AI Sovereignty Cascade: Cohere–Aleph Alpha, China's Manus Veto, Trump's Scrapped Order

Three news stories in thirty days. Each one significant on its own. Read together — placed on a calendar from April 24, 2026 through May 22, 2026 — they are the same story playing out in three jurisdictions, and they mark the moment that AI stopped being an open market and started being a sovereignty contest.

On April 24, Cohere and Aleph Alpha announced a roughly $20 billion merger to form a transatlantic "sovereign AI" stack, anchored by a $600 million commitment from the German retail conglomerate Schwarz Group and explicitly positioned as an alternative for nations and enterprises that no longer want their data, models, or critical workloads routed through American hyperscalers. Three days later, on April 27, China's National Development and Reform Commission formally blocked Meta's roughly $2 billion acquisition of the Chinese-origin agent startup Manus — the first state-level prohibition of an inbound AI acquisition by China, and the first time a major Chinese regulator has unwound an AI M&A deal that the parties had already partially integrated. And on May 21, after a series of late-evening phone calls from Elon Musk, Mark Zuckerberg, and David Sacks, Donald Trump scrapped a planned executive order that would have established a voluntary 90-day pre-release security review for the most advanced frontier models — the lightest-touch federal AI safety mechanism that has ever been on a US president's desk.

Three deals. Three jurisdictions. Three different surfaces — corporate strategy, foreign investment review, and US federal regulation. But the same underlying dynamic: each jurisdiction is now treating AI not as a technology sector but as a strategic industrial asset over which it must retain jurisdictional control. The April 24 to May 22 window is the inflection point where the geopolitical map of AI hardened from theory into operational reality.

This is the AI sovereignty cascade. It will define the next two years of enterprise AI strategy, cross-border M&A, and frontier-model competition more than any benchmark, model release, or funding round announced in the same period.

The Thirty-Day Cascade

The three events deserve to be laid out on a single timeline because the proximity is the signal. Each event individually could be read as a one-off. Together they are a coordinated pattern — not coordinated between the actors, who have no shared playbook, but coordinated by the underlying logic that each jurisdiction is now operating under.

The events are not connected by causation in the surface sense. Cohere's deal team was not coordinating with Beijing's NDRC. Musk was not on the phone with Trump because of the Manus block. But each actor is responding to the same structural condition: AI capability is concentrating in a small number of firms, the concentration is overwhelmingly American, the capability has clear national-security and economic-sovereignty implications, and every jurisdiction that is not the United States is now scrambling to ensure that it will not be permanently dependent on American AI infrastructure for the next thirty years.

The United States, meanwhile, is responding to the same condition by removing self-imposed constraints on its own AI sector, on the theory that any voluntary review regime is a competitive handicap in a race that the US is winning. That is the Sacks position, that is the Musk position, that is now — as of May 22 — the Trump position. The three events look unrelated because they sit on three different surfaces. They are not unrelated.

Event One: Cohere and Aleph Alpha — The Transatlantic Sovereign Stack

The Cohere–Aleph Alpha merger is the most consequential of the three events for enterprise AI buyers, because it is the first time a non-US AI provider has crossed the credibility threshold required to compete for serious sovereign workloads at scale.

The structure of the deal is worth dwelling on. Cohere, a Toronto-headquartered frontier-model company with a strong enterprise sales motion and existing deployments across financial services and defense, is the surviving entity. Aleph Alpha, a Heidelberg-based research lab with deep institutional relationships across German federal agencies and a multilingual European model family that has been the de facto European frontier model for several years, is the strategic asset. The combined company is anchored by a $600 million capital commitment from Schwarz Group — the parent of Lidl and Kaufland, one of the largest privately held corporations in Europe — and the combined valuation is roughly $20 billion. The merged entity is targeting the heavily regulated sectors where data sovereignty is not a marketing message but a legal prerequisite: defense, banking, healthcare, and the European public sector.

The number that matters is in the press release: the merged entity's addressable market for "sovereign AI" workloads — workloads where the purchaser requires that data, models, and inference all remain under their direct jurisdictional control — is roughly $600 billion of the $1 trillion total AI services market. The TAM is large not because every workload requires sovereignty, but because every workload in a regulated industry will be required, over the next three to five years, to demonstrate where its data lives and which jurisdiction can subpoena its model.

Sovereign-AI buyers historically had three options: build in-house (capital prohibitive), use a hyperscaler with a sovereign-cloud overlay (capability constrained, vendor-lock-in concerned), or use a regional provider that did not have a credible frontier model (capability gap unacceptable). Cohere–Aleph Alpha closes the third option's capability gap. Whether it closes the gap fast enough is the open question, but for the first time the question is open.

The Schwarz Group anchor matters more than the headline number suggests. Schwarz is not a tech investor. Schwarz is a $200 billion-revenue retail operator that has been quietly assembling its own European cloud and software stack — Schwarz Digits — under a strategy that explicitly assumes that dependence on US hyperscalers is an unacceptable long-term risk. The $600 million commitment is the first time that strategy has crossed the frontier-model line. Schwarz's commitment signals to the rest of European enterprise that the buyer demand for a sovereign stack is real, capitalized, and durable. That signal is what Cohere–Aleph Alpha was unable to send on its own.

The 78% US-hyperscaler concentration in EU enterprise AI workloads — even in regulated sectors — is the gap that Cohere–Aleph Alpha is now positioned to attack. The merged entity does not need to take all of that share. It needs to take enough of the regulated-sector share to validate the sovereign-stack thesis, and then the rest of the EU industrial base will quietly migrate over the following 18 to 36 months. That migration, if it happens at scale, is the single largest cloud-and-AI revenue re-allocation since AWS won the early enterprise cloud wave.

Event Two: China Vetoes Meta–Manus — The First Inbound AI Block

Three days after the Cohere announcement, on April 27, China's National Development and Reform Commission formally blocked Meta's roughly $2 billion acquisition of Manus. The block ended a months-long probe that began when Meta announced the deal in December 2025 — a probe that was, in the early months, widely expected to result in a structural-remedy approval rather than a formal veto.

The structure of the Manus situation is unusual enough that the precedent matters more than the deal value. Manus was founded in mainland China by an agent-systems research team that built an early autonomous agent that briefly captured global attention in early 2025. The company moved its operational headquarters from Beijing to Singapore in July 2025 — a move that the company publicly described as a localization decision but that the AI industry read as a preparation for an inbound acquisition by a US frontier lab. Meta announced the acquisition in December 2025. Chinese regulatory authorities — the NDRC, the State Administration for Market Regulation, and the Cyberspace Administration of China — opened a coordinated review in January 2026. The review took roughly four months. The veto came on April 27.

The substantive Chinese argument was that Manus, despite its Singapore headquarters, still relied on Chinese talent, Chinese training methodology, and Chinese-origin agent-research IP, and that those assets were strategic enough that they could not be permitted to migrate to a US frontier lab without Chinese consent. The procedural argument was that under China's foreign investment review framework, the agency had jurisdiction over the transfer of any technology that the state classified as strategic, regardless of where the acquiring or target entity was incorporated.

The implementation problem the veto creates is that Manus engineers have already moved. The deal was partially integrated. Meta's agent team has been absorbing Manus IP and personnel since January. The "unwind" remedy that the NDRC ordered is, in practical terms, impossible — you cannot un-hire the engineers and you cannot un-transfer the methodologies they brought with them. What the unwind order will produce in practice is a long-running compliance proceeding, an exit of Manus capital backers (Tencent, HongShan) from Meta's cap table to the extent they hold any, and a slow accommodation between Meta and Beijing that will probably resolve into a sub-scale technology-licensing arrangement rather than a full reversal.

The precedent the Manus veto sets is broader than the deal. China has now explicitly asserted jurisdiction over AI-IP migration across its borders, even when the company in question has relocated. The signal to every Chinese-origin AI startup is that the option of "move to Singapore, then sell to a US lab" is no longer reliably available. The signal to every US frontier lab is that acquiring Chinese-origin AI assets — even indirectly — now carries a months-long regulatory exposure that may end in a forced unwind. The signal to every cross-border AI M&A practitioner is that the deal calendar now includes a Chinese-foreign-investment review window by default, regardless of where the target company appears to be incorporated.

That is a meaningful constraint on AI M&A capital flows. The next three to four quarters will reveal how meaningful — whether US labs continue to acquire adjacent-Chinese-origin AI startups under heavier structural remedies, or whether the pool of acquirable AI talent now bifurcates along jurisdiction-of-origin lines.

Event Three: Trump Scraps the Voluntary Safety EO

The third event closes the cascade in the unlikeliest jurisdiction. The Trump administration had been preparing, for several weeks, a voluntary AI safety executive order intended to address the cyber-capability concerns raised by Anthropic's Mythos model release earlier in the spring. The order was the lightest-touch federal AI safety framework that has ever been seriously considered by a US administration. It would have established a voluntary mechanism — voluntary — for AI developers to submit their most advanced models to federal agencies for a 90-day security review prior to public release. No licensing regime. No mandatory hold periods. No statutory authority claim. A voluntary review process with no enforcement mechanism beyond reputational pressure.

On May 21, between Wednesday evening and Thursday morning, Trump received phone calls from Mark Zuckerberg, Elon Musk, and David Sacks. The Sacks argument — articulated in interviews after the fact — was that the voluntary review would, in practice, become a de facto licensing regime, because no AI lab would risk releasing a frontier model without going through the voluntary process, and the voluntary process would predictably grow in scope over time. The Musk and Zuckerberg arguments tracked the Sacks framing. Trump pulled the order Thursday morning. The framing in the Oval Office afterward: "we're leading China, we're leading everybody, and I didn't want to do anything to get in the way of that lead."

The substantive content of the scrapped order matters less than the demonstration effect. The US has now publicly demonstrated that any federal AI safety action, even one structured as voluntary and even one driven by a genuine cyber-capability concern from a major lab, can be rolled back by a small number of phone calls from frontier-lab principals. That demonstration — that there is no federal AI floor that the US will not lower in the name of "the lead" — is the signal that the rest of the world is now reading.

The chart is illustrative — the underlying frameworks are heterogeneous and no single score captures them well — but the relative ordering is the relevant signal. After May 22, the US sits at the bottom of the regulatory-floor ranking among major AI jurisdictions. The EU and UK are above it. China, through both its existing Generative AI Measures and its 2026 amendments to the foreign-investment review framework that produced the Manus veto, is well above it. Singapore, where Manus relocated and from which a credible sovereign-AI position is now plausible, is above it.

For the first time, the US is now the lightest-touch major AI jurisdiction. That is what "we're leading" means in practice in May 2026.

The Pattern: Three Surfaces, One Dynamic

The cascade is legible only if you read the three events as the same dynamic expressed in three different policy surfaces.

Cohere–Aleph Alpha is the M&A surface. Non-US capital and non-US enterprises are assembling a non-US-controlled frontier-model option, because the assumption that US-controlled AI will be safely usable by non-US enterprises has eroded to the point where a $20 billion transatlantic merger anchored by a single-family German retail conglomerate is the rational hedge.

The Manus veto is the regulatory surface. State actors are now claiming jurisdiction over AI-asset transfer, on the theory that AI capability is a strategic national asset that cannot be allowed to migrate to a competitor jurisdiction without state consent. The veto is not unique to China. The EU's foreign-subsidies regulation and the US's CFIUS framework will produce equivalent vetoes on inbound and outbound AI M&A in the next 12 to 24 months.

The scrapped EO is the industrial-policy surface. The United States has decided, at the highest level, that any constraint on its own AI sector is a competitive disadvantage in the strategic AI competition, and that the appropriate response is to remove constraints — including voluntary ones — rather than to negotiate them with other jurisdictions. That decision sets the US position for the rest of the Trump administration: the lead matters more than the floor.

The three surfaces are equally weighted because each one is what each actor controls. Schwarz controls capital and can assemble a sovereign stack through capital deployment. Beijing controls foreign-investment review and can use that control to block inbound capability migration. Trump controls executive-branch regulatory posture and can use that control to remove self-imposed constraints on the domestic sector. Each actor is using the surface available to them. The three surfaces add up to a single, coherent strategic shift.

That is why the cascade is best understood as a single event in three places, not as three coincident events.

What the Cascade Means for Builders

The implications for enterprise AI buyers and AI-product builders are concrete enough to be operational, not just rhetorical.

For US enterprise buyers, the implication is that hyperscaler dependency is now a geopolitical exposure that must be measured. The risk is no longer "vendor lock-in" in the conventional sense — it is the risk that the hyperscaler's home-jurisdiction regulatory posture becomes a discriminatory factor in cross-border data flows, in subsidiary operations in regulated jurisdictions, and in M&A activity. Enterprise architecture decisions that were made in 2023 and 2024 under the assumption of a stable global cloud substrate need to be re-evaluated against the assumption that the substrate is now jurisdictionally fragmented.

For EU enterprise buyers, the implication is that Cohere–Aleph Alpha is now the first credibly capitalized non-US frontier-model option for serious sovereign workloads. The question is no longer "is there an alternative" — it is "does the alternative get to capability parity fast enough to capture the migration window." Buyers should expect to be approached by sovereign-stack sales motions in Q3 and Q4 2026. The deals will be large, the architecture discussions will be substantive, and the buyer-side leverage will be larger than it has been in any prior AI procurement cycle.

The migration is not uniform across workload types. Customer-facing chat will remain primarily on US hyperscalers because the data sensitivity is low and the capability gap is high. Defense, healthcare, and public-sector workloads will migrate quickly because the data sensitivity is high and the regulatory floor in the EU is hardening. Code-generation, internal RAG, and financial analytics workloads will sit in the middle, and the actual migration rate will depend on Cohere–Aleph Alpha's capability trajectory and on whether European supervisory authorities formalize a "sovereign by design" preference in their procurement guidance. The likely answer to both is "yes, by 2027."

For multinationals operating across both blocs, the implication is a two-stack architecture is now a strategic necessity rather than a strategic option. Workloads that can serve EU citizens will need a non-US-controlled inference path. Workloads that touch Chinese-origin data will need a separate inference path that is auditable by Chinese authorities. The CFO conversation about AI infrastructure cost has just become a CISO conversation about architectural redundancy, and the cost premium for full sovereignty-compliant deployment is going to land somewhere between a 15% and 40% uplift on the single-stack equivalent over the next two years.

For startups, the implication is that the regulatory map matters as much as the product map. A US-founded AI startup that wants to sell into EU regulated sectors must now design its data architecture and its model-hosting strategy around sovereignty constraints from day one. A Chinese-origin AI startup that wants to sell into a US frontier lab now needs to plan for a multi-quarter regulatory exposure on any acquisition path. The optimal startup strategy depends heavily on jurisdiction of incorporation, jurisdiction of customer base, and jurisdiction of training data provenance — three variables that were second-order considerations in 2023 and that are now first-order.

The Compute and Talent Implications

The cascade is going to reshape the compute and talent markets within the next two to four quarters, on top of the M&A and procurement implications.

Sovereign stacks need sovereign compute. Cohere–Aleph Alpha's frontier-model training cannot run on a US hyperscaler if the sovereignty value proposition is to hold. The merged entity will need EU-resident or EU-controlled compute at scale — meaning that the European hyperscaler capacity that Microsoft and AWS have been building under sovereign-cloud overlays will be competing with genuinely European cloud capacity (Schwarz Digits, OVH, Scaleway, T-Systems) that can serve the same workloads with a stronger sovereignty claim. That shift will pull a meaningful share of frontier-training capex toward European hyperscaler capacity, with a corresponding investment cycle in EU-located GPU clusters.

The talent implications are sharper. AI engineers who have spent the last decade optimizing for compensation in a global market will now optimize for a jurisdictional market. Engineers with EU work authorization and willingness to work on EU-resident systems become structurally more valuable to Cohere–Aleph Alpha, to Schwarz Digits, and to the cluster of European frontier-AI startups that will form in the wake of the merger. Engineers who have worked on Chinese-origin agent systems become structurally less hireable by US frontier labs, given the Manus precedent. The talent market is fragmenting by passport and by training provenance in a way that it has not since the 1980s.

The Historical Comparison That Actually Fits

The cascade is best understood not by comparison to the cloud-computing era, which was an open-market expansion from roughly 2010 through 2022, but by comparison to the telecommunications and semiconductor industrial-policy era of the 1980s and 1990s. The structural condition is the same: a strategically critical technology, concentrated in a small number of national-champion firms, around which jurisdictions are unwilling to accept indefinite dependence on a foreign competitor.

The telecommunications example is the closer historical analog because the network-effects and standards-control dynamics map cleanly onto the foundation-model situation. AT&T's pre-divestiture dominance of US telecommunications produced a similar regulatory anxiety in Japan and Europe that drove the formation of national champions — NTT, Deutsche Telekom, France Telecom — and an aggressive industrial policy aimed at ensuring that the national telecommunications infrastructure would not be controlled by a single American firm. That industrial policy worked. By the early 2000s, the European and Japanese telecommunications sectors were genuinely competitive with the American sector. The cost was tens of billions of dollars in subsidies and regulatory protection sustained over twenty years.

The semiconductor example is the cautionary tale. Japan's MITI-led industrial policy in the 1980s successfully produced a national champion sector that briefly dominated DRAM. The dominance proved brittle. By the late 1990s, the Japanese DRAM industry had been hollowed out by Korean and Taiwanese competition, and the industrial-policy investment did not pay back the strategic dividend the architects had projected. Sovereign-AI proponents need to take the Japanese DRAM trajectory seriously — capability parity is necessary but not sufficient, and the long-run winners may not be the firms that the industrial policy initially backs.

The reading the cascade most clearly contradicts is the post-Cold-War-trade-liberalization reading of technology policy that has been the implicit assumption of US AI strategy through 2024 — that capability would diffuse, that markets would integrate, that the relevant boundary was the firm rather than the jurisdiction. The April–May cascade is the moment that reading became unsustainable as a basis for enterprise planning. The jurisdictional boundary is now the relevant boundary. Capability diffusion will be slower than the open-trade reading predicted. And the relevant unit of analysis for AI strategy through the end of the decade is the national-bloc industrial stack, not the global product.

The Counterargument

The case against reading the cascade as a structural shift is worth taking seriously, because the case is real.

The Cohere–Aleph Alpha merger could fail. Either the merged company fails to ship a credible frontier model in the 12 to 18 month window that the sovereignty thesis requires, or the Schwarz commitment proves to be a one-off rather than the leading edge of broader European industrial backing. European sovereign-AI projects have a long history of failing to clear the capability gap. Aleph Alpha specifically was, prior to this merger, the most visible example of that history.

The Manus veto could be narrow rather than broad. The substantive Chinese argument relied heavily on Manus's specific Chinese-origin posture, and the veto may not generalize to AI deals where the target is more cleanly non-Chinese in origin. The implementation problem — un-doing a partial integration — may produce an accommodation rather than a precedent.

The Trump EO reversal could be re-reversed. The Anthropic Mythos cyber-capability concern is real, the political pressure for some form of federal action will grow as more capable models ship through the rest of 2026, and a future EO that addresses the same concerns in a different structural form is still on the table.

Any one of those three failure modes would weaken the cascade reading. All three would invalidate it.

The base case, however, is that none of the three fully unwinds. Cohere–Aleph Alpha may not capture as much of the sovereign-AI TAM as the most optimistic projections suggest, but it will capture enough to validate the thesis and attract follow-on capital. The Manus veto may produce an accommodation, but the underlying regulatory authority is now exercised and will be exercised again. The Trump reversal may eventually be revisited, but the demonstration effect — that the US will not maintain even voluntary safety floors against sustained industry pushback — is already on the record and will shape non-US jurisdictional expectations for the rest of the administration.

What to Watch Through Q4 2026 and 2027

The cascade is a starting condition, not an end state. The next 12 to 24 months will reveal which of three trajectories the AI sovereignty story takes, and the early signals are visible in late 2026 and early 2027.

The first trajectory to watch is whether other major economies announce their own sovereign-AI capital commitments at the Schwarz scale. The UK has been building a national AI capability through the AI Sandbox and a series of public-private compute commitments. France has Mistral, which has been sub-scale relative to the sovereignty thesis but which now has a credible demonstration that the buyer demand exists. Japan, Korea, and India have all made noises about national AI infrastructure. Watch for at least two of those to translate noise into a Schwarz-equivalent capital commitment by the end of 2026.

The second trajectory is whether the Chinese veto framework is replicated. The EU's foreign-subsidies regulation is the obvious vehicle. Watch for an EU review action against a US-led acquisition of a European AI startup in Q3 or Q4 2026. If that happens, the cross-border AI M&A market enters a new regime in which every meaningful deal carries a multi-jurisdiction review exposure that did not exist a year earlier. CFIUS will become a more active gatekeeper for inbound foreign AI investment into US targets in parallel.

The third trajectory is whether the US position holds or whether Congress moves where the executive has retreated. Congressional movement on AI is unlikely in 2026 on existing patterns, but the Mythos-class cyber-capability concern is not going away. The path most likely to produce action is a major cyber incident — credible threat or actual breach — attributable to or mediated by a frontier model. If that happens, the Sacks/Musk veto power over voluntary safety frameworks weakens, and the prior frameworks return in a harder form.

The Bottom Line

April 24 to May 22, 2026 was the period in which AI sovereignty stopped being a thesis and started being a fact. Three jurisdictions, three policy surfaces, three events. Read individually each is a news story. Read together they are the moment that the geopolitical map of AI hardened from possibility into operational reality.

Enterprise AI buyers should re-evaluate their architectural assumptions. Multinational AI strategy teams should plan for two- or three-stack deployment. Cross-border M&A practitioners should price in multi-jurisdiction review windows on every meaningful AI deal. US frontier labs should expect a reduced acquisition pool of non-US AI assets. Non-US frontier labs should expect the buyer demand to be there for the first time in this cycle.

The next eighteen months will tell us how far the cascade runs. The base case is that it runs further than the most cautious readings suggest, because the underlying conditions that produced the April–May cascade are structural rather than cyclical. Capability concentration in a small number of US firms is not reversing in 2026. The strategic value of AI is not declining in 2026. The willingness of non-US jurisdictions to accept indefinite dependency on US-controlled AI is declining quickly. Those three conditions, together, produce more sovereignty-cascade events, not fewer.

The Cohere–Aleph Alpha merger, the Manus veto, and the scrapped Trump EO are the opening of the cascade, not the conclusion. The next year of AI strategy will be about what comes after.

Further reading

• Three-speed AI governance — the EU AI Act simplification, CAISI, and the UK sandbox — the regulatory baseline the cascade is now hardening.
• AI labs absorbing systems integrators — OpenAI deployment, Anthropic, Goldman, Blackstone — the parallel restructuring on the enterprise-services side of the same dynamic.
• Anthropic, SpaceX, and the Colossus compute deal — the compute layer that any sovereign-AI stack will have to either replicate or rent.
• My prediction: sovereign-AI infrastructure becomes a global regulatory mandate by mid-2027 — the falsifiable claim that the April–May cascade is now visibly validating.