Digital Sovereignty

The practical sovereignty laid out in the previous chapter, that sum of choices about tools, data and jurisdictions that a territory's economic actors make every day, emerged naturally at the end of our tour through the professions, because it is in the professions that sovereignty is tested and either holds or doesn't. But it isn't enough on its own. A sovereignty policy cannot stop at individual usage; it also requires choices about infrastructure, software and data governance that go beyond the scale of a fiduciary firm or a doctor's practice. Those choices are made at different scales altogether, and that is where this chapter begins. With it opens the fourth part of the essay, the part that turns diagnosis, inheritance and sectoral inquiry into a programme.

The subject is everywhere in speeches and almost nowhere in decisions, and everyone projects onto it whatever they wish to see. Some call digital sovereignty mere compliance with the nLPD, Switzerland's revised data protection act. Others see in it a programme of massive industrial relocation, as though it were possible to build a Swiss equivalent of AWS within the coming decade. Still others wield the term as a political slogan without ever specifying what it would mean in practice. In this chapter I want to say, as honestly as I can, what digital sovereignty does and does not mean for a canton like Valais. The result will, I suspect, be less rousing than some readers hope. But it will, I believe, be more accurate, and accuracy is what matters if we want to make progress.

Three levels to distinguish

Everything that follows depends on distinguishing the levels at which digital sovereignty actually plays out. They are too often conflated, and the conflation produces either oversized projects or unjustified resignation.

First, the cloud: what is commonly meant by the large infrastructures of computing and storage. Building and operating infrastructure at hyperscaler level, capable of rivalling AWS, Microsoft Azure or Google Cloud, requires investment in the tens of billions of dollars, technical expertise accumulated over decades, and a critical mass of usage large enough to amortise those costs. This level is out of reach for a canton, or even for a country the size of Switzerland acting alone. It plays out at a continental scale, and it plays out badly there: Europe, despite the Gaia-X project and national players such as OVHcloud, has failed to build a large-scale alternative to the American and Chinese hyperscalers on this front.

Next come national public infrastructures and sovereign AI models, sized for specific uses: government, research, healthcare, defence. This level is under active construction in Switzerland. The Swiss Government Cloud, approved by the federal chambers in late 2024 with a commitment credit of more than two hundred forty million francs and a total budget approaching three hundred twenty million⁴¹, is meant precisely to give the Confederation a sovereign cloud infrastructure. Its first functions have been live since early 2026, with full deployment planned by 2032. Cantons and municipalities will be able to use it at cost, but the private sector will be excluded. Alongside the SGC, sovereign-AI initiatives are gradually building a technical substrate that did not exist five years ago. Foremost among them is Apertus, the language model launched in September 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre under the Swiss AI Initiative⁴².

And finally there is everyday use, which the previous chapter addressed in relation to the intellectual professions but which concerns the entire economic and social fabric. At this level, sovereignty has nothing to do with who owns the servers; it depends on who decides which tools to use, who controls the data, who bears legal responsibility for its processing. A fiduciary firm that chooses an AI assistant compliant with Swiss law over a more capable American tool is exercising sovereignty at this level. A municipality that hosts its services on Infomaniak rather than AWS is exercising sovereignty. A doctor who keeps patient records on Swiss infrastructure is exercising sovereignty. These choices are individual, distributed, unspectacular — but taken together they form the most tangible of the three realities.

The distinction matters because it radically changes how we should read a canton's room for manoeuvre. At the first level, Valais can do nothing on its own, and it would be wrong to pretend otherwise. At the second, it can take part in federal and European projects and defend its interests within them, but it cannot carry a substantial project at its own scale. At the third, however, it can do a great deal, and that is where it should concentrate its efforts, because that is where decisions are within its reach and the benefits most direct.

Substitution or complement: the distinction that changes everything

Before going further, let me clear up a common misunderstanding that renders most debates about digital sovereignty incoherent. To say that a sovereign European hyperscaler is very hard to build, and then to suggest that a specialised infrastructure node might make sense at the scale of Valais, can look like a contradiction. If it's difficult at continental scale, how could it become possible at cantonal scale? The question is legitimate, and it deserves a precise answer, because confusion on this point undermines any serious discussion.

The answer is that these are not the same kind of project. The difference between Gaia-X's failure and the possibility of a targeted Valais node is not simply a matter of scale. It is a difference in kind.

When a European project sets out to offer every business on the continent a generic alternative to AWS, Azure and Google Cloud — an exhaustive catalogue of services, at comparable prices, in a market the American players already serve very well — it is pursuing a strategy of substitution. It claims to replace the dominant players on their own turf, where they hold a twenty-year head start, massive economies of scale, and costs amortised across tens of millions of customers. This strategy fails almost every time, because it attacks the hyperscalers precisely where they are strongest. No amount of public investment has yet produced a European player competitive on the general-purpose ground the Americans occupy.

When a targeted project sets out to offer specialised infrastructure for uses that hyperscalers serve poorly or not at all, it is pursuing a strategy of complement. It does not claim to replace AWS for a Swiss SME's e-commerce operation — a field where the Americans are competitive and where sovereignty is, for the customer, a secondary concern. It aims instead at fields where sovereignty is non-negotiable and where the market is too niche to interest hyperscalers: medical data under Swiss law, tax data held by public administrations, linguistic corpora of regional languages, specialised hosting for academic research. In these fields, profitability does not depend on price competition with the hyperscalers; it rests on captive use, on the legal value of sovereignty, and on the specificity of the service rendered.

This distinction explains why Gaia-X stagnates while the Swiss Government Cloud moves forward. The SGC does not claim to replace AWS for the Swiss private sector — it explicitly rules that out, and the parliamentary debate said so again and again. It serves a precise perimeter, defined by the public nature of the data it hosts and by the legal requirement of sovereignty attached to that data. It is precisely because it accepts this specialisation that it is viable. The Confederation is not playing the same game as the hyperscalers; it is playing a different game, on ground the hyperscalers don't really covet.

For a canton like Valais, the same logic must apply. A regional infrastructure project claiming to offer local businesses a general-purpose alternative to commercial cloud services would be doomed to fail, by the very mechanism that doomed European ambitions. A project targeted at uses specific to the territory — where legal sovereignty is non-negotiable, or where cultural, linguistic or sectoral specificity makes the offering uninteresting to global players — is, by contrast, defensible. Provided it accepts that it is not playing the same game as the hyperscalers, and does not pretend to.

This clarification of the nature of a project, even more than its scale, is what separates a realistic digital sovereignty policy from wishful posturing.

What the Swiss Government Cloud makes possible, and what it doesn't

A few more words on the SGC, since the project will structure Switzerland's digital sovereignty landscape for the coming decade, and since it is at once useful and insufficient for what a canton like Valais needs.

Useful, because it will give the cantonal administration and municipalities a credible alternative to foreign solutions for their internal services. A Valais municipality that today hosts its document management on an American service will, tomorrow, be able to switch to sovereign federal infrastructure. That is genuine progress, and the canton would do well to position itself actively to make use of these capabilities as they become available. The SGC also sends an important political signal: the Confederation is treating digital sovereignty not as a slogan but as a costed, scheduled investment, defended before the chambers. That signal changes the climate in which cantons make their own infrastructure decisions.

Insufficient, because the SGC is sized for federal, cantonal and municipal administrations, not for the private sector. A Valais fiduciary firm, a law practice, a doctor, an industrial SME will not be able to host their operations there. Yet it is in the private sector, as we saw in the previous chapter, that the greater share of practical sovereignty is actually decided. The SGC settles, perhaps, a third of the question. It leaves the other two-thirds to the judgment of economic actors and to the Swiss commercial operators able to offer them alternatives: Infomaniak, Exoscale, Swisscom, and a few others. These operators are respected in their field, they pay tax in Switzerland, they operate under Swiss law, but they remain small next to the hyperscalers, and their offerings do not cover every need. In certain segments — large-scale generative AI, high-performance scientific computing — there is, as things stand, no Swiss alternative at the cutting edge.

This state of affairs is not a failure; it is a fact. European territories that claim otherwise are either lying to themselves or setting themselves up for costly disappointment. Realism means acknowledging the limits, the better to work within the room for manoeuvre that actually exists.

The question of a cantonal data centre

There remains a question often raised in Valais public debate: that of a sovereign data centre at the scale of the canton itself. This is the moment to answer it clearly, putting things in their proper order.

A regional data centre, small or medium in size, is technically feasible in Valais. The ingredients are all there: abundant, decarbonised hydroelectric power, an alpine climate favourable to passive cooling, available land, competent regional public operators, including Genedis, and the legal capacity of the bourgeoisies — the traditional Valais civic corporations that hold communal land and other assets separately from the political municipality — to carry long-term land projects. No structural obstacle stands in the way of such a project.

The question, then, is not whether it is possible, but whether it is worthwhile, and for whom. My honest answer is: yes, under specific conditions, and precisely on condition that it follows a strategy of complement rather than substitution, as just laid out.

A general-purpose cantonal data centre, claiming to compete with Swiss commercial operators on the open market, would be hard to justify economically. Standard uses are already covered by Infomaniak, Exoscale, Swisscom and their equivalents, at competitive cost and with satisfactory service quality. Investing tens of millions of public francs to produce a redundant service makes little sense.

A specialised node, by contrast, can make sense under precise conditions. If the uses in question are genuinely specific to the territory — resilience against winter power cuts, processing of Valais medical data under reinforced cantonal law, hosting of AI models trained on Walliserdeutsch and on Valais Franco-Provençal, digital archiving of intangible heritage listed by Unesco, scientific computing for the cantonal institutes of higher education — then a targeted project can find its profitability in the very specificity of its uses. It is not competing against the hyperscalers; it is occupying ground they do not occupy.

The scenario that strikes me as most defensible over ten years is not a grand flagship cantonal data centre, but a combination of several targeted actions. First, coordination so that municipalities, the cantonal administration and public institutions migrate their priority services to the SGC as it becomes available. Second, a deliberate partnership with Swiss commercial operators — Genedis for the regional dimension, Infomaniak or an equivalent for the national dimension — to steer private actors toward sovereign solutions compatible with their needs. And one or two targeted projects of specialised regional infrastructure, conceived from the outset as complements to national schemes rather than as their competitors, on condition that rigorous analysis first demonstrates their relevance.

This approach is less spectacular than a grand cantonal data centre. It is probably more useful.

Energy, climate and computation

One question remains that this essay has so far skirted and can no longer avoid: the energy cost of artificial intelligence. The question has become central to international public debate, and a forward-looking essay on the relationship between AI and territory that failed to take it seriously would, rightly, leave its reader puzzled.

The orders of magnitude are alarming. Training a large contemporary language model consumes the equivalent of several hundred European households' electricity use over a year. Inference — the everyday use of these models by their users — consumes less per query, but at the scale of hundreds of millions of users it already represents a measurable share of global electricity consumption. Projections over the next three years suggest that data centres dedicated to AI could, on their own, consume as much as some mid-sized Western European countries. This trajectory is not sustainable as it stands. As it unfolds, it will raise sharp political questions about how to allocate electricity between AI, the climate transition of other sectors, and basic household needs.

On this front, Valais holds a position that few European territories can claim. Its hydroelectric power, abundant by Swiss standards, produces decarbonised electricity at a competitive marginal cost. This resource, which built the canton's industrial fortune in the twentieth century — aluminium smelting, Upper Valais chemistry, and later pharmaceuticals — could play a role in the digital transition comparable to the one it played in the industrial transition. Hosting decarbonised intensive computing, partnerships with European operators seeking to decarbonise their infrastructure, links with Valais research institutions on computationally intensive AI applications: these are all avenues that deserve serious study, and that have not, until now, been studied at a scale commensurate with the opportunity.

This opportunity should not, however, be read naively. Valais hydroelectric power is not indefinitely expandable; most major sites are already developed, and the concessions on the large hydroelectric works will expire in the coming years under political conditions not entirely within cantonal control. Allocating a significant share of this resource to AI would carry an opportunity cost relative to other uses: decarbonising transport, heating, existing industry. The canton cannot do everything with the same water. It will have to make trade-offs.

Those trade-offs involve choices beyond the scope of this essay. But two principles strike me as defensible. First, that any AI infrastructure project welcomed into the canton be sized consistently with local generation and distribution capacity, and build in long-term climate constraints from the outset. Second, that trade-off decisions be made explicitly and publicly: that the canton state openly why, within the limits of the decarbonised kilowatt-hours available, it devotes a given share to AI and a given share to other uses. The transparency of the trade-off matters democratically as much as its substance. The bourgeoisies of Valais have long known that finite resources are shared out through rules that are clear, transparent and revisable. The grammar of the bisse — the traditional Valais irrigation channel that carries water across long distances to the fields — can serve here too.

More broadly, there remains the relationship between the digital transition and the climate transition. The two are not opposed by nature, but they come into tension wherever material resources — energy, rare earths, cooling water — overlap. Valais, caught up in both transitions at once, may have a chance to think through their relationship earlier and more seriously than territories less exposed to the issue. This is one of the intellectual undertakings the cantonal institutes of higher education and the administrations should take seriously. It does not lend itself to media announcements, but it will decide what is or isn't possible in the decade that follows.

Software sovereignty, often forgotten

One dimension of digital sovereignty is routinely underserved in public debate, even though it is more structurally important than infrastructure itself: software sovereignty.

Building a sovereign cloud infrastructure is useful. Filling it with proprietary American software that imposes its own formats, interfaces and terms of use is less so. Switzerland's parliamentary debate over the SGC explicitly recognised this issue: the chambers asked that the project favour open-source software and suppliers headquartered in Switzerland. That is an important signal. But the task ahead is immense, and reaches well beyond the SGC alone.

For a canton, software sovereignty is built through choices that are concrete and, on their face, modest. Favouring open formats within cantonal and municipal administrations. Investing in training staff on open-source alternatives to standard proprietary tools. Supporting Swiss software publishers who, without rivalling Microsoft or Google across the board, offer credible alternatives in specific segments. Documenting the cantonal administration's technical choices so that they remain reversible and auditable.

None of these decisions makes headlines. Compounded over ten years, they build genuine software sovereignty, one that makes the canton less dependent on players whose commercial trajectory it does not control. They also carry a useful side effect: they support the Swiss software publishing sector, which remains, in Valais as elsewhere, a field where local expertise is competitive but where public procurement does not always follow.

Data sovereignty, which changes the equation

Beyond infrastructure and software lies the question of the data itself. This, to my mind, is where sovereignty takes on its full meaning for a canton.

The data a territory produces — administrative, health, economic, environmental, cultural — constitutes a form of heritage with no material equivalent. Once it has left the territory, such data is hard to retrieve. Once folded into the training pipelines of global models, it serves to train tools whose use the canton no longer controls. Once hosted under foreign jurisdiction, it becomes accessible to authorities whose interests are not necessarily aligned with Switzerland's own.

This question is anything but symbolic: it is legal, economic and strategic. The nLPD sets out principles that Valais actors must follow, and those principles alone justify certain hosting choices. But beyond regulatory compliance lies the question of heritage: what does the canton want to preserve of the digital trace of its own activity, and how does it want to govern its use over ten, twenty, fifty years?

This question cannot be settled by a single grand policy. It is settled through a series of sector-by-sector decisions that, taken together, sketch a coherent stance. Medical data on Swiss infrastructure, as already discussed. Tax and administrative data on the SGC as it becomes available. Scientific data produced within the cantonal institutes of higher education on open, auditable platforms. Heritage data — sound archives, linguistic corpora, cultural inventories — on regional public infrastructure explicitly dedicated to their preservation. None of these decisions is revolutionary. Their sum, over ten years, marks the difference between a canton that controls its digital heritage and one that has let it slip away.

A programme sized to the canton

The realistic, defensible programme for Valais over the next ten years seems to me to hold together in four strands. It contains no grand flagship project, because, as we have seen, that is not the scale at which the essential work gets done.

Make full use of sovereign federal capabilities as they become available. Migrate cantonal and municipal administrative services to the SGC, take part in the federal work defining sovereignty standards, and defend within it the interests of a canton that is scattered, multilingual and alpine, whose needs are not exactly those of Bern, Zurich or Geneva.

Equip the regional private sector, particularly the intellectual professions, with sovereign infrastructure and tools suited to their realities. The private sector has no access to the SGC: other arrangements — partnerships with Swiss operators, sector-wide pooling, targeted public support — must accompany it in its own choices.

Build the specialised infrastructure that makes sense at regional scale, and only that. The canton can carry a handful of targeted projects, typically a hosting node dedicated to uses where Valais specificity runs strong — mountain healthcare, regional languages, digitised intangible heritage — conceived as complements to the SGC, not as its competitors.

Invest patiently in software sovereignty and data governance. This is the least visible dimension, but likely the most structurally important over the long run: favouring open formats in public procurement, documenting systems so they remain reversible, archiving data on auditable infrastructure, supporting Swiss publishers in specific segments.

None of these strands is thrilling. Taken together, they sketch a realistic, achievable digital sovereignty policy, sized to a canton that is neither Bern, nor Zurich, nor Geneva, but that has something to protect and something to build. Over ten years, it will have the practical effect one expects of a genuine policy: preserving the canton's room for manoeuvre in a world where digital choices often close off without anyone noticing.

A final word on realism

I will close this chapter with an observation some readers will find disappointing and others reassuring: there is not, as things stand, any formalised cantonal strategy for digital sovereignty in Valais. There are elements — sector-specific choices, good practices, competent operators, a genuine political awareness in certain departments. But there is no articulated doctrine, and there is no roadmap. This can be read as a delay; it can equally be read as an opportunity, since territories that formalise their stance earliest often draw a lasting advantage from doing so. Valais could write this doctrine and this roadmap within a year or two, if it decided to. The moment is probably a good one: early enough to engage usefully with ongoing federal projects, late enough to benefit from their lessons.

This doctrine will not replace concrete decisions. But it will give them a framework, a coherence and a direction. And that is precisely what is missing most, today, from a debate that stirs without advancing.