Deutsche Fassung: Europa ist es leid. Technisch aufzuholen ist trotzdem noch keine Strategie
WIRED captures the European mood in a useful headline: “Europe
Is Fed Up and Wants Its Own AI”. The impulse is understandable.
Europe wants less dependency, more compute of its own, its own models
and infrastructure that does not become nervous every time the
geopolitical weather changes.
Yet the reaction contains a dangerous narrowing. A continent can
catch up technically and still miss the actual transformation. A
European fab can produce chips. It does not decide how schools organise
learning, how public administrations assign responsibility, how
companies protect human judgement or how the public sphere deals with
synthetic powers of persuasion.
AI is not one technology sector among others. It acts simultaneously
on work, education, public administration, the public sphere, value
creation, power architectures and the organisation of judgement. It
changes not only how we work. It changes what will count as work,
competence, career, institution and human contribution.
Technical catch-up is therefore necessary but insufficient. The more
important task is to shape the paradigm shift.
Catching up is not a
direction
The European debate likes to count data centres, parameters, chips
and investment totals. That is not wrong. Infrastructure determines who
can act at all. Without compute, energy, networks, open models and
skilled people, sovereignty remains a conference term with excellent
typography.
But infrastructure does not answer where a society wants to go. A
continent can own a capable model and remain dependent on a single cloud
provider, a proprietary orchestrator, an opaque supply chain or a
handful of consultancies that are the only ones who know how the system
works. Ownership alone does not create agency.
From a New Work perspective, that is the decisive distinction.
Freedom does not emerge from an object located somewhere in Europe. It
emerges from capabilities, rights and real alternatives. People and
institutions must be able to understand, limit, change and challenge a
system. Otherwise we will have European hardware and imported
powerlessness.
Sovereignty begins with
exit capacity
A useful working definition follows: AI sovereignty is the ability to
make dependencies visible, limit them, switch providers and continue
critical work when a contract, an API or a political relationship
fails.
This is not a demand for autarky. Europe does not need to manufacture
every model layer, chip and tool itself. It does need exit capacity.
Data, prompts, evaluations, logs and workflows must be exportable.
Critical processes need a degraded mode. Interfaces must be replaceable.
The first exit test cannot take place on the day of the crisis.
Procurement often asks about price, accuracy and speed. Sovereign
procurement also asks: How long does exit take? Which capabilities do we
lose while operating the service? Who can inspect the system without the
vendor? What remains functional when the connection disappears?
That sounds less spectacular than “Europe’s own AI.” It is much
closer to a Tuesday morning when a public agency or a company actually
has to keep working.
A server location is
a fact, not a strategy
EU hosting can be useful and necessary in sensitive contexts. It does
not automatically resolve questions of corporate jurisdiction, key
control, access, subcontractors and provider switching.
The EU
Data Act contains safeguards against unlawful third-country
government access and rules intended to make switching between
data-processing services easier. US law, meanwhile, can require covered
providers under specified conditions to disclose data within their
possession, custody or control regardless of whether the data is located
inside or outside the United States; the territorial principle appears
in 18 U.S.C. §
2713.
Which obligation applies in a specific case is a legal question, not
a LinkedIn punchline. The architectural conclusion is still clear: a
Frankfurt postcode does not replace analysis of jurisdiction,
encryption, key ownership, the supply chain and the exit path.
What colibri actually proves
The open-source colibri project expands
the technical design space. It can run the 744-billion-parameter GLM-5.2
mixture-of-experts model on a machine with roughly 25 GB of RAM and no
GPU. Part of the model remains resident in memory, while expert layers
are kept on an SSD and streamed when needed.
That is remarkable. It is not fast. The project documents roughly 370
GB of storage, about 11 GB of disk reads for each cold token and
approximately 0.05 to 0.1 tokens per second during cold decoding. This
is not a citizen-service chatbot and it is not proof that every standard
laptop can now run a frontier model in production. It is a proof of
feasibility: “too large to run locally” is not an immutable law of
nature.
The practical lesson is not to install colibri everywhere. It is to
build more differentiated architectures. Sensitive tasks may run on
smaller open models locally when quality and risk allow it. Other tasks
may use an external frontier service. Between them, organisations need
gateways, evaluations and replaceable interfaces. Sovereignty lies in a
designed portfolio, not in a heroic all-or-nothing decision.
The organisation of
judgement
This is where the actual New Work question begins. If only an
external vendor or a small specialist team understands how an AI system
reaches and operationalises decisions, the organisation is not merely
renting compute. It is renting judgement.
Workers can then become final-stage validators for a machine whose
assumptions they are not allowed to change. They remain responsible for
outcomes but cannot access the rules, logs or escalation paths. This is
not relief from work. It is responsibility without agency.
A small study of AI-assisted writing, “Your Brain on ChatGPT”,
found lower neural connectivity and a lower sense of ownership among the
LLM group across 54 participants in its first three sessions. This is
not proof of general cognitive decline and certainly not a finding about
every occupation. It is a bounded warning: when systems are designed so
that people merely accept outputs, learning and judgement can disappear
from the work process.
Sovereign organisations do the opposite. Domain experts define error
and action boundaries. Operations teams can inspect logs and roll
systems back. Workers and their representatives have intelligible rights
to challenge and stop systems. Learning does not occur only in an annual
course; it happens in the real work of shaping the system.
A Monday-morning test
Before using the label “sovereign,” six questions should be
answered:
- Can we name the full technical and legal dependency chain?
- Can we export data, workflows, evaluations and logs in usable
form? - Can we replace the model or provider without reinventing the
organisation? - Can critical services continue in a degraded mode during failure or
dispute? - Can the people who carry responsibility inspect errors, contest
decisions and stop the system? - Does our own capability grow during operation, or only the vendor’s
invoice?
This test connects technology, law and work. That is exactly what is
required for a transformation force that changes several social orders
at once.
Europe does not need a
victory pose
“Europe is fed up” can be a productive beginning. Fatigue does not
build an institution. A fab, a model and a European data centre can be
important freedom substrates. They do not replace the work of
redesigning education, public administration, co-determination, the
public sphere and the distribution of judgement.
Europe’s decisive capability is not to own everything. It is to set
direction together: to see dependencies, preserve alternatives, build
capability, assign responsibility and enable dissent both technically
and institutionally.
Then sovereignty becomes more than origin. It becomes practice. And
technical catch-up finally becomes what it should be: a tool for
societal design, not its substitute.
Further reading: New Work New Culture’s
Flowbook AI