Let's be clear: I'm French, so complaining is basically my cardio.

But this isn't a complaint. It's a war report.

On June 13, 2026, the US government blocked Fable 5 for anyone without a US passport. Just like that. No warning. No appeal. The first shot had been fired.

I saw the tension building the day before at Tech Fest Nancy, where François Zaninotto's talk on AI-assisted development sparked a debate that's now impossible to ignore: who controls the tools that control our code?

Tech Fest Nancy: Where It All Clicked

Last Friday, June 12th, I was at Tech Fest Nancy. Kalvad sponsored. We drank coffee, ate madeleines, and listened to talks worth the trip.

Then François Zaninotto took the stage. His topic: AI-assisted development tools. Claude Code, Copilot, Cursor. The tools that, in 2025, made such a qualitative leap that a lot of devs can no longer imagine coding without them.

How do we master these tools? How do they change our profession? Does it even make sense to learn programming anymore?

Good questions. Not the ones that made me sit up.

What got me was what he didn't say. The elephant he sidestepped in an already cramped room: sovereignty.

The debate detonated after his talk. Sovereignty. Security. Control. Who owns what? Who controls what? And above all: what exactly are we shipping to these LLMs?

Spoiler: most teams have no idea what they're shipping.

The Trigger: Fable 5 Goes Dark Overnight

The next day, Saturday June 13th, it stopped being theoretical.

The US Commerce Department hit Anthropic with an export-control directive: suspend all access to Fable 5 and Mythos 5 for any foreign national, whether inside or outside the United States. That includes foreign nationals working at Anthropic. Citing national security.

Anthropic's only way to comply was to pull the plug for everyone. Not just foreigners. Everyone. A model deployed to hundreds of millions of people went dark overnight.

The stated reason was a narrow cybersecurity jailbreak finding. Anthropic disputes it, calls it a misunderstanding, and says it's working to restore access. Maybe they win that fight. Doesn't matter for our purposes.

Because here's the part you need to sit with: it doesn't matter whether they were right. What matters is that a single directive, on a Friday evening, took a frontier model away from the entire planet with no warning, no transition, no recourse. Your access to a critical tool can evaporate because of a dispute you're not even a party to.

Welcome to the AI War. And you're already conscripted. Every prompt you send to a cloud LLM, every snippet you paste into Copilot, every log you pipe through a SaaS tool is a round in someone else's chamber.

You Do Not Own Your Data. Period.

But Loïc, we have contracts! Confidentiality clauses! GDPR up the wazoo!

Cute. You think a contract stops a government with an export-control directive? Ask Anthropic how much their Terms of Service mattered at 5:21pm on a Friday.

Take GitHub Copilot. You ship your code. All of it. The comments. The variable names. The secrets you almost all removed. Your architecture patterns. Your mistakes. Your fixes.

And you do it... why? Because it's convenient. Because everyone's doing it.

You're trading your technical soul for a 10% productivity bump.

And the worst part: you don't even know what you traded.

What You're Actually Sending

Quick exercise. Take your last project, the one wired up to Copilot or Cursor. What left the building this week?

• The code. Obviously. But not just the code.
• Filenames. internal/audit/tax_calculator.py tells a story about your business.
• Comments. "TODO: fix this before the audit" is a little too honest, no?
• Error logs. Stack traces that expose your stack, your versions, your vulnerable deps.
• Diffs. Exactly what changed, and why.
• Search queries. "how to bypass rate limiting on Stripe API." Really?
• Test data. With real card numbers, real names, real everything.

But Loïc, we have filters! We have rules!

Wonderful. And the day a dev skips the filter because they're in a hurry? The day a bucket gets misconfigured? The day a government decides "national security" means your data too? Your rules are a screen door on a submarine.

Your data is not yours.

Sovereignty Is Not a Luxury. It's a Decision.

OK, but we're a small shop. We can't afford to self-host.

Wrong. Sovereignty isn't a question of size. It's a question of will.

You think only the Fortune 500 can control their data? Most of them don't. They handed it to AWS, Azure, and Google Cloud because it was easier, because nobody ever got fired for choosing AWS.

So most companies, big and small, have already lost. You don't have to follow them off the cliff.

Self-Host or Die

At Kalvad, we self-host. As much as we can.

Not out of purism. Not out of cloud-hatred. Out of something for freedom.

When you run your own LLM:

• Your data stays yours. No leaks. No "oops, we trained on your private repos."
• Your latency is local. 4ms versus 400ms. Guess which one your users notice.
• Your costs are predictable. No month-end surprise because a prompt ran long.
• Your compliance is real. GDPR, HIPAA, classified: you own the infrastructure, you set the rules.

But Loïc, open models are worse!

The eternal excuse. "We can't self-host because GPT-5.5 is better."

First: less true every month. Mistral, Qwen or DeepSeek ships models that trade blows with the Americans, and they're not alone. Second, and more important: quality isn't the only axis.

A slightly weaker model that's 100% under your control beats a perfect model that can be confiscated on a Friday evening. Ask anyone who built on Fable 5 this week.

And you don't have to pick one. Self-host the sensitive work. Use the cloud for the rest. The point is that you choose, not a directive you never read.

Don't be the villain of your own story.

You Are the Product

You know the line: if you're not paying, you're the product.

With cloud LLMs you are paying. And you're still the product.

Every prompt, every result, every interaction is data. Data used to:

1. Improve their models. With your work. Without asking. Without paying you.
2. Sell insights. "French companies stick with this framework three years longer than Americans!" Thanks to whom?
3. Build lock-in. The more you use it, the more you depend on it, the harder it is to walk away.

It's a slow pauperization. You get poorer in control, in know-how, in independence. They get richer. And one day, the access just stops. Like Fable 5.

"But We Have No Choice!"

You do.

Option 1: Keep going. Use Copilot, Claude, and friends. Ship your data to the cloud. Hope nothing breaks. One day you get the email: "Sorry, your region is no longer supported." Or worse, "your account has been suspended." Risk: 100%. Control: 0%.

Option 2: Take it back. Self-host your tools. Run open models. Train your people. Become independent. Risk: low. Control: 100%.

But Loïc, it's complicated!

No. It's different. New skills, some invested time, some responsibility. You know what that's called? Being an engineer. Not a button-pusher clicking UIs they don't understand. An engineer who knows their tools, controls their infrastructure, and owns their data.

Where to Start

You're convinced. Or at least intrigued. Begin here.

Step 0: Audit what you're sending

Before you change anything, know what's leaving. Stand up a local proxy that logs every LLM API call. Analyze a week of traffic. You will be horrified. Good.

Step 1: Replace the worst offenders first

Start with whatever leaks the most sensitive data.

• Code completion: swap Copilot for Tabby against a local model.
• Chat: Ollama, LM Studio, or AnythingLLM on local weights.
• Search: Quickwit for self-hosted retrieval.

Step 2: Pick your models

No need to reinvent anything.

• Code: Kimi 2.7 Code, Qwen3-Coder-next
• General: Mistral Medium, Qwen3.5
• Embeddings: Qwen3-embedding

Step 3: Self-host properly

No Kubernetes. No Helm. No YAML cathedral.

• Inference: vLLM or llama.cpp
• Serving: FastAPI + Uvicorn
• Edge: HAProxy out front
• Provisioning: PyInfra, because Ansible is so 2015

"But It's Expensive!"

No. It's insurance.

A capable GPU runs USD2k to USD10k to buy. A 4x RTX6000 box, around USD40k. Real money. Now price the alternative: a breach is millions. Cloud lock-in is hundreds of thousands a year. Losing access is your whole business.

You pay a little now so you don't lose everything later. And you can start small: one RTX 4090 in the corner, a 7B or 13B model, and you've covered 90% of your use cases.

Choose Your Side

On June 13, 2026, the AI War stopped being a metaphor.

On one side, the people who want to control AI. On the other, the people who only want to use it. On one side, those who want to own your data. On the other, those who want to protect it. On one side, button-pushers. On the other, engineers.

At Kalvad we've chosen. We self-host. We control. We own.

Because sovereignty is not a luxury. It's a necessity. Because your data is not their oil. It's your gold. Because AI is not their revolution. It's yours.

Read that twice. Sit with it.

If you have a problem, and no one else can help, maybe you can hire the Kalvad-Team.

Need Engineers Who Actually Understand the Stack?

Tired of button-pushers who don't know what they're shipping? We don't do YAML cosplay. We don't do compliance theater. We build, we own, we control.

Talk to engineers, not button-pushers before the next directive leaves you on the wrong side of the firewall.

Contact Kalvad | Engineering & Technology Consulting

Get in touch with Kalvad to discuss your engineering, R&D, or technology consulting needs with our expert team.

Engineering & Technology Consulting