In Chapter 6, the guide mapped what's actually in the ground — 300-plus Canadian data centres, the city-scale megawatt loads in Toronto, Montréal, and…

In Chapter 6, the guide mapped what's actually in the ground — 300-plus Canadian data centres, the city-scale megawatt loads in Toronto, Montréal, and Vancouver, the hyperscaler regions opened by AWS, Microsoft Azure, and Google Cloud, and the two most-watched made-in-Canada storylines: the sovereign-compute push (Bell's AI Fabric build, the $240M federal Cohere commitment) and the contested O'Leary/Greenview Wonder Valley project in Treaty 8 territory. The chapter closed with a structural finding: "Sovereign AI infrastructure" as a strategic direction is meaningful; as a description of current operational reality, it would be misleading. This chapter goes inside that asymmetry. The question is not just whose buildings sit on Canadian land, but who actually holds AI power in Canada: operationally, financially, legally, and politically. The map turns out to be more complicated than either the federal strategy framing or the standard critique of the federal strategy framing suggests. There are at least four distinct actor categories with different kinds of power, and the federal AI for All strategy treats them asymmetrically in ways worth naming. You'll leave the chapter with: the four-category map of who holds Canadian AI power; the Cohere case study done honestly (neither as champion-narrative validation nor as critique target); the Public First commissioned-research pattern as a worked example of how economic-impact narratives are produced and what that means for reading them; the proxy-lobbying mechanism flagged in the Canadian Union of Public Employees' Senate submission; and a working answer to the question "who in Canada is in a position to actually shape what AI looks like over the next decade?" — which is one of the questions the strategy implies it answers but doesn't quite address directly. ---

The four categories

Most coverage of Canadian AI policy talks about "Canadian AI companies" as if they were a single category. The actual landscape has at least four distinct groups, with very different relationships to power and very different positions in the AI for All strategy. Naming them properly is the prerequisite for everything else in the chapter.

Category one: Foundation-model developers. Companies that train and operate large-scale AI models, the underlying systems that other applications run on top of. Globally this is a small group: OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, xAI, Alibaba's Qwen team, DeepSeek, and a handful of others. The single Canadian member of this category is Cohere, founded in Toronto in 2019 by Aidan Gomez (one of the co-authors of the original 2017 "Attention Is All You Need" paper that introduced the Transformer architecture underneath every current major LLM), Ivan Zhang, and Nick Frosst. Cohere is private, valued at roughly US$5.5 billion in its July 2024 Series D and about US$7 billion after August–September 2025 raises; in April 2026 it announced a merger with Germany's Aleph Alpha that would value the combined company at roughly US$20 billion, pending regulatory approval, with major investors including Inovia Capital, Index Ventures, NVIDIA, Oracle, Salesforce, and PSP Investments, the Canadian public-sector pension investment manager that led the 2024 round. Its product strategy targets enterprise customers (Oracle, Notion, Spotify, McKinsey) rather than the consumer market that ChatGPT and Gemini contest.

A note on Canadian AI companies that are not foundation-model developers but are part of the broader Canadian AI landscape. AI for All foregrounds Cohere as the national champion, which is technically accurate at the foundation-model layer but understates the breadth of the Canadian AI company ecosystem. Other named Canadian AI companies operating at meaningful scale include Ada (Toronto, customer-service AI for enterprise deployment), Waabi (Toronto, autonomous-trucking AI led by Raquel Urtasun), Coveo (Montréal and Québec City, enterprise search and AI), Sanctuary AI (Vancouver, robotics and physical-AI), and a longer tail of vertical-AI companies in healthcare, financial services, and industrial automation. These companies are not foundation-model developers; they build applications and specialized models on top of foundation models from Cohere, OpenAI, Anthropic, and others. The structural significance: Canadian AI capacity is broader than the foundation-model layer, and AI for All's national-champion framing centered on Cohere obscures a more distributed picture of Canadian AI activity. Several of these companies appear on the Voluntary Code of Conduct signatory list (Chapter 5), which connects them directly to federal voluntary-framework AI governance.

Category two: Foreign hyperscalers with Canadian regions. Amazon Web Services (Montréal + Calgary), Microsoft Azure (Toronto + Quebec City), Google Cloud (Montréal + Toronto), plus Oracle, IBM, and Tencent at smaller scale. These companies provide the compute infrastructure most other Canadian AI work runs on, including, for now, much of Cohere's own deployment infrastructure. Their Canadian footprints are operationally Canadian (data residency, Canadian jurisdiction for customer contracts) but corporately American or Chinese, meaning subject to home-jurisdiction legal frameworks including, in the US case, the CLOUD Act, which we'll come back to in Chapter 9.

Category three: Canadian telecommunications incumbents. Telus, Bell, and Rogers, with Telus and Bell most active in the AI conversation. Both market "sovereign cloud" offerings to enterprise and government customers, both have made significant data-centre investments, and both are central to AI for All's sovereignty framing: Bell directly through the Bell AI Fabric partnership with Cohere, Telus through its sovereign-cloud and AI-services positioning. These are Canadian-controlled corporations under Canadian regulatory jurisdiction (CRTC, Competition Bureau), with substantial existing relationships with federal and provincial governments. Their lean: pro-Canadian-infrastructure (the framing favours them), pro-incumbent (the framing also disadvantages newer entrants), with environmental and labour positions that have not been independently audited at the depth their public claims would require.

Category four: The research institutes and the federal alignment apparatus. Mila (Montréal), the Vector Institute (Toronto), and the Alberta Machine Intelligence Institute (Amii, Edmonton): the three national institutes coordinated through CIFAR under the Pan-Canadian AI Strategy. These are formally non-profit research organizations, primarily federally funded (the Pan-Canadian strategy committed $443 million in 2021; AI for All added over $200 million across the three institutes plus the Canadian AI Safety Institute). They train graduate students, publish research, and increasingly act as commercialization platforms. Mila CEO Valérie Pisano, Amii CEO Cam Linke, Vector CEO Glenda Crisp (appointed April 2025). Bias note from Chapter 1's methodology: these institutes are not independent observers of Canadian AI policy. They are major beneficiaries of it, and Mila CEO Valérie Pisano gave the on-record positive launch quotes ("ambitious," a strategy that "puts a stake in the ground") to CBC on launch day, with similar comments in trade-press coverage the day after. The guide will continue to label them as "federally funded, AI for All beneficiaries" rather than as neutral academic voices.

A fifth informal category, commissioned-research firms and consultancies, sits across the others, producing the economic-impact narratives that policy debates run on. The guide will treat this as a methodological category rather than a structural one, but it deserves its own section later in this chapter.

These four categories do not have equal positions in AI for All. The strategy explicitly names Cohere as the champion, names the three institutes as research and literacy delivery vehicles, names the telecom incumbents (especially Bell) as infrastructure partners, and treats foreign hyperscalers as somewhere between necessary partners and competitive threats depending on which pillar of the strategy is being discussed. The asymmetry is itself the policy choice, and reading the strategy is partly a matter of seeing which actors get framed which way.

---

The Cohere case study

Cohere is the single Canadian company most foregrounded by the federal AI strategy, and the test case for whether the "national champion" framing means what it implies. The fair read is more interesting than either the champion-narrative or the critique-target version.

The technical substance. Aidan Gomez was a co-author on the 2017 Google Brain paper "Attention Is All You Need," the paper that introduced the Transformer architecture, the foundational technical innovation underneath every current generation large language model (GPT, Claude, Gemini, Llama, Cohere's own Command series). The Canadian foundation-model-developer category exists because one of the Transformer's co-authors went home to Toronto and started a company in 2019. The Canadian historical claim on contemporary AI is real, not marketing.

The financial substance. Cohere's July 2024 Series D valued the company at approximately US$5.5 billion; an August 2025 raise led again by PSP Investments, plus a September extension, lifted that to about US$7 billion. Then, in April 2026, Cohere announced a merger with Germany's Aleph Alpha (with Germany's Schwarz Group leading the accompanying Series E), valuing the combined company at roughly US$20 billion, with close expected later in 2026 pending regulatory approval. A "sovereign Canadian champion" merging transatlantic is itself a sovereignty data point worth holding onto. Major investors include Inovia Capital (Montréal-based; the same firm partnering with Mila on the $100M Venture Scientist Fund — a private Mila–Inovia venture fund, not a federal program), Index Ventures, NVIDIA, Oracle, Salesforce, Cisco, and PSP Investments, the Canadian public-sector pension investment manager that led the Series D. The federal commitment of CA$240 million to Cohere's compute build-out in December 2024 is real public funding for real Canadian AI infrastructure, though the facility it backs is built and operated by the US cloud provider CoreWeave.

GOBLIN FACTS — a valuation is not a scoreboard. Cohere was valued around US$5.5 billion in its July 2024 round and roughly US$7 billion in 2025, with PSP Investments, a Canadian public-sector pension manager, leading the 2024 raise. Worth knowing. Also a private number set by investors, not a public market and not a measure of how much the company actually ships.

The strategic positioning. Cohere's product strategy is enterprise-focused rather than consumer-focused. Where OpenAI competes for chat-interface users at scale, Cohere competes for enterprise contracts (Oracle, Notion, Spotify, McKinsey, plus a growing federal-and-provincial-government client base in Canada and similar enterprise positioning in Europe). The positioning matters because enterprise foundation-model markets differ from consumer markets in operation: slower adoption cycles, larger contracts, more customization, more attention to data governance and jurisdiction. The Canadian regulatory environment is meaningfully more favourable to enterprise foundation-model providers than the US consumer-AI market environment.

Now the complications, and there are three.

The first is that Cohere is privately held, and its valuation, revenue, and growth trajectory are reported through company press releases and friendly press coverage rather than through public-company disclosure regimes. The US$5.5B valuation is a venture-capital valuation, not a market-tested public-company valuation; the company's actual revenue and burn rate are not publicly disclosed. The guide takes Cohere's existence and product seriously without taking its private valuation as the equivalent of operational scale.

The second is that Cohere's investor base is substantially American (Index Ventures, NVIDIA, Oracle, Salesforce, Cisco, all US-headquartered) and includes the major hyperscalers Cohere supposedly represents the sovereign alternative to. Oracle is both a Cohere investor and the operator of one of the cloud regions Cohere's products run on. NVIDIA is both a Cohere investor and the chip supplier whose GPUs Cohere's models are trained on. The "Canadian champion" framing is technically accurate at the headquarters and founding-team level. At the supply-chain, capital-stack, and operational-dependency level, it is more accurate to say Cohere is a Canadian-headquartered foundation-model company in an American-dominated capital and infrastructure ecosystem.

Third, the absence question. Cohere is not in the Stanford Center for Research on Foundation Models' 2025 Foundation Model Transparency Index. The Index scored 13 major foundation-model developers in its December 2025 edition. Cohere, Canada's named national champion in the federal AI strategy, does not appear on the list. The Index doesn't say why a developer was or wasn't included; possible reasons include scale thresholds, model-availability criteria, or non-engagement with the assessment. The absence is itself worth tracking. The most-cited foundation-model transparency benchmark globally is currently not measuring the company Canada has named as its sovereign champion.

The fair read of Cohere is that it is a real Canadian foundation-model company with a genuine technical pedigree, currently positioned successfully in the enterprise-AI market, operating with substantial American capital and infrastructure dependencies, and not yet at the scale where the "national champion" framing fully corresponds to operational reality. Whether it gets there over the next five years is one of the most-watched questions in Canadian AI policy.

---

The telecom incumbents — Bell, Telus, Rogers

Canadian telecommunications incumbents occupy a structurally interesting position in the AI conversation. They are Canadian-controlled (in a way the foreign hyperscalers are not), operationally large (in a way Cohere is not yet), and politically embedded (in a way the research institutes are not). They are also subject to long-standing public scrutiny on competition, pricing, and consumer-protection grounds that pre-date the AI conversation.

Bell has positioned itself most directly as the AI for All infrastructure partner. The December 2024 commitment of CA$240M to Cohere's compute build-out, and Bell's separate planned 500 MW AI Fabric (with Cohere signed on as model partner since July 2025), are the highest-profile Canadian AI infrastructure investments. Bell's broader "sovereign cloud" framing markets Canadian-controlled data and AI infrastructure to federal, provincial, and enterprise customers on the basis that home-jurisdictional control is structurally safer for sensitive data than reliance on US hyperscaler infrastructure subject to the CLOUD Act and other US legal mechanisms.

Telus has positioned similarly, with extensive sovereign-cloud and AI-services marketing aimed at health-sector, government, and regulated-industry customers. Telus's environmental claims (renewable energy commitments, net-zero targets) and sovereignty claims (Canadian data residency, Canadian jurisdiction) are corporate self-disclosure, not independently audited at the depth they imply.

EXAMPLE — the maple leaf on the AWS box. "Sovereign cloud" sometimes means Canadian-built and Canadian-run. Sometimes it means a Canadian flag on infrastructure that is owned, operated, and ultimately subpoena-able by a company headquartered in Seattle. The sticker is the easy part. Ask what's underneath it.

Rogers has been less prominent in the AI conversation specifically, though Rogers's enterprise services and data-centre operations are substantial.

The general bias label for all three: Canadian-controlled corporate incumbents with substantial existing relationships to government as both regulator (CRTC) and customer, with strong commercial incentive to frame "sovereign AI" as something they are positioned to deliver. Their environmental and operational claims should be read as corporate self-disclosure subject to the same caveats the guide applied to Google's August 2025 inference paper: claims may be technically accurate within their stated scope, scope is chosen strategically, and the absence of mandatory third-party verification means the claims cannot currently be tested.

A structural observation worth foregrounding: the AI sovereignty framing aligns commercially with the Canadian telecom incumbents in ways the public conversation rarely names directly. Every dollar of enterprise or government AI workload that shifts from AWS or Azure to Bell or Telus's sovereign-cloud offerings is direct revenue. Every regulatory tilt that incentivizes data residency, Canadian-jurisdiction processing, or domestic-hosting requirements moves capital toward the incumbents. That is the commercial geometry, not an accusation. When Bell or Telus advocates publicly for stronger Canadian sovereignty requirements, they have a real public-interest argument and a real commercial interest at the same time. The guide asks readers to hold both rather than collapsing the analysis into one or the other.

ALIGNMENT — public interest, or revenue, or both at once? When Bell or Telus argues for stronger sovereignty rules, they can have a real public-interest case and a direct commercial stake in the very same sentence, because every rule favouring domestic hosting moves money toward them. Notice whose revenue the rule moves. You don't have to assume bad faith to keep one eye on the invoice.

---

The commissioned-research firm pattern

In the source library work that produced this guide, one specific structural pattern recurred enough to deserve its own treatment: economic-impact narratives in Canadian AI policy are increasingly being produced by commissioned-research firms working for the entities being measured. The guide flagged this in Chapter 1's bias methodology as a distinct category, "industry/corporate via commissioned-research firm," and Chapter 7 is where we unpack what that means in practice.

The clearest example is Public First, a UK-headquartered research-and-strategy firm with growing Canadian and North American footprint. Public First produced the "AWS Canada Impact Report" (commissioned by AWS, as the report's own pages state) which put AWS's contribution at $8.5 billion in economic value for Canada in 2021 and projected that fuller cloud adoption could add over $40 billion to GDP by 2030. (The harder infrastructure numbers that circulate alongside it — $1.4 billion invested over the first five years, 687 jobs, $21 billion by 2037 supporting 5,195 jobs and $39 billion in GDP — come from AWS's own self-published economic impact study, one rung less independent than commissioned research.) Either way, these are not independent measurements of AWS's economic impact in Canada; they are numbers produced for, or by, the company being studied.

The same Public First produced the projected "$27 billion productivity boost to Alberta's economy" figure cited in Amii's launch material for its Google.org-funded National AI Literacy Initiative, the federally-aligned literacy delivery vehicle inside AI for All. The same research firm is producing the economic-impact narrative for the foreign hyperscaler footprint and for the federally-funded literacy delivery vehicle.

None of this is corruption. Commissioned-research firms doing this work is a legitimate professional category; research-and-strategy firms have produced economic-impact reports for paying clients since well before the AI era. But it is also not independent measurement, and the line between "research firm" framing and "primary source" framing is one the press and policy debates routinely blur. When a Globe and Mail piece, a Government of Canada press release, or a hyperscaler's own marketing material cites the "AWS Canada Impact Report" as a source for jobs and GDP numbers, the reader is being given commissioned-research-firm output and may not realize it.

🧌 GOBLIN CHECK — When an economic-impact number sounds like it was written by the birthday boy's mom, check who paid for the cake. Public First is a real firm doing legal, disclosed work — and the same firm produced glow-up numbers for two foreign hyperscalers, one set of which the federally aligned institute then reached for in its own launch material. The goblin doesn't cry foul. The goblin staples the invoice to the finding and files them together, where they belong.

The guide's bias-mapping methodology asks readers to do two things when they encounter this category:

Notice the commissioning relationship. Is the report cited as if it were independent, or is the commissioning entity disclosed? If a number sounds remarkably favourable to one specific entity, check whether that entity paid for the research.

Hold the numbers carefully. Commissioned-research numbers are not necessarily wrong; Public First is a reputable firm and its methodologies are mostly defensible. But they are not the equivalent of Statistics Canada figures, peer-reviewed academic work, or independent regulatory analysis. They sit somewhere between corporate self-disclosure and independent measurement, and the placement matters.

A broader observation. The ecosystem of commissioned-research-firm production has grown substantially in the AI conversation specifically because the conventional empirical sources for economic-impact claims (Statistics Canada, the Bank of Canada, the Parliamentary Budget Office, peer-reviewed labour economics) cannot produce numbers fast enough or specific enough to support the rapid-cycle policy framing the field operates in. The strategy launch needs a $200B-growth number; Statistics Canada will not produce that number on demand. So the framing space gets filled by commissioned research, and the framing space then anchors the policy debate. The guide asks readers to notice the substitution.

---

The proxy-lobbying mechanism

The Canadian Union of Public Employees' March 2026 Senate submission flagged a structural concern that deserves its own treatment, because it names a specific mechanism the broader policy debate often treats as background rather than as a thing to track.

CUPE's submission documented that "technology corporations have actively lobbied against regulatory safeguards despite documented harms," citing a Policy Options analysis from November 2025 titled "U.S. Proxy Lobbying Threatens Canada's Digital Sovereignty." The mechanism is straightforward: large foreign technology corporations cannot directly lobby Canadian governments at the scale they lobby in the United States (Canadian lobbying registries, foreign-influence rules, and procurement integrity regimes constrain direct foreign influence). What they can do is fund Canadian industry associations, think tanks, and policy intermediaries that produce positions sympathetic to their interests, which then enter the Canadian policy conversation as Canadian voices.

CUPE also flagged a more specific federal mechanism: Budget 2025 included changes to the federal interchange program that embed up to 50 private-sector workers in government roles: workers paid by their home companies but seconded into federal positions, ostensibly to bring industry expertise into policy development. CUPE's critique: the program "grants tech companies insider influence over technology adoption" inside the federal apparatus.

Interchange programs have a legitimate function and have existed across multiple governments in multiple forms. But the AI-policy moment is one in which the federal government is making rapid, large-scale decisions about AI procurement, AI training data rules, AI infrastructure investment, and AI safety regulation, and the embedded-industry-workers structure means that some of the people in the rooms where those decisions are made are paid by the companies the decisions affect. Whether the structure should be modified, restricted, or expanded is contested policy. Whether the structure exists is empirical.

The proxy-lobbying mechanism connects to the commissioned-research-firm pattern in a way worth naming. If the policy framing is anchored by economic-impact numbers produced by commissioned-research firms working for foreign hyperscalers, and the policy development process embeds workers paid by tech companies inside the federal apparatus, and the major Canadian-controlled actors (telecom incumbents, foundation-model champion) have substantial American capital and infrastructure dependencies, then the federated picture is one in which "sovereign Canadian AI" is being built in an environment where foreign-aligned actors and arguments have substantial structural presence at multiple layers of the policy stack.

Read that as structure, not conspiracy: it is simply what the map shows once you lay out the lobbying registers, the commissioning relationships, the capital stacks, and the interchange-program postings. What the map shows is that Canadian AI sovereignty is being pursued in conditions that make full sovereignty extremely difficult to achieve operationally, regardless of whether the strategic intent is sincere.

---

The actor map — who's in a position to actually shape Canadian AI

Putting the four categories and the connecting mechanisms together, the practical answer to "who holds Canadian AI power" breaks into a smaller set of named actors than the category framing suggests. The chapter will close with the named-actor map.

On the foundation-model-developer side: Cohere (Aidan Gomez, Ivan Zhang, Nick Frosst) is the only Canadian player. The relevant foreign actors are OpenAI (Sam Altman), Anthropic (Dario and Daniela Amodei), Google DeepMind (Demis Hassabis), Meta AI (Yann LeCun, whose Canadian Turing-Award credentials make him a relevant Canadian-history figure even though he operates from New York). Geoffrey Hinton and Yoshua Bengio, the Canadian Turing laureates whose work the field is built on, are increasingly active as public voices on AI safety and governance rather than as company-aligned actors, and their public positions matter for Canadian policy in ways the strategy doesn't always foreground.

On the hyperscaler side: AWS (Andy Jassy), Microsoft (Satya Nadella), Google (Sundar Pichai). At the Canadian-operations level, AWS Canada, Microsoft Canada, and Google Canada have their own leadership; these positions are influential on procurement and government-relations specifically. The recent NVIDIA leadership (Jensen Huang) shapes infrastructure conversations because NVIDIA's chip pricing and allocation choices effectively constrain what Canadian AI infrastructure can be built at what speed.

On the Canadian-incumbent side: Bell (Mirko Bibic, CEO), Telus (Darren Entwistle, CEO until June 30 2026, then Victor Dodig), Rogers (Tony Staffieri, CEO). These three CEOs are direct counterparts to the foreign-hyperscaler CEOs in the political-economy of Canadian AI sovereignty.

On the research-institute side: Valérie Pisano (Mila), Cam Linke (Amii), Glenda Crisp (Vector). Increasingly direct voices on federal AI policy, with the bias caveat that their institutes are major AI for All beneficiaries.

On the federal political side: Mark Carney (PM), Evan Solomon (Minister, AI and Digital Innovation), Shafqat Ali (Treasury Board, federal AI use), Sean Fraser (Justice, deepfake regulation). The cabinet structure splits AI policy across at least these four ministers, which produces both intra-government policy coordination challenges and multiple access points for stakeholder advocacy.

On the civil society side: the BC + AI coalition (Kris Krüg and contributors); the Indigenous data sovereignty leadership (FNIGC, ITK leadership, the Abundant Intelligences research network); the union leadership (Sarah Ryan at CUPE on AI policy, Mark Hancock as CUPE national president, Sean O'Reilly at PIPSC, the ACTRA leadership on creative-sector AI questions); the academic voices (Carys Craig at Osgoode, Michael Geist at Ottawa, Suzie Dunn at Dalhousie, the Luccioni research network at Hugging Face Montréal).

One more name to track: Canada's big public pension managers. PSP Investments led Cohere's 2024 round; CPP Investments' own holdings disclosures show its three largest foreign public-equity positions, as at March 31, 2025, were Apple, NVIDIA, and Microsoft (roughly C$15 billion across the three), and in late 2024 it took a 37.5% stake in a US$15-billion-plus joint venture with Equinix and GIC to build US hyperscale data centres for AI and cloud workloads. Canadian pensioners are economically exposed to the AI buildout on both sides of the border, regardless of where the AI for All strategy directs domestic public funds. Neither manager has published an explicit AI-investment framework as of this writing; the implicit exposure is substantial.

This map is more granular than the four-category abstraction but it's the operational reality. When AI for All commits to "Scaling Champions" or "Sovereign Foundations," the specific question for readers and citizens to ask is: which of these actors is the strategy actually empowering, and which is the strategy implicitly accepting as the constraining environment?

The competition question: is AI concentrating power, or spreading it?

The actor map answers "who's here." It does not answer the question a competition regulator asks, which is whether the structure of the AI market lets the powerful stay powerful. On that, the honest answer is that AI is doing both things at once, and which one wins is not yet settled.

The concentration case is strongest at the bottom of the stack. One company, NVIDIA, supplies somewhere around eighty to ninety percent of the chips AI is trained on, and more than ninety percent of training specifically. Three companies, Amazon, Microsoft, and Google, hold roughly sixty percent of the world's cloud capacity those chips sit in. And the firms with the chips and the clouds have bought their way into the firms that build the models: Microsoft into OpenAI, Amazon and Google into Anthropic, more than twenty billion dollars in combined investment that a US regulator's 2025 study warned can function like vertical integration, giving the cloud giants influence over their own customers' rivals and a window into their data. This is the "data flywheel" worry in concrete form: the players who already hold the inputs, the scale, and the capital are best positioned to keep them.

The democratization case is real too, and the guide has to say so. Open-weight models that anyone can download and run, from Meta, Mistral, DeepSeek, and Canada's own Cohere with its Aya family, lower the wall in front of new entrants. The cost of running a capable model has collapsed, by one Stanford measure roughly 280-fold in eighteen months. And in January 2025 a comparatively small Chinese lab, DeepSeek, released an open model that roughly matched the US frontier at a fraction of the cost, and wiped a record 589 billion dollars off NVIDIA's value in a single day, the market voting that the incumbents' moat was shallower than assumed. Cheaper inference and open weights genuinely do let a startup or a public agency build without renting permission from a hyperscaler.

Both can be true. The barrier to using AI is falling fast; the barrier to building the frontier is rising. For Canada the practical version of that tension is stark: the country has one scaled foundation-model firm, Cohere, and a startup economy that mostly trains on foreign clouds, which is the gap the sovereign-compute spending from Chapter 6 is trying to close. Whether a Canadian entrant can compete at the frontier is, on the evidence, structurally hard but not impossible, which is exactly the kind of contested question this guide refuses to resolve for you.

Into this, Canada's competition regulator has finally started moving, and it is worth tracking, because competition law is one of the few tools that bites on market power directly. The Competition Bureau ran a public consultation on AI and competition, publishing a discussion paper in 2024 and a "What We Heard" report in early 2025 that named the risks plainly: data and compute as barriers to entry, self-preferencing by platforms, algorithmic collusion, and deceptive AI marketing. A second 2025 paper went deeper on algorithmic pricing. The toolbox it would use was rebuilt at the same time: a wave of amendments to the Competition Act between 2022 and 2024 repealed the "efficiencies defence" that used to let large mergers off the hook, strengthened the rules against abuse of dominance, tightened deceptive-marketing and greenwashing provisions, and, from 2025, let private parties bring cases themselves instead of waiting for the Bureau.

The limits matter as much as the tools, and naming them is the point. The amended Act can reach a "hub-and-spoke" arrangement where competitors quietly align prices through a shared algorithm, but cartel and civil-agreement law still requires proving an agreement, a meeting of minds, which is precisely what purely autonomous algorithms coordinating on their own may never leave evidence of. The deceptive-marketing powers can in principle catch "AI-washing," the habit of stamping "AI-powered" on a product that barely is, but as of 2026 there is no Canadian enforcement case to point to, only the US regulators who have started bringing them. So read the Bureau's activity on the enforceability ladder this guide keeps climbing: the diagnosis is now sharp and the statute is stronger, but a sharpened tool that has not yet been swung is a promise, not a precedent.

---

Who's holding what

CHAPTER RECAP — you now have: - The four-category map of Canadian AI actors — foundation-model developers, foreign hyperscalers, Canadian telecom incumbents, federally-aligned research institutes — plus the commissioned-research-firm category that operates across all four. - The Cohere case study done honestly: a real Canadian foundation-model company with genuine technical pedigree, currently positioned successfully in enterprise markets, operating with substantial American capital and infrastructure dependencies, not yet at the scale where the "national champion" framing fully corresponds to operational reality, and currently absent from the most-cited global transparency benchmark. - The telecom incumbents (Bell, Telus, Rogers) read as Canadian-controlled actors with real public-interest sovereignty arguments and real commercial interests in the sovereignty framing, simultaneously. - The Public First commissioned-research pattern as a worked example: the same research firm producing economic-impact narratives for AWS's Canadian footprint and (commissioned by Google) the Alberta productivity figure that Amii's federally-aligned literacy initiative cites. - The proxy-lobbying mechanism and the federal interchange-program structure as concrete mechanisms by which foreign-aligned interests enter Canadian AI policy development, alongside the more direct industry-association and think-tank pathways. - The named-actor operational map across all the categories, including the public pension managers' AI exposure — PSP as Cohere's lead investor, CPP Investments' US technology holdings — as a structural commitment Canadian sovereignty advocates rarely engage directly.

The next chapter (Chapter 8) takes the landscape this chapter has mapped and stress-tests the strategy's environmental framing against independent measurement: the Luccioni/Strubell/Crawford Jevons-paradox critique, the 600%+ corporate-disclosure underreporting finding, and the Canadian-specific carbon arithmetic of building hyperscaler infrastructure in Alberta's gas-dominated grid versus Quebec's hydropower grid.

You can now read any future Canadian AI announcement with a structural map of who's actually in a position to deliver, constrain, or benefit from what's being announced. That map is more granular and more honest than the strategy framing itself provides, which is most of what the guide exists to do.

---

Bias label for this chapter: structural-political analysis of the Canadian AI corporate and institutional landscape. Author lean: critical-engaged with the asymmetries in how the federal strategy treats different actor categories; skeptical of commissioned-research-firm output cited as independent measurement; sympathetic to public-interest sovereignty arguments while attentive to the commercial interests that align with them; willing to name the proxy-lobbying mechanism and the embedded-industry-workers structure as features of the current Canadian AI policy environment rather than as anomalies. Corporate self-disclosure (Cohere, Bell, Telus, Public First reports) labelled and read accordingly. Civil society sources (CUPE, BC + AI) and academic sources (Stanford FMTI) treated as primary on their respective domains.

Primary sources cited or relied on in this chapter: Cohere corporate disclosures and December 2024 funding round coverage; Public First, AWS Canada Impact Report; Amii National AI Literacy Initiative launch material; CUPE Senate Brief (March 2026); Policy Options "U.S. Proxy Lobbying Threatens Canada's Digital Sovereignty" (November 2025); Stanford Center for Research on Foundation Models, 2025 Foundation Model Transparency Index; Vaswani et al., "Attention Is All You Need" (2017); BetaKit coverage of June 5, 2026 institute leader statements; Budget 2025 documentation on federal interchange program changes; ISED Voluntary Code of Conduct on Advanced Generative AI (September 2023) with original and subsequent signatory lists; Canadian AI company disclosures from Ada, Waabi, Coveo, and Sanctuary AI. Detailed citations in the Sources appendix.

---