Eighteen days, and Europe was not in the room

I want Europe to succeed. Genuinely.

In June, that wish met a practical test. A US government order made Anthropic's newest public model unavailable worldwide. The interruption lasted eighteen days. Later that month, OpenAI began the release of GPT-5.6 only after it gave the US government access to its plans and partner list.

European customers depended on both decisions. No European institution or customer took part in them.

During the same month, the European Commission published its Cloud and AI Development Act proposal. It is the most serious EU sovereignty proposal so far. It addresses real problems. What it does not, is provide a backup model during either incident.

Eighteen days, decided elsewhere

Anthropic released Claude Fable 5 and Claude Mythos 5 on June 9. Fable was the public model. Mythos had fewer safeguards and was available only to selected cyber-defence partners.

On Friday, June 12, the US government ordered Anthropic to block both models for every foreign national. The rule applied inside and outside the United States. Anthropic could not verify nationality in real time, so it disabled both models for all customers. The government lifted the controls on June 30. Global access to Fable returned on July 1. https://www.anthropic.com/news/fable-mythos-access, https://www.anthropic.com/news/redeploying-fable-5

While Fable was unavailable, OpenAI started a limited GPT-5.6 preview on June 26. OpenAI said they had shared plans, the model capabilities and the names of participating partners with the US government and that this process should not become the normal method for model releases. https://openai.com/index/previewing-gpt-5-6-sol/, https://techcrunch.com/2026/06/26/openai-limits-gpt-5-6-rollout-after-government-request-says-restrictions-shouldnt-be-the-norm/

The US Commerce Department cleared a wider release on July 8 after more tests by its Center for AI Standards and Innovation. OpenAI staff also travelled to Washington to answer questions. The wider rollout began on July 9. https://thenextweb.com/news/openai-gpt-5-6-broad-rollout-us-approval, https://openai.com/index/gpt-5-6/

Both cases ended after work in Washington. European procurement rules and assurance labels did not change the result.

Europe was not in the room.

These were also not normal cloud outages, where a second region can protect against a failed data centre, or another provider can protect against a failed control plane. Neither protects against a legal order that applies to the model itself. If the backup endpoint serves the same restricted model, it is not independent.

Europe was working on a different clock

The Commission adopted its CADA proposal on June 3, nine days before the Anthropic shutdown. The proposal covers research, infrastructure capacity and a common EU framework for cloud and AI sovereignty. It also aims to triple the European data-centre market within five to seven years. https://digital-strategy.ec.europa.eu/en/library/proposal-cloud-and-ai-development-act-cada, https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex%3A52026PC0502

Under Title III, each member state would have six months after the regulation enters into force to designate at least one data-centre acceleration zone. A qualifying project would enter a green corridor with a permit procedure of no more than twelve months. These clocks have not started because CADA is still a legislative proposal. https://www.euronews.com/next/2026/06/22/eus-cloud-and-ai-development-act-gets-mixed-reception, https://eur-lex.europa.eu/legal-content/EN/HIS/?uri=CELEX%3A52026PC0502

The EU already has a more immediate tool.
The Data Act has applied since September 2025. It requires cloud providers to remove obstacles to switching and to support the customer's exit plan. The standard transition period can still be as long as 30 days. More important, a right to export data does not create model weights, compatible behaviour or available compute. https://digital-strategy.ec.europa.eu/en/factpages/data-act-explained, https://eur-lex.europa.eu/eli/reg/2023/2854/oj?locale=en

Besides a powerful infrastructure, Europe needs an operational answer now.

European providers come into play

European cloud providers and AI labs have a good opportunity. They should host the strongest open-weight models on infrastructure under European operational and legal control.

OVHcloud offers a serverless catalogue with more than 40 models, including Llama, Qwen and DeepSeek. Scaleway offers open models from French data centers, OpenAI-API comaptible. https://www.ovhcloud.com/en/public-cloud/ai-endpoints/, https://www.scaleway.com/en/generative-apis-signup/

API compatibility is useful, but it is not model compatibility. Providers differ in tool calls, structured output, context handling, safety rules and version changes. A real fallback therefore needs a common internal interface and an evaluation set for the actual workload. Changing the base URL is only the first step.

Kimi K3 is the current test, which Moonshot released it's full weights under the Kimi K3 License.
The model has 2.8 trillion parameters and activates 104 billion for each token. It supports a context of ~1 Mio tokens. https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md, https://arxiv.org/abs/2607.24653

Inference is hard

K3 uses four-bit weights. The raw weights therefore need about 1.4 TB before metadata, caches and runtime overhead. The vLLM deployment recipe specifies at least eight GB300 accelerators and recommends several nodes for production traffic. https://recipes.vllm.ai/moonshotai/Kimi-K3

The license matters as well. A model-as-a-service company with more than $20 million in annual group revenue must make a separate agreement with Moonshot before commercial use. Kimi K3 is open-weight, but it is not an unrestricted public asset. A European provider must solve the contract and the serving system. https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE

Demand provides the final warning. Moonshot stopped new subscriptions within days of launch because its compute capacity was almost full, I'm on the waitlist ever since and still haven't secured a subscription. https://apnews.com/article/4c66a2e0f557ce79d3cc2d769c9a6226 Z.AI released GLM-5.2 with MIT-licensed weights and a 1 Mio. tokens context window. https://huggingface.co/zai-org/GLM-5.2

European hosting must still be a good service. Customers need the exact model version, clear prices, useful quotas, measured throughput, a service target and a retirement policy. An old model with poor throughput is not sovereignty. It is a weak product with a flag on it. Europe does not have to win the foundation-model race, but it can become the best place to run strong open-weight models, at least for europeans. The customers are here. (Some weights are here toohttps://www.soofi.info.)

Upgrade the office laptops

A smaller form of the same problem is on our desks.

People discuss on-device AI as part of European resilience. Yet I still see many 16 GB Excel-machines in the European tech sector. Come on. That is not enough for a useful local-model strategy.

The basic calculation is simple. A 27-billion-parameter model at four bits needs about 13.5 GB for its weights alone. One published MLX measurement reported about 14.5 GB of resident memory for a quantised Qwen 27B server. A 16 GB machine then has almost no space for the operating system, applications or a larger context. Local inference also needs memory bandwidth and a supported accelerator. On Apple silicon, MLX lets the CPU and GPU use the same memory without copying arrays between separate pools. On many other laptops, system memory and accelerator memory are separate. https://github.com/ml-explore/mlx

It's an expensive moment to ask for more memory, when DRAM contract prices are to rise by more than 100 percent from the previous quarter in early 2026. Micron reported that its average DRAM selling price for the first nine months of fiscal 2026 was about 140 percent above the same period in 2025. https://www.trendforce.com/presscenter/news/20260202-12911.html, https://investors.micron.com/static-files/e18b3c93-8b84-411b-94eb-517b018d9dab

That changes the price of the decision. It does not change the memory requirement.

Companies, universities and public institutions should specify local-AI hardware by workload. They need enough memory, enough bandwidth and an accelerator that the selected runtime supports. Then the system with a small local model and a stronger cloud adviser, which I described a few weeks ago, can work. https://philipbrembeck.com/writings/2026/07/only-as-much-intelligence-as-you-need

I want Europe to take this seriously

I want European labs at the frontier. But sovereignty cannot exist only in future plans.

European providers can host strong open-weight models now. European organisations can keep state portable, reserve independent capacity and test failover now. CADA can build the larger base over the next five to seven years.

During those eighteen days, Europe did not need to own the world's best foundation model. It needed the legal right, model weights, compute capacity and prepared systems to run a strong alternative here.