I have been scratching my head on this “bubble” business for a long few months since starting this company.
It’s becoming more and more apparent that frontier AI is a bubble. At the same time, the work you can get done with AI these days, especially in coding and tech, is nothing short of incredible. I know a lot of people will scoff at the concept of “vibe coding” and “10x engineers” but the truth is, working with AI can so much more complex than the model, the prompts, or event the content you give it to work with.
I’m talking about things like skills, context management, coding CLIs, IDEs, guardrails, sandboxes, subagent driven development, loop engineering, harness engineering… it’s a lot more complicated than saying “Hey Claude, build me an app, and make no mistakes”.
One thing I see a lot of people focus on is the model. Fable 5, GLM 5.2, GPT-5.5, Qwen 3.6… there are a ton of favourites there (and don’t trust anyone that trusts Grok). But, is the model really the most important part of performance?
The thing is, we’re reaching a point with the models that we’re seeing less and less return on investment for further training. And that’s not good, since so many investors have already thrown trillions of dollars at this thing. That’s why they’re building so many datacentres. Polluting our water. Burning our planet. To reach standardized benchmarks. To avoid this bubble popping.
But there’s no way to stop it because the returns you get from a bigger datacenter are just not what they used to be. The open weight models are rapidly closing the time gap behind the frontier ones at this point. We will get to a point where Fable-level models will be able to run on consumer-grade technology (think a laptop under €3k).
There is just no form of economics that supports the madness behind frontier-model investment at this point. Even though I believe in using AI, I really cannot fathom why anyone would be investing in this particular part of innovating in AI right now.
I am trying to find ways to use AI responsibly, and that’s why I have focused so much on EU Sovereignty. The bubble as described above will pop and I’m glad that it’s everyone but Europe who managed to build a truly impressive frontier model, because the real “value” of AI is not owning a model that works well, but what problems you solve with the models you have available (such as running the more boring parts of your business so you can focus on creativity), and keeping your data safe while you do so.
So, while Europe still hasn’t succumbed to the rest of the world’s AI madness about opening datacentres, spending trillions of Euros, and actually getting very little out of it, I’d really like to suggest pivoting to our tooling, helping those with an entrepreneurial spirit build non-AI businesses with the help of AI, and actually creating something of value (you know, that humans can use).
Where to start: the biggest gains in AI performance right now are found in AI Harnesses#
Multiple experiments are showing that AI Harnesses can take a regular model and upgrade it to something much more impactful and useable.
An AI Harness is a concept that first got traction in December 2025. It’s hard to say what the first “Harness” really was because to some people, it needs to be near-to-fully autonomous, and to others, it’s as simple as an LLM “wrapper”. Depending on who you ask, it can change.
The most popular harness and model, by my perception of the AI world, is Claude Code and its Opus/Sonnet/Fable models. It’s what I have used to build most of my projects. However, it’s a major PITA to use Claude models in a way that is acceptable to EU standards.
- You need an enterprise account (5 seat minimum) or expensive API-based usage. Without this, there is no DPA.
- If you’re a solopreneur, like me, and therefore using a consumer-level subscription to Pro Max, you cannot let it know anything about your clients. So, only some work can be done via Claude (we’ll go over this in the first step below).
- Even with a DPA, Claude’s infrastructure is all American. Yes, they have reasonably tight security guardrails in place, but you’re still putting your data in the USA. Even if they’re not training the model on it, a lot of people don’t trust that.
So, how can you use Claude Code as a Harness, but on EU infra?
Well, if you already have Claude Code installed (as a CLI, not within the Claude Desktop App)…
1: First, decide if you want to use Scaleway, Mistral, or both for your API key#
You can get 1 million tokens free on Scaleway to start testing that your API calls work, as well as other new signup benefits. They don’t do affiliate links, but I’d put one here if I could.
Mistral also has a free tier for their API, but you need to have a paying contract (such as pay as you go) to benefit from their DPA.
My work is now divided into tiers:
- Tier S: Scaleway for my sensitive clients.
- Tier C: Mistral for my clients who need GDPR (all of them) but can tolerate a frontier model. Still, I have toggled off the “train on data” option.
- Tier D: Claude for my personal work.
Some suggest using it on AWS with EU data residency for this hidden Claude GDPR trap. But, it’s more expensive than Scaleway for the same open weight models (such as GLM) and despite having EU data residency, and, it’s still technically vulnerable to the American CLOUD Act.
They DO offer the Claude models, but they’re hilariously expensive for the performance advantage you’ll get as compared to GLM on Scaleway or Mistral direct.
2: Next, connect your API to the backend of Claude Code.#
In order to make this happen, you will need to set up a worker that will take the expectated Anthropic API format and convert it to the OpenAI standard (which is what Scaleway uses). For this, they recommend y-router. To set it up locally, you’ll also need docker. The y-router documentation goes over this.
You can choose whichever models you want to “mimic” the setup of Claude code.
export ANTHROPIC_DEFAULT_HAIKU_MODEL=<scaleway model name>
export ANTHROPIC_SMALL_FAST_MODEL=<scaleway model name>
export ANTHROPIC_DEFAULT_OPUSMODEL=<scaleway model name>
export ANTHROPIC_DEFAULT_SONET_MODEL=<scaleway model name>You can also set up Opus Plan, where you use your most powerful (expensive) reasoning model for planning, and you can use GLM as a powerful coding model with reasoning switched off (Scaleway documentation).
I’d recommend working through your choices during a conversation with the real Claude. Talk about how you like to work, and Claude can make some recommendations.
It does feel funny to use Claude to get off of Claude, but here we are.
Claude also helped me set up a few helpers for token usage, which can quickly mount up on Scaleway (no prompt caching for example).
These include:
And we set up prompt caching for Mistral, enabled via y-router. I also have different subagent directives for Scaleway vs Mistral.
Mistral is really my “workhorse” for client projects. It’s not economical currently to use the Scaleway infra for day-to-day work due to the prompt caching issue.
3: Test it out. Change models if they don’t suit you. Research how you’d like to modify your harness or add other prompts to your Claude.md files.#
This isn’t the end-all-be-all of harnesses. There are a ton more options out there for improvement.
You’re also giving up some benefits from Anthropic, like its cost-optimized caching. You can “kind of” do this on Mistral but it’s not as good. Mistral also sometimes forgets to follow some of your “rules” in your Claude.md files, which means you need extra guardrails in place. Skills are also a bit more of a headache - you need to really decide which ones you want to include in your project, because if you bloat your repo, you will end up burning through tokens at a crazy rate. Some skills, like reducing verbose output via caveman or another method, can be reduced to a line in a Claude.md file.
This setup will force you to be a bit more thoughtful about your vibe coding, agentic engineering, or whatever you want to call it. But isn’t that why you’re here anyways? To be more thoughtful about the data that you want to protect?
I can’t just dump entire webpages into this anymore and say “help me use the best practices from this page on my site”. I need to expertly provide the exact guidance. I’ve gotten closer to my code, and I think it’s actually a great way to push back on the trend of letting Claude Code run the whole show.
It’s still much faster than doing this without AI.
If you need any help setting this up for solo reasons, or if you want a consultation on getting your team onto a Claude Code or custom harness with EU sovereign LLM model delivery infrastructure, please get in touch.