By Jen Evans, Principal, Pattern Pulse AI and co-founder, Tech Reset Canada

It was an eventful two weeks in Canadian AI sovereignty. A major funding announcement for our biggest foundation model and a sovereignty influence shift. Pivotal House committee testimony on AI from labour leaders. Announcement of a seemingly unrelated sovereignty fund. Increased instability for two major players in our domestic ecosystem. The gap between a functional AI strategy and the active AI procurement supposed to be governed by it continues to expand, with no delivery date in sight. The moving parts add up to growing complexity and opacity.

Where the Series Left Off

The series began by arguing that sovereignty is not a label but an architecture: compute, capital, data flows, procurement rules, operational control, and failure contingencies. Paper 1 introduced the Sovereign AI Maturity Model and argued that Canada missed the chance to use domestic telecoms as sovereign compute anchors against the Japan/SoftBank counterfactual. Paper 2 argued that the AI infrastructure boom is a pricing problem disguised as a valuation story, with hyperscalers selling below cost on debt against assets that depreciate faster than they amortize. Paper 3 modeled four dependency triggers across three cost layers and four scenarios. The triggers were ownership, pricing, security and operational temperature, and capability degradation. The paper proposed five instruments to govern them: a Sovereign Exposure Registry, a Sovereignty Trigger Framework, a Contingency Architecture Requirement, a Tempo-Matched Monitoring Function, and a Provincial Sovereignty Backstop. Paper 4 looked at Palantir exposure in the ecosystem, and Paper 5 looked at the implications of a paper showing that any downstream functions of a frontier model would experience unexpected shifts and safety drift.

The Sovereign AI Architecture for Canada Visual Walkthrough rendered the dependency map as a network diagram. CoreWeave appeared as a recurring structural node across multiple Canadian chains. Chain 1, captioned “Canadian label, American plumbing,” ran from NVIDIA through CoreWeave to Cohere to the Cambridge data centre supported by $240 million in federal funding.

The series was designed to make the architecture visible before the federal AI strategy landed. The strategy is now eight months delayed. The last two weeks have turned the scenarios from theoretical exposure into live policy problems.

What Has Happened in the Last Two Weeks

Inside two weeks, four announcements landed across technology, funding, partnerships, and labour. Read together, they describe a sovereignty architecture restructuring in directions the federal AI strategy has not yet acknowledged.

Technology

Cohere is acquiring Aleph Alpha in a transaction framed publicly as a transatlantic sovereign AI merger. Reuters reports Cohere will own approximately 90 percent of the combined entity, with Aleph Alpha shareholders holding approximately 10 percent. Schwarz Group, Aleph Alpha’s major German backer, is investing $600 million in Cohere’s funding round. The combined firm will keep the Cohere name and Canadian headquarters. Cohere’s chief AI officer Joelle Pineau told The Logic it is premature to say whether the firm will integrate Aleph Alpha’s technology.

This is Scenario 2 from Paper 3 materializing in real time. Paper 3 described three variants of how Cohere’s role could change: collapse, pivot, or acquisition. The Aleph Alpha transaction is a hybrid, with Cohere as the acquirer and Schwarz Group taking a sovereignty-relevant capital position in the resulting entity. Three weeks after Paper 3 was published, the scenario it modeled is in motion.

The “Canadian-owned” claim is the sovereignty label. The Schwarz Group lead position in the funding round is the sovereignty-relevant fact. The transaction materially changes the ownership, customer, and capital context around Cohere, even if the headquarters and name remain Canadian. The four-trigger framework applied:

The ownership and influence structure has materially changed. Pricing exposure for Canadian customers, including the recent ISED deployment of Cohere’s North platform to 1,400 employees, is unaddressed. Temperature is unknown because the combined entity will operate under Canadian and German governance with Schwarz Group influence on security posture, data handling, and model behavior. Capability is pending because the integration of Aleph Alpha’s technology is not yet decided.

If Cohere is the federal government’s anchor sovereign AI bet, which is what the procurement footprint, the Cambridge funding, and Solomon’s Berlin appearance all imply, then the substance of the Aleph Alpha transaction is material to the national strategy.

Aleph Alpha brings a German operational footprint, European customer relationships, and the local technical staff that European procurement processes effectively require. The combined entity inherits sovereign-AI credibility in Europe that takes years to build independently. The $600 million from Schwarz Group is significant against Cohere’s previous $1.64 billion total raised, extends runway in a tightening capital market, and brings a major German conglomerate into the customer and capital base. Aleph Alpha’s existing relationships with the German military through NATO’s Maven framework give the combined entity a defence-adjacent pathway in two NATO countries. For Canada, Cohere is a stronger European competitor and a more capitalized Canadian-headquartered firm than it was a month ago.

The deal’s reach has clear limits. Neither company is a frontier-class lab. Anthropic, OpenAI, and Google operate at a scale of compute and research investment two orders of magnitude beyond what the combined entity can sustain. The federal government’s most demanding AI workloads will continue to depend on US frontier models. The Aleph Alpha transaction reduces Cohere’s dependence on US capital. It does not reduce Canada’s dependence on US frontier capability or US compute infrastructure.

Solomon’s “hyperscaler or hegemon” framing in Berlin positions the combined entity as the only non-American, non-Chinese option. Canada has neither a hyperscaler nor hegemonic posture, so the framing describes a choice Canada is not positioned to make. What Canada is actually choosing between is which foreign hyperscaler’s terms it accepts, what residual sovereign capacity it builds alongside that dependency, and how it governs the gap between the two. The Cohere-Aleph Alpha transaction is one input to that question. The Berlin framing pretended otherwise.

The actual analytical question is narrower: does the combined Cohere-Aleph Alpha entity have the capability, the capital, and the operational independence to serve as the technical foundation of a sovereign Canadian AI strategy in workloads that matter? On capability and operational independence, the answer is partial. On capital, the answer is improved but still small relative to the frontier labs. On strategic foundation for the Canadian government’s AI deployment, the answer depends on what the strategy specifies.

If the federal AI strategy is being built around Cohere as the Canadian champion, the strategy needs to acknowledge what the champion can and cannot do. A champion that depends on American compute, runs at the enterprise tier rather than the frontier, and now answers to a German lead investor on major decisions is a real but partial asset. The strategy has to be built around what the asset actually is.

Infrastructure

CoreWeave’s financial position has worsened. On April 17, the company priced $1 billion in 9.750% senior notes due 2031, near junk-bond territory, per CoreWeave’s own announcement. Fitch has referenced approximately $30 billion in fiscal 2026 capex against the prior year’s $14.9 billion. Public investor materials disclose a contracted revenue backlog in the range of $66 billion, Q1 revenue guidance of $1.9 to $2.0 billion, adjusted operating income guidance of zero to $40 million, and quarterly interest expense guidance of $510 to $590 million, which annualizes to over $2 billion. Total debt is reported above $11 billion. The Masaitis securities fraud class action remains pending. Q1 2026 earnings report is May 7.

The architectural risk Paper 3 identified is sharper than it was at publication. The infrastructure layer Canadian sovereign AI sits on is taking on more debt at higher rates with thinner margins to support a CapEx ramp that depends on contracted revenue continuing to materialize. The Cohere-Aleph Alpha transaction does not change Cohere’s compute dependency. The Cambridge facility is still operated by CoreWeave. Bell’s Saskatchewan facility is still built around CoreWeave and Cerebras compute. The Cohere of after the deal will run on the same American infrastructure layer.

Funding

The Canada Strong Fund does not name AI as a dedicated mandate category. On April 27, Mark Carney announced a $25 billion sovereign wealth fund modeled on Norway. The named sectors in the official announcement are clean and conventional energy, critical minerals, agriculture, and infrastructure. Some media coverage uses the broader term “technology,” but AI was not named as a dedicated sector or mandate category in the government’s announcement.

The structure is the one John Ruffolo had been advocating publicly since October 2025. Ruffolo’s mandate proposal explicitly included data, communications and AI; health care and biotech; mobility and transportation; defence and cybersecurity; and the financial system. The implemented mandate includes none of those except defence-adjacent infrastructure.

The exclusion is a banker’s read of what belongs in a sovereign wealth instrument. Carney spent his career in central banks and investment houses where sovereign-grade solidity is built around hard assets with predictable cash flows and long depreciation curves. Energy infrastructure, critical minerals, agricultural land, transport corridors. AI venture is the inverse: short hardware lifecycles, volatile pricing, infrastructure that depreciates faster than the debt against it amortizes, business models still in formation. Carney is making the case that AI does not match the instrument, not that AI does not matter. That is a defensible read for a banker positioning Canada as an investment haven at a moment when the long-standing financial partnership with the United States has become unreliable.

The analytical problem this creates is that the highest-growth, highest-leverage sovereignty play has been routed away from the country’s biggest new capital pool. Ruffolo’s response on the record: “the announcement made no mention of innovation or growth investing.” The shortfall, he said, is in “high-risk, high-reward sectors that require serious money and long time horizons. It’s placing the big bets in AI, quantum, these big, big things that take a lot of capital.” Lucy Hargreaves, CEO of Build Canada, the same organization that published Ruffolo’s blueprint in February, said the announcement is “a sovereign wealth fund in name only.”

Whether a different vehicle exists with the mandate, governance, and capital scale required to address what Ruffolo named as the big bets is the unanswered question. The federal AI strategy has not yet provided one.

Partnerships

Solomon was in Berlin blessing the Cohere-Aleph Alpha deal. AI Minister Evan Solomon stood next to Cohere CEO Aidan Gomez at a press conference in Berlin, called Cohere “an essential tech company in the world,” and used the phrase “either a hyperscaler or hegemon” to describe the choice the merged entity is meant to provide an alternative to. The deal was framed as proof of the Canada-Germany Digital Alliance.

Germany has, for now, drawn a sovereignty line around Palantir in military AI procurement. On April 28, Vice Admiral Thomas Daum, Germany’s highest-ranking officer in cyber and information space, told Handelsblatt that the Bundeswehr will not award contracts to Palantir for now. He cited data sovereignty and the firm’s operational model. “Granting industry staff access to the national database is inconceivable.” Three European alternatives were shortlisted: Almato and Orcrist from Germany, ChapsVision from France. Defense Minister Boris Pistorius had previously raised concerns about Peter Thiel’s minority stake in German drone manufacturer Stark Defense.

The Pentagon, in the same period, formalized Palantir’s Maven program as a permanent program of record under a 10-year framework agreement valued at up to $10 billion. Germany declined to award Palantir contracts at the exact moment the United States locked them in.

This is the architectural critique Solomon used in Berlin coming back from Germany. Germany made a sovereignty decision about an American AI vendor. Canada has not made the same decision about the same vendor, even though Palantir’s footprint inside the Canadian government runs through third-party procurement arrangements that make the actual deployment surface invisible. The “Palantir: Five Jurisdictions and One Vendor over 48 Hours” piece, published in April, mapped what could be reconstructed from procurement disclosures and partner announcements. The full picture remains unmappable without procurement reform.

The question this raises is direct. If Germany applied the scrutiny that produced the Bundeswehr decision, why has Canada not applied the same scrutiny? The answer cannot be that the procurement structure is opaque, because the opacity is the problem the scrutiny is meant to address.

Labour

Bell has cut unexpectedly, and Rogers has opened the door to deep workforce reductions. The company announced on April 27 that it is offering voluntary departure and retirement packages to roughly 12,500 employees, half of its 25,000-person workforce. The Globe and Mail reported it as the largest round of buyout offers in Canadian telecommunications in recent years. Excluded from the offer: MLSE staff, Sportsnet, on-air talent, and unionized workers. The buyout offer follows Rogers’ April 22 announcement of a $1.2 billion reduction in 2026 capital expenditure, a 30 percent year-over-year cut. CEO Tony Staffieri described the year ahead as a “flat” market. An offer to 12,500 people does not equal 12,500 cuts. The signal is the door opening at that scale.

Bell’s pattern is multi-year and severe. 4,800 jobs cut in February 2024. Net job losses across 2025. 690 management positions eliminated in November 2025. 60 more at Bell Media in February 2026, including 11 journalists. The three-year strategic plan announced in 2024 targets $1.5 billion in cost savings by 2028.

Bell’s cuts and Rogers’ buyout offer point in the same direction: telecom cost reduction at the very moment telecom capacity is being rhetorically repositioned as sovereign AI infrastructure.

The structural advantage Paper 1 named, that Canadian telcos already controlled the networks, already served government, and already operated under domestic jurisdiction, was the foundation of the SoftBank-equivalent sovereign compute play that Canada did not make. Bell and Rogers are reducing the workforce that would have built it, while the sovereignty branding continues. Bell AI Fabric is positioned as Canada’s sovereign infrastructure layer. Whatever AI Fabric delivers operationally has to be delivered by the people Bell has not cut yet.

The supply side of the sovereignty equation is contracting at the same moment the demand side is being announced. The federal government as the first customer of Canadian sovereign AI is buying from a market where the largest potential domestic suppliers are in cost-reduction mode and the most prominent Canadian AI champion is being recapitalized by German capital. Whether the strategy that lands acknowledges this, or treats Bell AI Fabric and Cohere as stable suppliers, is a measurable test.

The CUPE testimony at INDU on April 30 made this layer of the architecture explicit. Nathalie Blais told MPs that generative AI systems being adopted across Canadian telecom are mostly hosted south of the border, and that Canadian data flows through non-Canadian networks every time those systems are used. Unifor and the United Steelworkers added the workforce dimension. The unions described the same architecture from the labour side that Balsillie has described from the capital and IP side. The diagnosis is shared across labour, capital, and policy.

What the Federal Strategy Still Has to Address

The federal AI strategy has been delayed past every deadline it has set. Originally promised by the end of 2025. Solomon told the science and research committee in February it would launch “this quarter,” meaning Q1 2026. Q1 ended without release. The April 28 spring economic update revealed six pillars without releasing the strategy itself. Solomon’s office now says “very soon” and “shortly.”

The pillars use the language of the published architectural critique. “Building the Canadian sovereign AI foundation.” “Scaling Canadian champions.” “Building trusted partnerships and global alliances.” Whether the strategy that lands operationalizes the language or treats it as branding is the test.

The strategy operates on a premise that has not been stated openly: the Government of Canada is the first and anchor customer of whatever Canadian sovereign AI gets built. The premise is already operational. The Innovation Department is deploying Cohere’s North platform to 1,400 employees. CanChat runs on Cohere models. Shared Services Canada is procuring AI tools for cross-government deployment. Provincial governments are digitizing health records and benefits adjudication on AI infrastructure procured under existing rules. The federal government is buying AI now. The strategy that is supposed to govern how it buys is still in draft.

If the government is the first customer, every procurement decision is a sovereignty decision. Whose compute the government buys determines whose jurisdiction the data falls under. Whose models the government deploys determines whose terms of service govern how it operates. Whose orchestration layer the government adopts determines what visibility Canadian institutions have into how decisions get made.

The gaps in the existing procurement framework are documented. They have not been closed.

There is no automatic review trigger for ownership change at a foreign infrastructure provider. The Investment Canada Act review threshold for national security is discretionary, applies to acquisitions of Canadian businesses, and would not be triggered by a debt restructuring or a foreign acquisition of a US company that happens to operate Canadian facilities. CoreWeave’s debt position now makes this risk concrete.

There is no procurement requirement for infrastructure origin disclosure. A government department can deploy AI services without knowing who owns the underlying compute, where the data resides, or what jurisdiction governs the infrastructure. Defence procurement requires this. Critical infrastructure procurement requires this. AI infrastructure does not.

There is no sovereign exposure registry. The EU AI Act mandates registration of high-risk AI systems but tracks the application layer rather than the infrastructure layer. A registered system can be running on a US-owned hyperscaler subject to the CLOUD Act with no obligation to disclose that fact. Canada has the opportunity to build what the EU did not.

There are no data portability mandates and no tested migration paths. The Ontario provincewide medical records announcement specified no timeline, no funding, no infrastructure provider, and no contingency architecture.

There is no tempo-matched monitoring function. Quarterly reports to Parliamentary committees are inadequate when the underlying infrastructure can change ownership, pricing, security posture, or capability within days.

There is no labour governance infrastructure operating at the rate of AI deployment. PIPSC is negotiating AI clauses for roughly 20,000 federal IT workers against an agreement that expired in December 2025. CUPE has confirmed there is no single AI clause and that coverage is patchy and contested.

There is no resolution to the trade-treaty constraint. Any sovereign compute move that disadvantages US providers can be challenged under USMCA. The CUSMA review window opens in July.

What the Strategy Must Include

The five operational instruments proposed in Paper 3 are now sufficiently developed to be tested directly against the federal strategy as it is released.

A Sovereign Exposure Registry that records, for every public-sector AI deployment, the service involved, the provider, the compute owner, the governing jurisdiction, the relevant trigger exposures, the replacement path, and the institution responsible for monitoring it. Procurement-driven, not voluntary. The registry’s “Unknown” fields are procurement design gaps made visible.

A Sovereignty Trigger Framework that defines thresholds across the four trigger types and automatically initiates review when changes affect critical AI infrastructure. The CFIUS model applied to infrastructure dependency rather than acquisition alone. The Cohere-Aleph Alpha transaction would have triggered review under this framework. The CoreWeave debt restructuring risk is the textbook case the framework is designed to govern.

A Contingency Architecture Requirement that no AI-dependent government service can be deployed without a documented, tested migration path, a pre-qualified alternative provider, and data portability guarantees embedded in the original contract.

A Tempo-Matched Monitoring Function that operates at the speed of the infrastructure it monitors, with dedicated analytical capacity tracking the financial stability, policy changes, and technical performance of every infrastructure provider on which Canadian government services depend.

A Provincial Sovereignty Backstop that prevents initiatives like Ontario’s medical records digitization from proceeding without infrastructure sovereignty requirements, data portability mandates, contingency architecture, and a tested migration path embedded in design.

These instruments are derived from the four-trigger framework, the three cost layers, and the four scenarios laid out in Paper 3. They are testable against the strategy when it lands.

What the Architecture Tells Us Now

A striking convergence is now visible across institutional positions that do not normally align.

John Ruffolo from the capital markets side. Jaxson Khan and Sean Mullin from inside policy design. Jim Balsillie from CIGI and the historical analysis side. CUPE, Unifor, and the United Steelworkers from the labour side. The “Whose AI Runs the Government?” series from the architectural and operational side. The German government, in the Bundeswehr’s Palantir decision, demonstrating what applied scrutiny looks like.

The architecture is restructuring across all four dimensions simultaneously. Cohere is becoming a Canada-Germany hybrid with German capital influence. CoreWeave is taking on more debt at higher rates while serving as the structural backbone of Canadian sovereign AI infrastructure. The Canada Strong Fund has been operationalized for hard assets and excludes AI as a dedicated mandate category. The two largest Canadian telcos are reducing or offering to reduce their workforces while their sovereignty branding continues.

The federal AI strategy will land into this market. It will be measured against the documented record of what the architecture actually does. The strategy’s language will be tested against that record. The five operational instruments are the test. The Sovereign Exposure Registry, the Sovereignty Trigger Framework, the Contingency Architecture Requirement, the Tempo-Matched Monitoring Function, and the Provincial Sovereignty Backstop are now on the public record.

If the strategy treats sovereignty as a label affixed to whatever ends up located in Canada, the gaps stay open and the architecture continues to restructure without governance. If the strategy treats sovereignty as a property of who owns the compute, who governs the data flow, who controls the capital, which suppliers can sustain the infrastructure, and which institutions monitor for trigger events, the gaps begin to close.

The experts are testifying. The transaction has been blessed. The fund has been announced without its most consequential sector. The infrastructure provider is reporting earnings on Thursday. Bell is cutting and Rogers has opened the door. The strategy is still in draft.

Procurement is now well ahead of strategy. ISED is deploying Cohere’s North platform to 1,400 employees. CanChat runs on Cohere models. Provincial governments are digitizing health records on infrastructure procured under existing rules. Each of these decisions is locking in trigger exposures, jurisdictional dependencies, and migration costs that the strategy, when it lands, will have to govern after the fact rather than before. The longer the strategy stays in draft, the more of the architecture is being built on procurement rules that predate the sovereignty framework the strategy is supposed to introduce. The deployments now in production were specified, contracted, and rolled out without any of the five instruments this analysis identifies as necessary. They will continue to operate under those original terms unless the strategy retroactively imposes new ones, which is a harder political and contractual problem than imposing them at procurement. The cost of governing the architecture rises with every week the strategy is delayed.