Dependence is not a fault, but treating it as an asset is one — the moment that asset does not belong to you. On June 12, on Washington's order, Anthropic had to cut off its two most powerful models — Fable 5 and Mythos 5. The instruction targeted only non-American users; unable to isolate them, the vendor cut off everyone, everywhere, with no notice. "The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance," the vendor writes in that day's statement, followed by "We apologize for this disruption to our customers." A vendor apologizing for cutting off its own product, because it was ordered to. The model lived only three days — too little for any Maison to have built anything on it. But the day a cutoff hits a model in service, it is not the storefront that will go dark: it is the operational core — the agent that forecasts demand and arbitrates stock allocation across points of sale, the one that adjusts prices market by market, the one that drives replenishment and logistics. None of this speaks to the client — and that is precisely where, in the invisible that keeps the Maison running, the cutoff would hurt the most.
This is not an outage, it is a precedent. When the availability of what runs your business no longer rests on a commercial relationship you reread and renegotiate, but on a political decision you have no hold over, trust no longer protects anything. Architecture does.
A stack of agents is not a back-office tool: it is a brand asset, on a par with distribution or savoir-faire. I call this the patrimoine agentique (the agentic estate) — the intelligence a Maison owns, documents and passes on, where others merely rent it. The trap is insidious: the lock-in stays invisible as long as you experiment, and it turns deadly once you go to scale — exactly where Maisons stand today. A Maison always keeps a spare set of keys to its workshops; for the intelligence that runs its business, that spare set does not yet exist.
The figures tell the illusion. According to the Harris Poll survey conducted in late 2025 for Dataiku among 600 chief information officers, 74% of them regret at least one major AI-vendor choice from the past eighteen months. You think you are choosing freely; a year and a half later, you are defending before your board a choice you can no longer undo. And the price of exit is never merely financial: it is quality that drops, service that breaks off, a process you can no longer reproduce identically, maintenance that drags on. For a Maison, the lesson is bare: the right to exit is not the ease of exit. Psychologists have a name for this gap between the exit you believe possible and the one you endure: the illusion of control — you always overestimate your grip on what you have handed to someone else.
Last Thursday, in "The War on Fakes," I told you about the launch of Fable 5. By the time you read me, Washington had had it cut off. The risk, at bottom, is not American: it is single-vendor — it is entrusting what runs a Maison, its operations, its e-commerce and its client knowledge, to a single hand you do not hold. The answer is not a spare flag, it is the intelligent diversification of your agentic stack: the one that preserves, precisely, the right to roll back.
A question. If your model vendor cut you off tomorrow morning, how many hours before your agents are running again — the ones that anticipate your sales, drive your logistics, run your e-commerce? And do you know, or do you merely hope?
Why "The spare keys"? Because a Maison never entrusts the only copy of a key. Of its workshops, its vaults, its reserves, it always keeps a spare — to get in the day the first lock jams. Its agentic stack should follow the same rule: a second vendor already wired in, data ready to redirect, enough to reopen the door without waiting on a third party's goodwill. The spare keys are not distrust. They are the condition for staying master in your own house.
Welcome to LUXE ÆTERNAI: my weekly decoding of what AI agents change, or do not change, for Luxury Houses. I am Mickaël Tsakiris. Twenty years in Luxury, on the House side and the digital-agency side, from Saint Laurent to LVMH by way of Dior, Chanel and Hennessy. I help Houses and leaders turn agentic AI into a competitive advantage without betraying their DNA. Enjoy the read!
Sources:
The week in brief.
- On June 12, on Washington's order, Anthropic cut off Fable 5 and Mythos 5 for all its customers, everywhere — the first frontier model already in service unplugged by a state. Not an outage: a precedent (the full story below).
- It is not "a tool" that is under threat, it is a Maison's agentic estate — the invisible intelligence that runs its operations, its logistics, its pricing and its e-commerce, and that it believed it owned when it was only renting it.
- The trap is not American, it is single-vendor. Replacing Anthropic with Mistral because it is French is not becoming independent: it is changing landlords. Sovereignty rests on an architecture with several exit doors, not on a flag.
- Dependence on the model takes three forms: the price that doubles without warning, the bundled-in that becomes paid (Anthropic, June 15), the model a state or an outage cuts off. Each has its tooled-up counter — routing across several models, vendors wired in as fallback, a written continuity plan.
- Multi-model is no longer the avant-garde, it is the infrastructure: 81% of CIOs plan at least two vendors in 2026 (Harris Poll/Dataiku), and gateways like LiteLLM or OpenRouter make switching possible without rewriting the agent.
- Residency is not sovereignty: an American cloud hosted in Europe remains subject to the CLOUD Act. The only test — who can be compelled by a judge to cut off access?
- This week, Luxury parades its agents at VivaTech (June 17-20), while the real question stands: who drives the agent, the Maison or the platform that sold it to them?
- Next Thursday (#18): the heaviest bill, the human one — rebooting a model takes months, rebuilding a craft you let atrophy takes years.
It provisions for everything. Except what runs all the rest.
A Maison provisions for everything it sees coming. The price of gold soaring and eating into the jewelry margin, the exchange rate that swallows a collection sold in dollars, the sole supplier of a rare material whose access you secure by taking a stake in its capital, the counterfeiting you hold at bay with legal task forces, the stock you insure down to the last square of silk: for every visible risk, a plan, a hedge, a second source. This has been its discipline for two centuries, and it practices it better than any other industry.
The one risk it has not provisioned for is the one that runs all the others. Agentic intelligence is on its way to becoming the Maison's invisible conductor — it forecasts demand, allocates stock, orchestrates logistics, runs e-commerce — and almost no one has written the plan for the day it goes dark. We insure the material; we forgot to insure what drives it. June 12 sent the bill, or rather "a warning with no charge," as the saying goes.
Three signals this week.
Three strong signals this week.
Shoptalk Europe brought Luxury up onto the agentic stage. In Barcelona, June 9-11, under a theme that says it all ("Where AI and human ingenuity meet"), the Houses sent their operators, not their futurists. Mark Elkins, who runs L'Oréal's global e-commerce, shared the stage with Google for a session whose title is disarmingly frank: "Agentic Commerce: What's Here, Real, and Next?" Nadine Graf, EMEA president of Estée Lauder, came to defend the results of her "Beauty Reimagined" transformation. Alexis Rollier, Global COO of Sephora, made the case for unified commerce — stores as experience hubs, data aligned behind them. The signal that matters is not in the announcements, it is in the makeup of the table: when L'Oréal talks agents, Google is sitting next to it. A Maison's real question is not "do I have an agent?", it is "who drives it — me, or the platform that sold it to me?" The whole subject of the week is in that neighboring chair.
VivaTech came down to the Champs-Élysées, and Guerlain turned it into an AI showcase. On Sunday, June 14, for its tenth anniversary, the show left the Porte de Versailles for an open-air, public edition on the most beautiful avenue in the world. Guerlain, long established at 68 Champs-Élysées, deployed there an LVMH × VivaTech activation dedicated to AI and content creation — the House of the beehive amid the robots and the prompt contests. Anecdotal in substance, revealing in form: Luxury no longer watches AI from afar, it puts it in the window, at curb height. Two centuries of controlled narrative, on display one Sunday between two demos — the Maison that controls its desire must now decide what it entrusts to the machine, and before whom.
And at the show, June 17-20, Luxury rolls out its agents for real. LVMH deploys "Dreamscape" there — twelve projects, eleven Maisons and their technology partners. CELINE presents CELIA, an agent powered by MaIA, the group's in-house platform, which assists the sales teams and in-store client service. Louis Vuitton unveils an app wired to ChatGPT. Another Maison shows an agent that generates its marketing content while holding the House's tone. Luxury no longer watches agentic AI from afar: it stages it. But when Vuitton's app runs on ChatGPT and LVMH debates agents with Google on the main stage, this edition's question comes back, sharper: who drives the agent — the Maison, or the platform that supplies it?
In support, a figure that frames the stakes without summing them up: according to the AI Luxury 25 ranking (5W / Haute Living) published on June 9, more than a third of luxury buyers now start their search with an AI rather than a classic search engine. The Maison's front door is shifting — and the one who holds the door is not the Maison. That is precisely where this week's subject begins.
Sources:
Outside Luxury, what weighs on Maisons anyway.
Three signals that appear nowhere else in this edition — and that all speak to the Maison, by ricochet.
1. "All-agentic" is a staging, not a state of the world
Menlo Ventures put a number, in its annual report "The State of Generative AI in the Enterprise" (2025 edition), on what the demos know how to hide: across all enterprise AI deployments, only 16% are true agents — able to plan and act in a loop. Everything else is fixed-sequence automation, workflow dressed up as intelligence. For a Maison, the consequence is direct: before dreading — or promising — a client journey "driven by an agent," you have to check that the thing still deserves the name, because it is exactly this confusion that pushes you to hand over too quickly an adviser, a storefront, a client file to a system that orchestrates nothing. You demand the proof, not the label.
2. The AI Act is not pushed back — the misreading, on the other hand, thrives
It was read everywhere that Europe was "easing up" on AI. False. Only the obligations on high-risk systems slip to December 2027 (Digital Omnibus, May 7 agreement); the rules on general-purpose models and the Commission's enforcement powers do fall, as planned, on August 2, 2026 — in six weeks. For a Maison industrializing a clienteling, moderation or client-file-sorting agent, the deadline that exposes it has not moved a single day. The reprieve it was sold does not exist.
3. A market for independence is taking shape — and it is already billing
Deliverance AI came out of stealth on June 9: a multi-vendor "agentic OS," designed to run outside the CLOUD Act and in a partitioned environment, claiming £6M in annual recurring revenue (self-reported, source PRNewswire). The exact figure matters little. The signal, though, is clear: customers are already paying not to depend on a single flag or a single vendor. For a Maison used to doubling each of its critical sources, the reversibility of its agent stack stops being a virtuous intention and becomes a line on the bill — the one no executive committee will be able to ignore for long.
Sources:
Three days online, then Washington cut the power.
On June 9, Anthropic puts two models online: Fable 5, open to the public, and Mythos 5, in restricted access — the absolute frontier of the generation. Three days later, they are unplugged. Not down: cut off, on a government's order.
On June 12 at 5:21 p.m., New York time, U.S. Commerce Secretary Howard Lutnick sends a directive to Dario Amodei, head of Anthropic — it is the press that reports the exchange (Fortune, NBC, TIME). In the name of export controls and national security, access must be closed to every foreign national. The grounds: an alleged security circumvention on Mythos, which Anthropic disputes. Only, Anthropic cannot isolate the targeted users alone. So it cuts everything. For everyone, everywhere. Only Opus 4.8, the previous generation, is spared. In its statement, the vendor apologizes for unplugging its own product.
The timing adds the irony. On June 10 — two days before the cutoff — Amodei published an essay calling for the state to be granted the power to block or withdraw frontier models judged dangerous. He got the stick, and was promptly struck with it: on the 12th, Washington cut off his own. In the background, the stock market — Anthropic filed its IPO paperwork on June 1, OpenAI a week later, each valued around 900 billion dollars. Showing the power to support the valuation, asking that it be regulated, watching it get cut off: three beats in twelve days.
The reversal is there. It is not a client complaining about its vendor: it is the vendor announcing it no longer has the controls. Dependence played out between two — a client and its provider, bound by a contract you reread, renegotiate, break. On June 12, a third actor sat at the table, a state, and it ruled in place of the other two. The contract was worth nothing anymore, because Anthropic was no longer master of what it was selling.
This is a precedent, not a news item. Export controls on AI have existed since 2022, but they targeted the chips, the hardware, the upstream. For the first time, it seems, a state has a frontier model cut off that was already deployed, in service, at work in the agents and products of millions of users. The geopolitics of AI no longer plays out only in the semiconductor plants: it plays out in access to the model, which a government can shut from one morning to the next.
And what goes dark with the model is neither a storefront nor a website: it is the operational core. The agents that, backstage, forecast demand, distribute stock across boutiques, adjust prices market by market. What never speaks to the client — and what, that morning, fell silent for whoever had bet everything on that single model.
Sources:
Changing the model without ever changing the Maison.
The June 12 cutoff did not only unplug agents. It raises a fundamental question: what does a Maison truly own of the intelligence that runs its business — and what is it merely renting?
A Maison lives on two clocks. The clock of the quarter — listed, accountable to its shareholders, results to hold. And the clock of the decade: its value rests on a patrimony that passes on and that no quarter manufactures — the savoir-faire, the archives, the client relationship. The intelligence that now irrigates its business belongs to that second clock. This is what I call the patrimoine agentique, the agentic estate: the stack of agents a Maison owns, documents, governs and passes on. Most do not own it. They rent it, and believe the opposite.
The misunderstanding stems from a sound management habit. Everywhere, a good technical leader pools: a single client-relationship tool, a single e-commerce solution, one provider per function. Maintaining the same brick twice is costly, so you simplify. Agentic AI is the first exception, and it is unprecedented. Where the client-relationship tool stays confined to one function, the model that runs the agents is transversal: it touches operations, logistics, pricing, service, every flow at once. Entrusting that layer to a single vendor is a dependence of unprecedented scale.
And these deployments are no longer pilots. At Pandora, an agent built on Salesforce Agentforce resolves 60% of client requests with no human intervention, across 45,000 conversations a month. Mandarin Oriental has its kitchen waste weighed by a vision AI — down 36% in waste, 207,000 dollars saved a year across four hotels (Winnow) — and saw its room-service revenue climb 54% with an ordering assistant (IRIS). Prada has driven its leather-goods production and procurement on the o9 platform since April. Coty is targeting twenty million dollars in savings by entrusting its demand forecasting to AI. Service, kitchen, production, supply chain: these are no longer storefronts wired to a model, they are crafts — each rented from a vendor, each a dependence that almost no one governs as an asset.
This dependence turns against you in three ways — each with its counter, already tooled, all on the same floor: that of the brain that plans and orchestrates, which can be decoupled from the model that executes it.
THE PRICE THAT SPIKES WITHOUT WARNING
A vendor revises its rates, you take the hit. OpenAI bills its latest generation, GPT-5.5, at double the previous one. For a Maison that has entrusted an agent with demand forecasting across fifty markets, this is not a spreadsheet line: it is a margin redrawn overnight. The counter: routing that distributes the work across several models by task — routine computation to the cheap engine, the hard arbitration to the frontier — and switches if one goes off the rails. Gateways already do it without rewriting the agent: LiteLLM, open source, or OpenRouter. Gartner anticipates that 70% of teams running several models will go through this kind of gateway by 2028, against a quarter in 2025. Multi-model is no longer the avant-garde, it is the infrastructure becoming standard.
THE BUNDLED-IN THAT GOES PAY-AS-YOU-GO
More insidious, because we lived it this week. A service included in a plan is pulled out and goes to full price: on June 15, Anthropic pulled automated usage out of its Claude plans to bill it at full rate. Boris Cherny, at Anthropic, owns it — these plans "were never designed for the continuous, automated demand of these tools." The counter: separate orchestration from the model, measure what each agent costs, keep an alternative ready to plug in. The argument is airtight, because the vendor itself builds the fallback — Claude Code ships a setting that switches to a backup model under overload, and Anthropic's documentation notes that "several large enterprises" go through LiteLLM to govern several vendors at once.
THE MODEL THEY SWITCH OFF
The cutoff, the outage, the regulatory decision — June 12 is the extreme case, told above; last year, OpenAI was down for more than fifteen hours straight. The counter is not called "a spare vendor": it is a second, then a third vendor already wired in and tested, whatever its country, doubled with a written continuity plan — not thought through, written, the way a Maison writes its product-recall plan before the crisis. Mistral takes its place here as one brick in the portfolio: its open-license models wire in as fallback, and its European anchoring counts — but it remains one option among others, not the sovereign replacement that would settle everything. Deliverance AI, out of the shadows in early June, already sells a multi-vendor agentic system outside the CLOUD Act: customers pay not to depend on a flag or a vendor.
One word remains to be put back in its place: sovereignty. An American cloud with a data center in the Île-de-France sells residency — your data sleeps in Europe — not sovereignty. As long as the parent company answers to an American court, the CLOUD Act applies wherever the data sits. The only test that counts: who can be compelled by a judge to cut off access?
Owning your agentic estate is not internalizing everything — it is keeping the controls in the right place. An agent you cannot unplug from its vendor in a day is not an asset, it is a liability on the balance sheet. The trap is not American, it is single-vendor: the only sovereignty that holds is to multiply providers and own what cannot be rented. Compute, for its part, is rented without shame — from several, precisely. What is owned and passed on is the data, the distribution, the client relationship, and the logic that orchestrates the agents. A Maison's agentic estate is exactly that: being able to change the model without ever changing the Maison.
Sources:
Everything with one, or several exit doors?
To a Maison's CIO/CTO.
Diversifying your AI vendors or pooling everything with a single one: which way do you really lean — and is it the same trade-off whether you run a large group or an independent Maison?
The question is not rhetorical. The large group has the means to keep several providers running at once, but also everything to gain from the economies of scale of the single window. The independent Maison does not have the resources to maintain ten stacks in parallel — and finds itself tempted by the single vendor, exactly where it is most exposed. The risk and the capacity to guard against it are not aligned. That is the whole problem, and I do not have a single answer that holds for both.
The ground decides better than the doctrine: it is on this trade-off that the Maisons who decide will part ways with those who endure.
To dig deeper.
- Empire of AI, Karen Hao (Penguin Press, 2025). The investigation that shows how a handful of players concentrated computing power. To grasp why your vendor holds, for its part, all the levers.
- Chip War, Chris Miller (Scribner, 2022). The war over semiconductors and export controls. The June 12 precedent is merely its software extension — and it reads like a geopolitical thriller.
- The Coming Wave, Mustafa Suleyman (Crown, 2023). The DeepMind cofounder on a technology that can no longer be contained. Asked by someone who is building it, one question: can we still keep the controls?
- Technological dependence on American software and cloud services, CIGREF / Asterès (April 2025). The pricing of what we call "sovereignty": 264 billion euros leaving each year toward American vendors, and a CLOUD Act that prevails even over data hosted in Europe.
The sector goes to the test of the numbers.
The coming weeks put the sector to the test of the numbers. From June 23 to 28, menswear shows in Paris; Haute Couture follows, July 6-9. Then come the accounts: LVMH publishes its first half on July 28, the other groups in its wake — there we will read, or not, the trace of AI in the margins. And on August 2, Europe arms its powers over general-purpose models: the first deadline where the dependence we are talking about stops being a debate and becomes an obligation.
Behind every date, the same question: who holds the keys? Luxe oblige!
Sources:
— Mickaël Tsakiris
Paris, Thursday, June 18, 2026