Sweden Chose Its Favorite American

Sweden’s ministries grounded two American chatbots for coming home without a contract. Both are still allowed over through a side door, and a third American got the house keys. If your sovereignty policy is a list of approved vendors, you didn’t choose independence. You chose a favorite.

Sovereign AI | theCUBE Research | Amit Eyal Govrin | October 2026

The week Sweden picked a favorite

On Friday, October 23, Sveriges Radio’s Ekot reported that Sweden’s Government Offices, the Regeringskansliet that houses the prime minister’s office and every ministry, would block ChatGPT and Claude on staff work computers [1][2]. On Monday, October 26, the block went live [2][3].

The reason fit in one sentence. The digitalization director, Mikael Wettercrantz, warned staff that using AI services the Government Offices has no agreement with can have information-security consequences (my translation) [3]. He declined an interview [2].

The favorite fit in one sentence too. “All employees have access to Microsoft’s AI tool,” the Government Offices told SVT by email (my translation), after an internal review that tightened Copilot’s security [1][3].

Then the same email added the line that turns a security memo into a family sitcom: “Within Intric, employees still have access to the language models ChatGPT and Claude” (my translation) [1].

Pinch yourself for a second and read that again..

Sweden did not ban ChatGPT and Claude. It banned the browser tab.

My work is done here- that’s the whole edition. Two kids got grounded for coming home without a permission slip. They’re still allowed over through the side door. And the favorite, the one from Microsoft, got the house keys, the car, and the allowance.

Every family has some history. In August 2025, Prime Minister Ulf Kristersson told Dagens industri that he uses ChatGPT and France’s Le Chat to get a second opinion on political questions [4]. Fourteen months later, his own office grounded one of them. Neither announcement mentions Le Chat. The European kid wasn’t grounded. It just wasn’t anybody’s favorite.

Credit where it’s earned

Drumroll – first for the credit section, because Stockholm got three things right, and pretending otherwise cheapens the rest of the argument.

First, shadow AI is a real problem, and this is a real answer. A civil servant pasting a ministry draft into a consumer chatbot with no contract is the Operational pillar failing at the keyboard. Sweden is not alone. The Swedish Tax Agency and the National Board of Institutional Care already block AI services, and the Enforcement Authority is considering it [1]. In February, the European Parliament disabled built-in AI features on lawmakers’ work devices, after its IT department said in an internal email that it could not guarantee the security of data uploaded to AI companies’ servers [5].

Second, the Government Offices bought a platform, not a chatbot. In July, it awarded Intric, a Stockholm company, a contract for a platform where staff build assistants on the ministry’s own data, delivered as cloud and on-premises [6]. Intric says the platform runs on Swedish infrastructure without depending on foreign cloud services, offers open models that can run entirely on Swedish soil, and can run fully air-gapped [6][7]. Its CEO calls the approach “sovereignty through architecture” [7]. That’s the right sentence.

Third, the contract is short. Twelve months, with an option for twelve more [7]. In this newsletter, a short contract is the Financial pillar doing push-ups.

Then you look at who got the keys.

How the favorite got picked

The test Stockholm actually ran was simple: do we have an agreement with you? ChatGPT and Claude in a browser: no. Copilot: yes. Intric: yes.

That is not a sovereignty test. That is an accounts-payable test.

A contract tells you what the vendor promised. It does not tell you which court can overrule the promise. Contracts lose to statutes. In both directions.

Microsoft, to its credit, has put more on paper than most. In April 2025, Brad Smith pledged that Microsoft would contest any order to suspend its European cloud operations in court, and would challenge government demands for European customer data wherever it has a legal basis to do so [8]. On October 20, Microsoft added Data Guardian, a European partner that reviews its engineers’ remote access to EU customer systems, plus a tamper-evident access log [9]. Real engineering. I said so in the custody edition, and I’ll say it again.

But on June 10, 2025, under oath before a French Senate inquiry, Microsoft France’s director of public and legal affairs was asked whether he could guarantee that French citizens’ data entrusted to Microsoft would never be handed to US authorities without the explicit agreement of French authorities. His answer: “Non, je ne peux pas le garantir.” No, I cannot guarantee it. He added that it had never happened [10]. Fair. A pledge to fight is still not a promise to win.

Last week I wrote that Europe got visitation rights and Microsoft kept custody. This week Sweden went one step further. It made Microsoft the favorite and called it house rules.

Here’s the part nobody said out loud

The ban does not remove a single model. It removes a door.

Door one, the browser: locked.

Door two, Intric: open, with ChatGPT and Claude on the menu next to Gemini and the open models [1][7]. Intric’s own storage is Swedish. But every frontier GPT or Claude model on the market is served by an American-headquartered company. Whatever country the GPU sits in, the parent company that answers the subpoena has a US address.

Door three, Copilot: open to every employee. And the favorite has a guest room. Microsoft now offers Anthropic’s Claude models inside Copilot. In the EU they’re off by default, an administrator can switch them on, and Microsoft’s own documentation says those models “are currently excluded from the EU Data Boundary” [11]. The Government Offices has not said whether anyone flipped that switch [12].

Sit with that for a second. The grounded kid may already be sleeping over in the favorite’s room.

So here is the policy Sweden actually shipped. ChatGPT: grounded, unless you go through Intric. Claude: grounded, unless you go through Intric, or through Copilot if an admin says yes. Copilot: approved for everyone.

That’s not a ban. That’s a seating chart. Good week for Microsoft.

The five pillars, under pressure

I’m grading the dependency, not the company. The buyer is the Swedish Government Offices. Same test as always. What holds when someone pushes?

01 Territorial

The question. Where do the prompts and the answers physically live, at rest and in motion?

The story. Copilot runs inside Microsoft’s EU Data Boundary [11]. Intric stores data in Sweden and offers open models that never have to leave Swedish soil [7]. The exceptions are the guests. Claude inside Copilot sits outside the EU Data Boundary once it’s switched on [11], and the frontier models inside Intric run wherever their provider runs them.

Under pressure. A minister’s aide picks “Claude” from a dropdown in either house. Where did the draft go?

Verdict. Partial. The house is Swedish. Some of the sleepover guests aren’t.

02 Operational

The question. Who runs it, who holds the keys, who gets paged, and who can pull the plug?

The story. This is the pillar Sweden actually moved. Tools without a contract are blocked, the favorite was reviewed and hardened, and the government now decides which doors open [1][3].

Under pressure. Microsoft changes a default, the way Anthropic models went on by default for most commercial tenants outside Europe [11]. Who in Stockholm finds out first?

Verdict. Passes on house rules. The favorite still writes its own curfew.

03 Technological

The question. Can you audit it, fork it, self-host it?

The story. Copilot: no, no and no. ChatGPT and Claude, through any door: no, no and no. The one row that passes is the open models on Intric, on Swedish servers or fully air-gapped [7].

Under pressure. The model your Copilot assistants were tuned on are no longer your roadmap discretion.

Verdict. Fail for the favorite. Pass for the quiet kid nobody made the default.

04 Legal

The question. Which court can compel whoever holds your prompts, your keys or your model?

The story. Microsoft, OpenAI and Anthropic are all American companies. My standing rule applies: a hosted frontier model anywhere in the request path is an automatic fail on this pillar. Microsoft has pledged to fight [8]. Its French legal director told the Senate, under oath, that he could not guarantee the outcome [10].

Under pressure. Three words: US Cloud Act.  A US order for one ministry’s Copilot history lands at 5:21pm Eastern.

Verdict. Fail. Grounding two Americans doesn’t help. The favorite shares their courts.

05 Financial

The question. How much control do you have? Are you able to exit on your timeline, at a survivable cost, without a meter you don’t own setting your budget three years out?

The story. Before Monday, a civil servant could open any of three American chatbots in a browser. After Monday, the only one open to every employee is Microsoft’s [1]. A Swedish investor blog put it more bluntly than I would: the rule rewards whoever already holds the contract [12]. Every assistant built in Copilot and every workflow shaped around Copilot habits is switching cost, accruing on someone else’s meter. The 12-month Intric contract stays on the board as credit [7]. That’s an exit with a date on it.

Under pressure. Renewal season. Microsoft knows it’s the favorite.

Verdict. Fail on the default. Partial for the short contract. The favorite always negotiates the allowance.

If the budget doesn’t survive the renewal, you never had the fifth pillar, whatever the other four score.

The executive TCO

Three line items, none of which show up on the invoice:

• The favorite’s allowance. Whoever is the default gets the usage, and whoever gets the usage gets the renewal. Model your unit economic alpha at 3x token prices, and price the renewal the same way.
• Two allowances for the same kids. Frontier models paid for twice, once inside Copilot and once inside Intric. Two doors, two bills, the same American models.
• Moving out. Assistants, agents and prompt libraries built inside Copilot don’t pack themselves. Every month on the default is another month of switching cost.

Same playbook: steady-state-and-burst strategy as always, applied to a government. Baseline on open models on Swedish soil, which the Government Offices already bought. Burst to the frontier through a gateway you own, with every call logged. Own the crown jewels. Rent the edges.

Sweden bought the crown jewels in July. On Monday it made the edge the favorite.

What to do Monday

• Open the Microsoft 365 admin center and check “AI providers operating as Microsoft subprocessors.” Know whether Anthropic is switched on, who switched it on, and for which users [11].
• Ask every AI platform you buy which company serves each model, from which country, under which law. Put the answer in the contract annex, not the sales deck.
• Run the 72-hour “prohibited, not down” drill on your default assistant. If the honest answer is “we’d go back to email,” write it into the risk register.
• Make the sovereign row the default, not the exception. If the open model on your own soil is option four in a dropdown, almost nobody will pick it.
• Keep your contracts short. Sweden got this part right.
• Price the favorite. Write down what leaving would cost in year three, then fund a sovereign sibling the way Bern funded its 3,000-seat pilot.
• Two questions for your favorite vendor. “If a US court asks for our prompts, who tells us?” Then: “What does leaving cost in year three?”

Calculated risk, priced in

Sweden is not the first family in Europe to pick a favorite. In July I wrote that Switzerland built a sovereign AI, then wired the money to Microsoft [13]. The pattern hasn’t changed since.

Just before Christmas 2024, the Federal Office for Buildings and Logistics signed three more years of Microsoft office licenses, reported at CHF 150 million, without a public tender [14]. On September 2, 2025, EPFL, ETH Zurich and the Swiss National Supercomputing Centre released Apertus, an open model with its weights and training data published [15]. Credit in full. Then Bern kept paying the favorite.

A month later, the outgoing army chief, Thomas Süssli, wrote to the Federal Council’s delegate for digital transformation that Microsoft 365, in its current configuration, was largely unusable for the Defense Group. He asked for an exit strategy toward a redundant platform independent of Microsoft (my translation) [16].

In September, the Federal Chancellery answered with a price tag. A first phase of CHF 9 million puts about 3,000 of roughly 47,000 federal employees on an open-source workplace by the end of 2027, running alongside Microsoft 365, not replacing it [17].

Read that the way a CFO would.

That’s not a purity test. That’s a hedge. Switzerland kept its favorite and bought a second set of house keys.

And that’s the honest pattern across Europe this year. Nobody in Bern or Stockholm is running a sovereignty purity test. They’re making calculated bets: keep the favorite for the work that needs it, build or buy a sovereign sibling for the work that can’t afford the exposure, and decide how much risk the family can live with. Sovereignty is not binary. It’s control with an acceptable risk sidecar.

I’m fine with that. Picking a favorite is allowed. The test is whether the risk is priced in.

Switzerland put a number on its hedge: CHF 9 million, 3,000 seats, a date. Sweden put a number on its contract: twelve months with Intric. Neither has published what it would cost to leave the favorite. That’s the line item that turns a calculated risk into an uncalculated one.

Calculated risk is a strategy. Uncalculated risk with a favorite’s name on it is the ostrich.

Every family has a favorite

Sweden grounded two American chatbots for staying out without a permission slip. Fair enough. House rules matter, and most ministries in Europe still don’t have any.

But a permission slip is not a passport. ChatGPT and Claude are still in the building through the side door. The favorite got a key to every room. And the one genuinely Swedish row on the scorecard, the open models on Swedish servers, is the kid you have to go looking for.

Hybrid is fine. Mislabeling isn’t.

Picking a favorite is fine too. Not pricing it isn’t.

Sweden didn’t choose sovereignty this week. It chose its favorite American. Until the favorite answers to a Swedish court, Sweden isn’t sovereign. It’s a parent with a favorite.

Amit

Amit Eyal Govrin is lead analyst for sovereign AI at theCUBE Research and CEO of Agentcy Labs, which designs and builds sovereign AI platforms for enterprises and governments. His research for theCUBE Research is conducted under theCUBE Research’s editorial standards and is not sponsored or directed by Agentcy Labs. Agentcy Labs may hold commercial relationships with vendors covered in this research.

References

[1] SVT Nyheter, “AI-tjänster spärras på Regeringskansliet,” Oct 23, 2026 (updated Oct 24; cites Ekot’s original report). Link

[2] The Next Web, Alina Maria Stan, “Sweden’s Government Offices block ChatGPT and Claude on work computers,” Oct 26, 2026. Link

[3] Techtidningen, “Regeringskansliet blockerar AI-tjänster,” Oct 26, 2026. Link

[4] PC Gamer, “‘We didn’t vote for ChatGPT’: Prime Minister of Sweden admits he uses AI chatbots for second opinions,” Aug 2025. Link

[5] TechCrunch, Zack Whittaker, “European Parliament blocks AI on lawmakers’ devices, citing security risks,” Feb 17, 2026. Link

[6] Intric, “Intric Wins Procurement with the Swedish Government Offices for Sovereign AI Platform,” Jul 8, 2026. Link

[7] Techtidningen, Linus Bengtsson, “Regeringskansliet väljer svensk AI-leverantör: ‘Något vi brinner för’,” Aug 11, 2026. Link

[8] Euronews, “Microsoft pledges to take legal action if governments try to seize EU data,” Apr 30, 2025. Link

[9] The Next Web, “Microsoft adds EU access oversight and offline AI to its sovereign cloud,” Oct 2026. Link

[10] Sénat, Commission d’enquête sur la commande publique, compte rendu, audition de MM. Anton Carniaux et Pierre Lagarde (Microsoft France), Jun 10, 2025. Link

[11] Microsoft Learn, “Anthropic models in Microsoft Online Services,” updated Oct 23, 2026. Link

[12] Marketmate, “Microsoft – vad Regeringskansliets förbud faktiskt betyder,” Oct 25, 2026. Link

[13] theCUBE Research, Amit Eyal Govrin, “Switzerland Built a Sovereign AI. Then Wired CHF 140M to Microsoft.,” Jul 20, 2026. Link

[14] Blick, “Bund vergibt Microsoft Lizenzen für 150 Millionen freihändig,” Jan 20, 2025. Link

[15] EPFL AI Center, “Apertus: a fully open, transparent, multilingual language model,” Sep 2, 2025. Link

[16] Netzwoche, “Armeechef wünscht sich Microsoft-Alternative für Gruppe Verteidigung,” Oct 31, 2025 (reporting Republik). Link

[17] The Register, Liam Proven, “Switzerland tests a FOSS escape route from Microsoft 365,” Sep 9, 2026. Link