cohere investment announcement, October 29, 2026
- Reuters — Canada’s Cohere investment, October 29, 2026
- The Information — DeepSeek’s $16 billion funding, October 29, 2026
- Semafor — Trump administration’s potential AI stakes, June 17, 2026
- Reuters — Anthropic government stake discussions, July 2, 2026
