By: Lisa Damast
Originally published: June 26, 2026
Last updated: July 1, 2026 (to reflect subsequent developments involving Anthropic, OpenAI, and related legal proceedings)
This piece draws on interviews with IronSpark’s analyst team who cover open source and responsible AI, DevOps, data and analytics, and workforce, alongside Chief Analyst Salvatore Salamone, conducted in the days following the events described below.
Takeaways
- A US export-control directive forced Anthropic to take its most advanced AI model offline worldwide on June 12, with no warning and no date for restoration
- JPMorgan and Goldman Sachs have cut Claude access for Hong Kong staff over licensing terms that exclude Greater China; Morgan Stanley has taken a different approach
- The EU, South Korea, and Canada built sovereign AI infrastructure over the past year, well before the shutdown made the case for them
- Just 37% of CIOs say they have full visibility into the AI tools already in use across their organization, according to Logicalis’s 2026 Global CIO Report
On the evening of Thursday, June 12, at 5:21 PM ET, a cutting-edge AI model simply went dark. Not crashed. Pulled by a government export control directive whose justification Anthropic disputed from day one, and which has shifted at least twice since. Anthropic had no choice but to disable its most advanced models, Fable 5 and Mythos 5, for every customer worldwide, because it had no way to segment foreign-national access on short notice.
It was, by many accounts, the first time a leading AI company had taken a publicly deployed frontier model offline due to a federal directive, but it wasn’t the first sign something had shifted. Earlier this year, a US-sanctioned International Criminal Court judge found her Apple ID, iCloud, Amazon, Airbnb, and PayPal accounts blocked overnight, with no advanced warning. The mechanism was the same in both cases: access to infrastructure modern business runs on, switched off by administrative decision, instantly, with no visible appeal.
Governments have spent the past year or more building sovereign AI capacity, because they already recognized the risk of depending on foreign-controlled technology. Most enterprises have not started preparing for that same risk, a risk that should worry CTOs and CDOs.
“The biggest thing to consider is that access to AI models and tools can be instantly cut off at the whim of an administration,” says Salvatore Salamone, IronSpark’s chief analyst. He notes that several observers have questioned whether the stated rationale for the shutdown matches the actual risk, suggesting the action was as much punitive as protective. Either way, he argues, “sovereign alternatives are critical moving forward.”
Not every analyst reads the moment as a five-alarm fire. IronSpark’s Hardware and Open Source analyst Gordon Haff is more measured. “Most enterprises are not sufficiently locked into AI to the degree some are locked into AWS. I’m not sure AI is all that different from understanding dependencies on CRM and other systems,” he explains.
Whether June 12 was a uniquely AI problem or a familiar vendor-dependency problem with a new label, every technology leader now has to answer the same question: what happens the day a system you depend on simply isn’t there, and did you choose that risk, or did someone else choose it for you?
Most haven’t answered it. A global Logicalis 2026 survey of more than 1,000 CIOs found that 16% have no continuity plan if a key AI provider becomes unavailable, only 37% say they have full visibility into the AI tools and services already in use across their organization, and 62% admit they’ve already compromised on AI governance standards because they didn’t have the knowledge or capability to do otherwise.
June 12 didn’t come out of nowhere. In early March 2026, the Pentagon designated Anthropic a “supply chain risk,” a label previously reserved for foreign adversaries. “Even though these services are billed as sovereign, it does not matter if a country or person crosses a current administration,” Salamone says, pointing to the ICC sanctions as the clearest illustration. “A similar thing could happen to a foreign government or non-domestic enterprise.”
The concentration behind that risk is stark. According to Epoch AI‘s tracking of global GPU cluster performance, the US and China together host the overwhelming majority of the computing power needed to train and run frontier AI models, about 90% combined, per analysis cited in the Center for a New American Security’s Sovereign AI Index. The same index notes that, per LMArena’s model leaderboard, those two countries’ labs own all 50 of the top-ranked AI foundation models.
What Just Happened
Regardless of Fable 5, some organizations are beginning to address concerns. On June 18th, the Financial Times broke that JPMorgan Chase removed Claude from its approved tools for Hong Kong staff, following Goldman Sachs in April. Both cited Anthropic’s licensing terms, which exclude Greater China, per Reuters.
As Theo Lau, founder of Unconventional Ventures, told American Banker, the move, combined with the Fable and Mythos shutdown, is “an early indicator of an increasingly fragmented AI world with three blocs: U.S.-aligned, China-aligned, and a hedging middle,” meaning AI vendor selection “is no longer just a tech decision, but a geopolitical positioning one.” Bank of America Europe board member Oliver Bussman read the same move differently. It’s less a sign of fragmentation than of institutions “moving first on containment rather than waiting for regulators to force the issue,” he said in a LinkedIn post, the publication reported.
Morgan Stanley took a different approach to managing the same broader category of US-China jurisdictional risk. The bank has issued more than 300 of its Hong Kong investment bankers separate, restricted-use devices, iPhones and iPads stripped down to email and meeting apps, specifically for trips into mainland China, where cross-border data rules are far stricter than in Hong Kong itself. Goldman and JPMorgan, by contrast, haven’t adopted a device policy like it.
Three banks. Different controls. No shared sovereign AI playbook.
None has publicly articulated a formal sovereign AI strategy, yet all are grappling with how to operate across jurisdictions where technology, regulatory, and security requirements increasingly diverge. Their responses are being implemented through existing frameworks for risk, compliance, and operational resilience rather than dedicated sovereignty programs.
In that sense, sovereignty is being addressed more as a constraint than a strategy. The institutions with the greatest exposure are encountering these questions first. The rest of the enterprise market may not be far behind.
Governments Got There First
While banks are beginning to address AI dependency through governance and vendor controls, governments are addressing it through sovereign infrastructure and domestic AI ecosystems. Both are asking the same underlying question, how much control do you have over a capability that has become strategically important to your operations?
Governments are further along, and have been for a year. The European Commission unveiled a sovereignty package, estimated, as reported by Fortune, at €422 billion, on June 3, with Commission President Ursula von der Leyen warning that Europe “cannot afford to depend on others for the technologies that keep our hospitals running, our energy grids stable and our services secure.” South Korea’s NAVER and NVIDIA expanded sovereign AI infrastructure on June 7. Canada released a national AI strategy on June 4. These initiatives were already underway before June 12 because governments had concluded that dependence on externally controlled AI systems represented a strategic risk well before that shutdown made the case for them.
Even committed advocates disagree on the path forward. Economist Cristina Caffarra, per the Fortune article, called the EU package “very feeble,” arguing that only a small portion of cloud contracts would meet a meaningful sovereignty standard. Oxford’s Sandra Wachter, per the article, argues Europe should not try to match American frontier-model scale at all, suggesting smaller, more efficient models may provide sufficient capability without triggering an infrastructure arms race.
Wharton’s Jonathan Iwry raised a different concern, noting that while restrictions may advance short-term policy goals, they could also encourage allies to reduce their dependence on American AI providers over time, Fortune reported.
The sovereignty question does not yet have a single answer. As such, the most valuable capability may be foregoing a permanent strategy and instead building the flexibility to reassess the tradeoffs as technology, markets, and geopolitics continue to evolve.
What “Sovereign AI” Actually Means
Sovereign AI is an organization’s, or a country’s, ability to develop, host, deploy, and govern AI systems using controlled data, infrastructure, workforce, and ecosystems, rather than relying on external providers who can make key decisions without your consent.
Two misconceptions dominate. First, it doesn’t mean building your own frontier model. Fine-tuned models on infrastructure you control can achieve real sovereignty without replicating the largest labs’ R&D budgets. Second, global cloud providers aren’t automatically incompatible with sovereignty. Many hyperscalers now offer sovereign cloud with isolated infrastructure and local data residency. Whether that’s enough depends on whether you believe, after June 12, that US law can still reach into those arrangements regardless of where the servers sit.
Open-source models are one real path. “Open-weight models let you hone LLMs down to just the parameters your solution requires… and insulate you from policy and regulation changes imposed by a vendor’s country or region,” explains IronSpark’s open source and responsible AI analyst, Elisabeth Strenger.
But one question determines whether a CTO is actually solving the problem rather than relocating it. “Do you have the expertise to tune models and integrate them in-house? If not, you are going to be dependent on consulting firms,” she warns.
Asked which of the four sovereign AI pillars – data, infrastructure, workforce, governance – leaders most underestimate, Salamone doesn’t hesitate, “Data sovereignty is the biggest.” The cost isn’t theoretical. It’s the gap between believing a vendor’s “sovereign” label protects you, and discovering, the way that ICC judge did, that it doesn’t survive contact with a determined administration.
Organizations that have already built GDPR compliance have a head start they may not realize they have. “Organizations subject to the EU’s GDPR early on realized that compliance is most easily and credibly done on sovereign IT systems that locate data on hardware and data centers physically located within EU nations,” says IronSpark’s data and analytics analyst, Philip Russom. “Firms that have achieved that kind of compliance have valuable experience they can apply to sovereign AI.”
The instruction for a CDO assessing this Monday morning, in his view, is to “keep doing what the best ones already do: periodically review and revise their governance programs, under the assumption they must extend governance as new requirements arise.”
What Modularity Actually Costs an Engineering Organization
A new IronSpark Analysis report on sovereign AI, authored by Salamone, recommends that organizations “rearchitect for modularity.” That is easy to write and hard to execute. IronSpark’s DevOps analyst Haff grounds it in Conway’s Law: “Organizations which design systems are constrained to produce designs which are copies of the communication structures of these organizations. If you want a modular design, you pretty much need a modular organization to support it, one that’s almost certainly different from a globally integrated organization.” You cannot bolt federated infrastructure onto a centralized team and expect it to hold. The org chart has to change, not just the architecture diagram.
There is a practical middle path, though. “Certain things can be kept centralized in a sovereign environment, but there’s often a good argument to have standardized templates that are distributed locally, if only for reasons of keeping data and other assets local too,” he says.
But the bigger failure mode, in Haff’s view, isn’t technical at all. “CTOs have a tendency to focus on the tech and not the people and the culture,” he says. “To the degree that AI does fundamentally change workflows, hiring patterns, and how employees advance, that’s a major point I’m not sure anyone has a great handle on at the moment.”
Modularity also shows up at a smaller scale than org charts. Russom points to the common practice of embedding AI copilots directly into existing CRM, ERP, and sales tools, where “the AI tool should stay within the sovereignty established by the calling application or by enterprise governance policies,” rather than introducing a new, ungoverned dependency every time a team adopts a new AI feature.
The Pillar Nobody Talks About
Haff’s last point lands on the pillar that gets the least airtime: workforce and talent. An organization can acquire every piece of sovereign-compliant infrastructure on the market and still lack the human capital to operate it. Less than 5% of AI training data involving African languages is sourced or governed by African institutions, showing how quickly infrastructure can outpace the human ecosystem around it. The enterprise version is that adopting AI infrastructure without building internal competency, doesn’t make you sovereign. It just trades one external dependency for another.
Sovereign AI also isn’t a one-time migration; it’s an ongoing compliance obligation. IronSpark’s people and business empowerment analyst, Joe McKendrick, points to the EU’s AI Act, a framework mandating “ethical use, transparency, and localized compliance,” likely to keep evolving. “This requires constant human oversight to manage and assure that AI systems are in compliance,” he says.
The obvious corporate instinct is to bury that oversight in a compliance department. That’s exactly where it breaks down. “In Europe especially, trade unions have a strong presence and influence in corporate life. There is, and will be, considerable pushback against what are seen as job-killing or negative AI consequences,” McKendrick argues, pushback that works its way back to the legislators writing the next round of rules.
“The push for AI sovereignty needs to be woven into the fabric of the organization itself, meaning managers and workers at all levels will have a say.” Handing the whole problem to legal and compliance won’t solve the problem.
What to Do Now
Not every organization carries the same exposure, and the first useful question is whether the issue applies at scale at all. For multinationals, that exposure runs through customer data, supply chain relationships, and financial reporting, all potentially governed by different rules in different places, which Russom calls “a pressing issue.”
But of the roughly half a million manufacturing firms in the US, he notes, three-quarters are small businesses selling only to domestic partners. For them, “sovereignty as a governance parameter is not a pressing issue, and may not be an issue at all.”
The first real diagnostic question, in his framing, should be, “does our footprint even create this exposure in the first place.”
For organizations where the answer is yes, map your AI stack against the four sovereign AI pillars, Salamone recommends. Identify your highest-exposure workloads. It’s a different exercise than the data-sensitivity work most organizations have already done, since the relevant question isn’t who inside your organization can see the data, but which external authority could compel access to it, or shut off the model processing it.
The fastest way to start is to identify the single AI vendor whose sudden disappearance would stop a critical workflow tomorrow, and build a real answer for that one dependency before mapping the rest.
Asked about the timeline, Salamone is blunt: “Take the issue seriously immediately. There is no time for years-long evaluation. CTOs and CDOs must act now looking for alternatives.”
Update: June 28, 2026
This article was published on June 26th and the story continues to develop rapidly, including the following that took place on June 26th.
- The U.S. government partially reversed its June 12 export-control directive, allowing Anthropic to restore Mythos 5 to more than 100 approved U.S. organizations while broader restrictions remained. Anthropic said it continues working with the government to restore access to Fable 5.
- OpenAI announced that at the U.S. government’s request, it delayed the full public launch of GPT-5.6, initially limiting access to a small group of vetted partners while it worked with the Administration on a broader framework for future model releases. OpenAI described the arrangement as temporary and said it should not become the long-term default.
Join us for live discussion and Q&A with our analysts on Wednesday, July 1st at 1pm ET for their perspectives and guidance. Register here
Update: July 1, 2026
Claude Fable 5 and Mythos 5 redeployed
Anthropic announced on June 30th that the export controls on Fable 5 and Mythos 5 were lifted. They were redeployed on July 1st.
Related Resources
Research Report
What Every CTO and CDO Needs to Know About Sovereign AI
Foundational research on sovereign AI, including strategic drivers, governance models, market implications, and recommendations for enterprise technology leaders.
→ View the research report
CDO Checklist
Sovereign AI Checklist for CDOs
A practical checklist to help Chief Data Officers evaluate AI sovereignty risks across models, infrastructure, governance, vendors, and regulatory exposure.
→ View the checklist
Lisa Damast
Founder, IronSpark Analysis
Lisa Damast is the founder of IronSpark Analysis, where she works with a team of veteran industry analysts, decades of combined experience in data infrastructure, AI, and enterprise technology, to produce independent research for technical and business leaders. This piece draws on reporting and interviews with IronSpark’s full analyst team. Connect on LinkedIn →
AI disclaimer: AI assisted with the editing of this article