Was this forwarded to you by a friend? Sign up, and get your own copy of the news that matters sent to your inbox every week. Sign up for the On EdTech newsletter. Interested in additional analysis? Upgrade to the On EdTech+ newsletter.

A few weeks ago, in an On EdTech+ post titled The Resistance Is Real, I argued that the AI resistance on display at ASU+GSV wasn't a passing annoyance. It's likely to be one of the major EdTech stories of the next year or two. Some of that resistance rests on weak arguments and outdated assumptions. But not all of it — there are legitimate concerns underneath, particularly about AI's impact on deeper learning and on students' future job prospects.

That resistance has mostly been B2C: students and faculty reacting as end users. Two weeks into the Anthropic Fable 5 saga, I think the B2B version is emerging, even if it won’t look on the surface like the B2C resistance. It will show up as procurement friction, institutional hesitation, regulatory pressure, and pushback against American-dominated AI infrastructure.

Get ready for AI sovereignty to become an EdTech issue.

The short version of the Fable 5 story is that for several months Anthropic touted its new Mythos class of AI models as having a new level of risk for cybersecurity. Two weeks ago, Anthropic released Fable 5, the first Mythos product release, with safeguards to supposedly remove the cybersecurity risks along with biological weapons development and distillation attempts of other AI models.

Based on my two-day experience with Fable 5, I agreed with most analysts who called it a meaningful leap in capability, significantly beyond OpenAI's GPT 5.5 and Anthropic's own Opus 4.8. Within days, the U.S. government imposed export-control restrictions on Anthropic's most advanced models, citing national security concerns partially triggered by Amazon’s testing. Anthropic responded by pulling access to Fable 5 and Mythos 5 for all customers, arguing that the directive made it impossible to comply cleanly while continuing to operate the models.

There are real arguments to have on the merits. Maybe Anthropic was right to be cautious with the injected safeguards. Maybe the Trump Administration was right to intervene. Maybe the government overreached based on an incomplete grasp of how probabilistic AI systems actually work. Those debates matter — but for institutions and EdTech vendors, the takeaway is simpler: access to a major AI capability that customers had already started using was pulled back.

Not because a customer misused it. Not because a university changed its mind. Not because a vendor failed to renew a contract. The model was effectively taken away because Anthropic and the US government each showed they had both the power and the willingness to decide what AI capabilities could remain available.

That will likely get the attention of organizations building strategy around AI, even if the immediate issue gets resolved, with Fable 5 re-released, in the near term.

Anthropic has already shown that it will change model behavior, adjust prompts behind the scenes, downgrade model access, and make unilateral product decisions in the name of safety. And that Anthropic leadership believes it is solely positioned to determine. The US government has now shown that it may treat advanced AI models as export-controlled strategic technologies, with access subject to national-security judgment. In its sole discretion. Agree or disagree with either side — the combined lesson makes it viscerally clear that customers do not fully control the AI capabilities they are beginning to depend on.

In a technology context, sovereignty is about control. Who controls the infrastructure? Where does the data go? Which legal regime applies? Who can inspect the system, change it, or turn it off?

For AI, that set of questions is broader than it was for cloud computing, where the focus was mostly about data residency. AI sovereignty is about control over the models, the data path, the hosting environment, the allowable uses, the continuity of access, and the ability of institutions to rely on those capabilities over time.

Two angles matter for EdTech.

The first is national sovereignty. Non-US governments, universities, and systems of higher education will increasingly ask how much of their AI strategy can safely depend on American companies and American policy. What happens during the next trade dispute? What happens if the US decides a capability used in advising, assessment, research, or student support is too sensitive to export? What happens if a model is available one semester but restricted the next?

This is where the Fable 5 saga becomes bigger than Anthropic. It gives non-US institutions a concrete reason to ask whether American AI can be trusted as infrastructure. That doesn't mean they'll all abandon the leading AI tools — in many cases they won't have realistic alternatives. But it could mean that procurement rules, national AI strategies, and institutional risk reviews may start looking for different answers: local hosting, country-of-origin requirements, model transparency, open-weight options, or preferred support for national and regional providers — France's Mistral, the UAE's G42, and the national AI pushes now underway from India to the Gulf.

The second angle is organizational sovereignty, and this one applies inside the US as well.

What happens when a college builds operational processes around AI — not just classroom experiments, but advising workflows, tutoring, accessibility support, research assistance, student services, compliance, and administrative automation? What happens when an LMS, assessment platform, or student-success system embeds AI based on one model provider, and the vendor's pricing, roadmap, and revenue projections all assume that access will continue?

Here, sovereignty is not a geopolitical issue. It's an operational-risk question: is there trust that AI capabilities will continue to be available, who controls these decisions, and what happens if policies changes? Put another way — is the institution buying a capability, or taking on a dependency?

EdTech went through a somewhat related version of this with cloud computing.

In the 2000s and early 2010s, many non-US institutions were wary of cloud-based EdTech, fearing that student data could pass through US data centers or be exposed to government surveillance. Privacy rules, data-domicile requirements, and procurement expectations slowed cloud adoption, particularly in Europe. Vendors could insist that cloud hosting was secure, scalable, and better than local infrastructure — but that didn't answer the additional questions of where does the data live, and which government can reach it?

One inflection point came when AWS opened its Frankfurt region in 2014. The AWS Ireland data center had been open for a few years, but there was no acceptance of that providing a viable and long-lasting solution to the cloud data residency issue for EU countries. AWS Frankfurt had a much bigger impact. It gave vendors and institutions a practical and believable answer they'd lacked: European data could be hosted in Europe under a more acceptable and viable operating model. Cloud EdTech didn't become frictionless, but the infrastructure changed the nature of European higher ed procurement decisions.

Seen in this light, AWS Frankfurt wasn't only a relief for nervous buyers. It was a win for the vendors who built it and who adopted it. Amazon turned a procurement objection into a competitive advantage and took business that hesitant or US-only competitors couldn't. Similar changes came from EdTech vendors ready to adopt its usage. The lesson for AI isn't merely defensive. Whichever vendors give institutions a credible answer to "who controls this capability, and what happens if it's withdrawn" will gain a competitive advantage.

Assuming that AI sovereignty becomes a much bigger EdTech issue, AI will need its own AWS Frankfurt moment. The analogy isn't perfect — cloud sovereignty was mostly about data location and legal access, while AI sovereignty is about something harder: control of capability.

There's also a risk that EdTech answers a legitimate question with a rearview-mirror understanding of AI itself. Much of the sector is still operating from an AI conversation two years out of date — chatbots, cheating, syllabus statements, generic privacy worries — and procurement rules built on those assumptions could slow adoption without solving the sovereignty problem. Cloud went through a similar pattern: real issue, understandable fear, some blunt early regulation that delayed adoption, and eventually new infrastructure compromises that largely unlocked the market. AI sovereignty looks likely to follow the same arc.

The open question is what those compromises will be. Local model hosting? Open-weight models? Regional AI clouds? Vendor-neutral AI gateways? Contractual guarantees of model continuity? Viable back-up options? Government-backed national models? Some combination of all of them? Whatever it is, I doubt the current "just plug in the best available model and move fast" phase stays acceptable in every market.

This matters for vendors in ways that are easy to underestimate.

AI is no longer just an end-user tool sitting outside institutional systems. It's starting to be built into products. The LMS isn't just adding a chatbot; advising tools aren't just adding better search; assessment products aren't just adding writing detection. Vendors are rebuilding around AI-assisted workflows — and institutions may find themselves depending on capabilities they never separately procured or fully understood.

So adding AI features is no longer just a product-design decision. In some markets it will be seen as a procurement risk. Institutions may start asking which model provider is being used, where prompts and outputs are processed, whether data is retained for training, whether the feature can be disabled, whether a local or open-source model can be substituted, and what happens if the underlying model is withdrawn or restricted.

Vendors will want to market AI as a capability layer. But I suspect that institutions will also evaluate it as a dependency layer. That complexity could slow down adotpion, but it could also provide competitive advantage for whoever answers it first.

I don't think AI sovereignty will dominate EdTech procurement in 2026. The immediate story in higher education will still be faculty resistance, student use, academic integrity, workflow redesign, and vendor messaging. It will be fascinating to watch the July LMS conferences try to thread it: promoting AI capabilities while navigating the B2C resistance minefield.

But I'd add the B2B question to the radar.

The main On EdTech newsletter is free to share in part or in whole. All we ask is attribution.

Thanks for being a subscriber.