Anthropic Just Got Switched Off Overseas

So, What Actually Happened?

So, the week the headlines fought over which model is smartest, the real news was about who can switch the models off. We scanned 190,000 articles this week so you don't have to, and the pattern underneath was control, not capability. A US export ban quietly switched off Anthropic's models for users outside the country, proving a government can reach into your AI stack from another continent. A European regulator ruled Securitas' AI driver cameras unlawful, deciding that ”it keeps people safe” does not buy a pass on surveillance. Meanwhile the money voted with its feet: Baseten tripled to a $13 billion valuation in five months, not for building models but for running them. And a commodities desk warned the grid may not power the AI boom everyone is underwriting. Capability was last year's fight. This year's fight is control.

The Bottom Line: When everyone is arguing about how smart AI is, the smart money is asking a quieter question: who can turn it off, and is it you?

The Tracks That Matter

1. Baseten Triples to $13 Billion as Money Chases Inference

Here is a number that should reset a few assumptions. Baseten nearly tripled to a $13 billion valuation in five months on a raise of about $1.5 billion. The company does not build frontier models. It runs them, doing the unglamorous work of inference: turning a trained model into fast, reliable answers at scale. Five months ago it was a quarter of the size. The market just decided the place to put money is not the brains, it is the engine room where the brains actually do the work.

The $1.5 billion inference mega-round is the clearest signal yet that AI's center of gravity has shifted. For two years the headline money chased training: bigger models, more parameters, record compute bills. Baseten's raise says the next bottleneck is delivery. Every company bolting AI into a product now needs somewhere fast and dependable to serve it. Tripling in five months is not hype froth, it is demand for the layer that sits between a clever model and a working feature. When the picks-and-shovels company outgrows the gold miners, the gold rush has entered its industrial phase.

So the so-what for your roadmap: inference is becoming a dependency you cannot hand-wave. The model you fine-tuned is worthless if the thing serving it is slow, down, or owned by someone who can raise the price at renewal. Baseten's valuation is really a price tag on reliability, proof that ”where does this run, and who runs it” is now a board-level question rather than a DevOps footnote. The companies treating inference as strategic infrastructure will ship faster than the ones still treating it as plumbing they rent and forget.

Here's what works: Before your next AI feature ships, ask where it runs at scale and what happens if that provider doubles its price or goes dark for a day. If you cannot name a second option, you do not have an AI strategy. You have a single-supplier bet with a nice demo.

2. A US Export Ban Just Switched Off Anthropic Abroad

This is the story that belongs on every CIO's whiteboard. US export controls disabled Anthropic's models for users outside the country, and overnight, companies that built on Claude found a core dependency switched off by a policy decision made in Washington. Not a bug. Not an outage. A government reached across borders and turned off software a business was running, because the vendor's home country changed its mind about who is allowed to use it.

The geopolitics got loud fast. One analysis warned the ban sets a risky precedent: if the US can switch off one frontier vendor abroad, every government now knows AI is a lever it can pull. In Washington, the fight over the ”Fable” affair has Congress drafting mandatory testing of frontier models for national-security risk. Read together, the message to anyone outside the US is blunt. The most capable AI you can buy comes with a foreign off-switch attached, and the people holding it do not answer to your business.

So the strategic signal, and it is an uncomfortable one: model choice is now a sovereignty decision. For a European or Asian enterprise, ”best model” and ”model my government and my vendor's government will both let me keep using” are no longer the same question. That is why Europe's scramble to build its own AI heavyweights stopped reading as protectionism this week and started reading as risk management. Concentration on one foreign supplier was always a single point of failure. Now it has a name and a precedent.

Here's what works: Map every production system that depends on a single foreign-controlled model. For the critical ones, stand up a fallback, an open-weight model you host or a second vendor in a different jurisdiction, before the next policy shift makes the choice for you. Sovereignty is not a slogan. It is a backup plan.

Try It Yourself

3. GDPR Rules Securitas AI Driver Cameras Unlawful

When a regulator tells a security company its safety cameras are illegal, pay attention to the reasoning. A European data authority ruled Securitas' AI driver-monitoring unlawful, rejecting the company's argument that watching drivers with AI cameras was justified by safety. ”It keeps people safe” did not survive contact with GDPR. The cameras were not hacked or broken. They were switched off by a legal finding.

This is the workplace-surveillance shoe everyone knew would drop. AI monitoring has quietly spread into vans, warehouses, and call centers under the banner of safety and quality, and this ruling says the banner is not enough. GDPR treats constant AI observation of workers as a heavy intrusion that needs a far stronger justification than ”it might prevent an accident.” With the EU AI Act now sitting on top of GDPR for high-risk systems, the compliance bar for any AI that watches people just got a lot higher, and a lot more enforceable.

So the so-what: if your AI roadmap includes watching humans, whether employees, drivers, or customers, ”safety” and ”efficiency” are no longer a legal force field. The same capability that makes the monitoring useful is what makes it a liability. The companies that win here will treat consent, proportionality, and data minimization as design constraints from day one, not as a compliance review bolted on after the cameras are installed.

Here's what works: Inventory every AI system that observes a person, and for each one write down the legal basis in a single sentence. If that sentence is just ”safety” or ”quality,” you have a Securitas problem waiting for a complaint. Fix the justification, or turn the system off before a regulator does it for you.

4. The Grid Becomes AI's Quiet Off-Switch

Every AI valuation you have read this year quietly assumes one thing: the electricity will be there. A commodities analysis this week says it will not, and lays out an awkward gap between the power AI is being promised and the power the grid can actually deliver. The boom is being underwritten on an assumption nobody put on the slide, that you can plug in as much compute as the models demand, whenever they demand it.

The demand side is exploding from places most forecasts ignore. The Consumer Goods Forum just flagged that AI adoption will accelerate energy demand tied to data centers across ordinary industries: groceries, logistics, manufacturing, not just the frontier labs. So the load is rising everywhere at once while the build-out of generation and transmission moves at the speed of permits and concrete. That is why operators are signing behind-the-meter power deals and chasing reactors and remote hydro. They have done the math and do not trust the public grid to keep up.

So the strategic read: power is becoming the real constraint on AI, not chips or talent. The off-switch here is not a government or a regulator. It is physics and a substation. For most companies this lands as cost and latency, AI that gets more expensive and less available exactly when everyone wants more of it. The teams planning AI capacity the way they plan cloud spend, with power and location as first-class variables, will avoid the brownout the optimists are pricing out of existence.

Here's what works: When you model the cost of an AI initiative, model the energy too. Ask your providers where the compute physically runs and how that power is secured. ”We will scale when we need to” is a hope, not a plan, the day the grid says no.

5. Kalshi Eyes an IPO as Revenue Hits $2 Billion

Prediction markets just grew up. Kalshi, the platform where you can bet on real-world events, is eyeing an IPO as revenue hits $2 billion. That is not a fringe crypto curiosity anymore. That is a real financial business with real money flowing through it. A company that lets people put cash on what happens next is suddenly big enough to think about the public markets.

The timing tells you something. The IPO rumors landed in the same week Bitcoin slid below $64K and regulators sharpened their focus on crypto crime, yet the prediction-market story cut through the noise. Why? Because Kalshi sells something the rest of the space struggles to deliver: a clear, regulated product with obvious demand. Underneath the bets is a data business, a continuous, liquid stream of real-money signals about what a crowd actually believes will happen. That is a different and more honest thing than what people say in a survey.

So the so-what for anyone building with data: a $2 billion revenue line says markets-as-data has crossed from novelty to infrastructure. Pricing, forecasting, and risk teams now have a live, money-backed signal source that did not exist at this scale a few years ago. The contrarian read is that Kalshi's real product was never the gambling. It is the forecast, and forecasts you can trade are worth more than opinions you can only argue about.

Here's what works: If your team makes bets on the future, whether on demand, supply, or geopolitics, start treating prediction-market prices as one input alongside your internal models. They will not be right every time, but a number with money behind it beats a confident guess in a meeting.

6. Wayflyer Buys Conjura to Make Its Data AI-Ready

Here is a small acquisition with a big tell. Wayflyer, the small-business finance platform, acquired Conjura to accelerate its AI product offering for small businesses. Conjura is a data-analytics company most people have never heard of, which is exactly the point. Wayflyer did not buy a flashy AI startup. It bought the data plumbing that makes AI features actually work.

This is the quiet pattern under all the loud AI news. Companies are buying their way to clean, connected data because that is the part you cannot fake. Wayflyer lends to e-commerce businesses, a decision that lives or dies on the quality of the data behind it. Bolt a model onto messy data and you get confident, wrong answers fast. Acquiring Conjura is a bet that owning the analytics layer, the unglamorous pipes that turn raw transactions into something a model can trust, is cheaper and faster than building it from scratch.

So the strategic signal: in this cycle, the M&A that matters is not the headline mega-deal, it is the boring data-layer tuck-in. When a finance company spends real money to own its analytics rather than rent it, it is telling you where the durable advantage lives. The frontier model is rented by everyone. The clean, proprietary data underneath is the only part a competitor cannot also buy off the shelf next quarter.

Here's what works: Look at your last three AI initiatives and ask what stalled them. If the honest answer is data, scattered, dirty, or locked in someone else's system, your next investment is not another model. It is the pipes. Wayflyer just paid for that lesson. You can learn it cheaper.

7. Healthcare AI Left the Lab; the Hard Part Starts Now

Healthcare AI is having its reckoning. After years of dazzling demos, the field has left the lab and hit the hard part: getting a model to work safely inside a real hospital, with real patients, real liability, and real regulators. The breakthrough was never the bottleneck. The bottleneck is everything around it, the workflow, the trust, the accountability, and the messy clinical reality a benchmark never sees.

And the money is following the gap, not the hype. US health regulators are offering $2 million for next-generation healthcare technologies aimed squarely at the unglamorous integration problems: interoperability, data exchange, the parts that decide whether a model helps a doctor or just adds another screen. That is the tell. When public funding shifts from ”build a smarter model” to ”make the model usable in a clinic,” the field has moved from the demo stage to the deployment stage, where the boring problems are the whole game.

So the so-what, and it generalizes far past medicine: the lab-to-production gap is where most AI value gets stranded. Healthcare just hits it hardest because the stakes are highest. The pattern is the same in finance, legal, and operations. The model is the easy 20%, and the workflow, governance, and trust are the 80% that decides whether anyone actually uses it. The teams that budget for the 80% ship. The ones that fall in love with the demo stall.

Here's what works: When you greenlight an AI project, split the plan in two: the model, and everything around it. If 80% of your time and budget is not going to integration, change management, and accountability, you are funding a demo, not a deployment. The hard part is the part you are tempted to skip.

Signal vs. Noise

🟢 Signal: The compliance and security desks took the wheel. While the feeds argued about models, regulatory compliance and data security gained the most real pull across the week, the same days a regulator switched off Securitas' AI cameras and a government switched off Anthropic abroad. The decisions that actually killed or saved AI projects came from legal and security, not from a benchmark. Most coverage is still watching the labs.

🔴 Noise: ”AI,” ”Machine Learning,” and ”Agentic AI” as headline labels. All three pulled big volume again this week while their grip on the real conversation slipped. The terms are everywhere; the fundable, decision-shaping work moved one floor down, into compliance, risk, and the question of who controls the system. Anyone tracking ”agentic AI” as one signal is reading the brochure, not the budget line.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

A US export ban switched off Anthropic's models abroad, a GDPR ruling switched off Securitas' AI cameras, and a commodities desk warned the power grid could switch off the AI boom itself, all in one week.

Read alone, each story belongs to a separate desk: geopolitics covers the export ban, privacy law covers the camera ruling, the energy press covers the grid. Read on the same morning, they rhyme. For the first time, the binding question about an AI system is not ”does it work” but ”who can turn it off, and is it me?” A vendor's government, a regulator, or a substation can each end your deployment without asking permission. The move on Monday is to stop auditing your AI only for accuracy and start auditing it for control: which of your systems has an off-switch you do not hold.

By The Numbers

• Baseten tripled to a $13 billion valuation in five months — on a raise of about $1.5 billion, for running models rather than building them. The inference layer just repriced itself as critical infrastructure.

• Kalshi's revenue hit $2 billion as the prediction market eyes an IPO. Markets-as-data crossed from novelty to a real financial business.

• The deep learning market is projected to reach $1.96 trillion by 2035 — a forecast that quietly assumes a decade of uninterrupted compute, and the power to run it.

• 78% of enterprise leaders hit data-readiness problems that forced significant rework or slowed their agentic-AI rollouts. The model is rarely what breaks; the data underneath it is.

• 80% of CFOs expect AI-enabled business models to feature significantly in their organization within the next 12 months. The budget conversation has already moved from ”if” to ”how fast.”

• See what's rising in our 190K-article corpus this week →

Deep Dive: The Off-Switch

Every DJ has the same nightmare, and it is not a bad song. It is the moment the power cuts mid-set: the floor packed, the bassline about to drop, and then nothing. You did not make a mistake. Someone you have never met flipped a breaker you did not know existed. This week, enterprise AI met its breaker panel.

The money moved to where models run, not where they are made

Baseten tripling to $13 billion for inference says the industry has stopped fetishizing the model and started caring about the engine room, the place an AI feature actually gets served. That is healthy. It is also a concentration story. When everyone serves their intelligence through the same handful of layers, those layers become a dependency you share with your competitors and do not control. The brains are cheap to rent. The room they run in is the part that decides whether your product is up at 2 a.m.

Three hands reached for three switches

In one week, a US export ban switched off Anthropic abroad, a GDPR ruling switched off Securitas' cameras, and a commodities desk warned the grid could switch off the boom. None of those are model-quality problems. They are control problems: a government, a regulator, and physics, each able to end a deployment that works perfectly. The thing that breaks your AI in 2026 increasingly is not the AI. It is someone with a hand on a switch you forgot was wired to your business.

Dependency is the risk nobody benchmarks

We measure models to four decimal places and barely map who can unplug them. Concentration on a single foreign model, a single inference provider, a single power-constrained region: each is a single point of failure dressed up as a convenience. The export ban turned an abstract risk into a dated precedent. The companies treating ”who controls this” as a real question are quietly building fallbacks while everyone else admires the demo.

What Actually Works

1. Map the off-switches: For every production AI system, name who can disable it (vendor, government, regulator, or grid) and whether that someone is you.
2. Second-supplier the critical layers: One model, one inference host, one region is a bet, not a strategy. Have a tested fallback for anything you cannot afford to lose.
3. Make compliance a design input: Bake consent, proportionality, and data minimization into any AI that touches people, before a regulator names the date.
4. Plan power like you plan cloud: Treat energy and location as first-class variables in every AI capacity plan, not an afterthought.

The set only stops if you let someone else hold the breaker. Map your switches, wire your backups, and keep one hand on the power yourself. Then the music does not stop when someone you have never met decides it should.

What's Coming

The Sovereign-AI Fallback Becomes a Buying Requirement

European buyers are already reframing sovereign cloud as a strategic enabler, not compliance theater. After this week's export ban, expect ”who controls this model, and what is our fallback” to move from a footnote to a line item in every serious AI contract. The vendors that can answer it will close enterprise deals the ones that cannot will lose.

The Power Bill Lands on the AI Roadmap

Construction and energy desks are already wiring the AI build-out into the energy transition. The grid warning will not stay a commodities-desk story. Expect data-center power and physical location to show up as hard constraints in 2026-2027 capacity plans, and expect ”we will scale when we need to” to quietly disappear from the optimistic decks.

Workplace AI Meets the Compliance Wall

Privacy professionals are openly borrowing privacy frameworks for AI ethics. The Securitas ruling is the first of many. As the EU AI Act bites, any AI that watches employees or customers will need its legal basis written down before it ships. The companies treating governance as a feature, not a disclaimer, win the enterprise.

For Your Team

Strategic purpose: This week belongs on the leadership table because it changes the question you ask about AI. The headlines argued about which model is smartest. The real story was that smart is no longer the risk. Control is. Your edge this quarter is naming every off-switch in your AI stack out loud, and wiring a backup before a vendor, a regulator, a government, or the grid forces the issue for you.

Monday's meeting prompt: ”For every AI system we now depend on in production, name who can switch it off without our permission (a vendor, a government, a regulator, or the power grid) and tell me what our fallback is if they do.”

The Off-Switch Audit Framework:

1. Name the switches — For each production AI system, list everyone who can disable it. If you cannot draw that list in 30 minutes, that is your first finding.
2. Second-supplier the critical few — Anything the business cannot run without needs a tested fallback model, host, or region. One supplier is a bet, not a plan.
3. Write down the legal basis — Any AI that watches a person gets one sentence of justification. If that sentence is just ”safety,” you have a Securitas problem waiting.
4. Budget the power — Treat energy and location as real constraints in every AI capacity plan, not someone else's problem.

Share-worthy stat: Baseten tripled to a $13 billion valuation in five months, not for building AI models but for running them. The market just priced reliability higher than genius.

Go deeper: Track who controls the AI you depend on →

The Track of the Day

”Every AI boom forecast being published right now seems to be making the same assumption. The electricity will be there to power it when they need it. It won't.”
— from this week's commodities analysis on AI and the power grid

The models keep getting smarter. The question this week was quieter and harder: when it matters, who is holding the off-switch, and is it you?

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: June 20, 2026 | Curated by Yves Mulkers @ Ins7ghts