Homepage | Bulgaria AI Strategy

Introduction

Compute has become physical. For most of the internet era, digital infrastructure was treated as weightless — software scaling from a handful of technology capitals while the machinery behind it stayed invisible. Artificial intelligence has ended that. Frontier AI is a heavy industry that consumes electricity at the scale of national grids, needs land, water and resilient power, and is redrawing the economic map of Europe around wherever those things can be found. Physical things have locations, and the question of where Europe builds its AI has become one of the most consequential industrial decisions the Union will make this decade.

It matters because Europe does not yet control enough of its own. The models, chips and cloud platforms its companies, governments and universities depend on are built and governed elsewhere — and in 2026 that dependency stopped being an abstraction. A single foreign export-control order was able to cut off access to the most advanced AI models for every user outside one country, with no notice and no appeal. The lesson was not that Europe should decouple from American technology; that would be neither possible nor wise. It was that Europe needs enough of its own sovereign capacity to keep functioning if external access is ever restricted, degraded or repriced. After that, AI policy and physical infrastructure policy are the same policy — and Europe has, rightly, decided to build.

The AI Continent Action Plan, InvestAI, and the commitment to fund up to five AI gigafactories are that decision made concrete. The money is committed. The strategic rationale is accepted at the highest level of the Union. What remains is execution — and, first of all, the choice of where. This paper is about that choice, and it makes an argument that is meant to feel less like persuasion than like stating the obvious once someone finally looks at the map.

The logic of AI infrastructure runs away from Europe's crowded, expensive capitals and toward its energy-rich periphery, because these campuses follow power, not people. And within that periphery, one location assembles the full set of advantages the others can only offer in part: northern Bulgaria, with EU jurisdiction, operating nuclear baseload power, Danube cooling, the cheapest land and labour in the Union, and a region whose steep economic decline turns the cost advantage into a regeneration case at the same time. The demand to fill it already exists — much of it compute that Europe cannot legally run anywhere but inside its own borders. The supply is scarce. The best place to close that gap is not in serious doubt.

This paper sets out the case in full: why AI is reshaping Europe's economic geography, why the site should be on the Danube rather than in Sofia, why Bulgaria beats the competing national bids, how the surrounding economy should be built so the value stays in Bulgaria, who the customers are, what the site physically requires, and what could go wrong. It does not pretend the project is easy, and it does not hide the risks — they are given their own section. But its central contention is simple, and it is aimed squarely at two audiences. To the Bulgarian government: the case is yours to win and yours to lose, and losing it will take nothing more than moving too slowly. To the European Union: the strategy you have already written points to this place, this asset and this moment — now fund it, with your own members' money, and build.

Executive Summary

Europe has decided it needs to build sovereign AI infrastructure. It has committed the money — €20 billion for up to five AI gigafactories, inside the wider €200 billion InvestAI programme. It has accepted the reason: in 2026, a single foreign export-control decision disabled access to frontier AI for every user outside one country, overnight, proving that a continent which does not own its compute does not control its own AI policy. The only question left is where to build. This paper argues that the answer, for at least one of those gigafactories, is so clear it barely needs arguing — and that the real risk is not choosing wrongly, but moving too slowly to choose at all.

Northern Bulgaria has what AI infrastructure now competes for, in one place: operating, low-carbon nuclear power at Kozloduy; the Danube for cooling; the cheapest land and the cheapest labour in the European Union; and room to expand for decades. It has one thing more, which no rival site can match — it is the poorest region in the EU, which means the same project that delivers sovereign compute for Europe also delivers regeneration to the part of the Union that needs it most. This is not a data centre that happens to sit in a poor region. It is a rescue that happens to take the form of a data centre. EU cohesion funding and EU AI-sovereignty funding are each designed to reward exactly this — and almost no competing bid can claim both at once. Spain's leading site sits beside a nuclear plant facing closure, in a wealthy region. Vienna and Prague offer capital-city sites with no energy anchor and no regeneration case at all. On the merits, this is not close.

The demand side is just as settled. Europe already has compute it cannot legally or safely run on foreign infrastructure — government data, defence workloads, regulated industry, health and genomic research — and it does not have enough sovereign capacity to serve it. The buyer is not speculative. In large part it is the EU and its member states themselves, who have already decided they need this and are being asked to fund it. Supply is scarce; the demand is captive and growing. The two facts belong together.

So what is missing? Only will and speed. Comparator countries — Spain, Romania, Austria, Czechia — already have delivery vehicles, committed capital and formal submissions in motion. Bulgaria, as of the latest public information, has an exploratory conversation. The Commission's formal gigafactory call is expected within weeks. That gap is not a problem of case; it is a problem of urgency, and it is fixable now. This paper calls for two things. First, the Bulgarian government must stop treating this as an aspiration and start treating it as a national priority: one decision on the site, one accountable delivery vehicle, one leader, taken in weeks rather than years. Second, the European Union must put its own members' money where its stated strategy already is — funding this as a public act of sovereignty rather than waiting for private capital or foreign hyperscalers to decide Europe's strategic autonomy for it.

There are risks. The nuclear expansion carries execution risk; the local skills base is thin; heavy electrical equipment has long global lead times. They are set out honestly in this paper, in a dedicated register. But not one of them is disqualifying, and all of them are smaller than the prize — and smaller, too, than the cost of hesitating while the window is open. The case is strong enough to win. The only way to lose it is to be too slow. This paper is an argument for moving.

Why AI Infrastructure Is Changing the Economic Geography of Europe

For most of the internet era, digital infrastructure was treated as almost weightless. Software companies could scale globally from a small number of major technology centres, while the physical infrastructure behind cloud services remained largely invisible to governments, investors and the public. Artificial intelligence changes that.

Europe’s sovereignty wake-up call

Until recently, Europe’s AI strategy rested on an unstated assumption: that frontier AI models, chips and cloud capacity built outside the EU would simply remain available, indefinitely, on commercial terms. That assumption was tested directly in June 2026.

On 12 June 2026, the United States government issued an export control directive instructing the AI company Anthropic to suspend all access to its most advanced models, Fable 5 and Mythos 5, by any foreign national, whether located inside or outside the United States, and including Anthropic’s own foreign national employees. Because the directive could not practically be segmented by nationality, Anthropic disabled both models for every customer globally, days after Fable 5 had been released. The government cited national security concerns related to a method of bypassing the model’s safeguards; Anthropic stated it had received limited written detail of the underlying concern and that other publicly available models appeared susceptible to a similar technique. Access to Anthropic’s other models was unaffected, but the episode marked a clear escalation: the United States had, for the first time, used export control authority to restrict access to an AI model itself, rather than only the chips used to run it.

The controls were lifted eighteen days later. The Department of Commerce withdrew them on 30 June, and Fable 5 was restored to users worldwide the following day, following negotiations between the company and the US government to which no European party was privy; access to Mythos 5 remains limited to a small number of vetted US organisations. That the interruption proved temporary does not diminish what it demonstrated. For eighteen days, every European user, government and business had no notice, no appeal and no standing, and access was returned not because Europe had any leverage but because a foreign administration revised its own position. A capability that can be withdrawn and restored at another state's discretion is not one Europe controls.

The reaction across Europe was immediate, and instructive. The UK’s AI minister publicly stated that the episode underlined how easily a country could be cut off from the world’s most advanced AI capability, and argued it should prompt deeper investment in domestic AI capacity. Whatever view one takes of the underlying dispute, the structural lesson for Europe is unambiguous: if a single national security directive can disable an entire tier of frontier AI capability for every foreign user on a Friday afternoon, with no advance notice and no mechanism for an unaffected ally to be carved out, then AI policy and physical AI infrastructure policy are the same policy. A continent that does not control meaningful compute capacity inside its own jurisdiction does not control its own AI policy, rather it operates AI policy at the discretion of whichever government controls the infrastructure it depends on.

There is a second, less obvious implication that matters specifically for this thesis. For the purposes of AI sovereignty, the European Union behaves less like 27 separate countries and more like a single jurisdiction. It operates a single market, a shared regulatory framework for AI and data protection, free movement of capital and labour, and for most members a shared currency. The protections, legal standing and policy leverage that AI sovereignty actually requires exist at the EU level, not solely at the level of any one member state. That has a direct consequence for where sovereign compute should be built: the right question is not which national economy most needs this investment for its own sake, but where, inside the Union as a whole, sovereign compute can be built fastest, cheapest and most durably, for the benefit of the entire bloc. Europe does not need all 27 member states to each build their own sovereign AI capacity. It needs enough capacity to exist, securely, somewhere inside its borders and the member state best placed to host it need not be the member state with the largest economy.

This is precisely why Europe is now building, not just regulating, and it is the lens through which the rest of this section, and this thesis, should be read.

AI is turning compute into physical infrastructure

AI is not only a software layer. It is a physical infrastructure industry. Large-scale AI models require enormous quantities of electricity, advanced chips, high-density data centres, cooling systems, fibre connectivity, land, skilled operators and resilient grids. The result is that the economics of AI are increasingly shaped not just by software talent and capital, but by the availability of power and physical infrastructure.

The International Energy Agency estimates that global electricity consumption from data centres is projected to double to around 945 TWh by 2030, with demand growing by roughly 15% per year between 2024 and 2030. This is more than four times faster than the growth rate expected from other electricity-consuming sectors. The IEA also reported that electricity demand from data centres increased by 17% in 2025, with AI-focused data centres growing even faster.

This matters because AI infrastructure is no longer a marginal extension of the digital economy. It is becoming a major source of electricity demand and a meaningful component of national industrial strategy.

Traditional data centres followed cities; AI campuses follow power

Historically, data centres were often located close to major cities, financial centres, telecoms hubs and enterprise customers. This made sense because many workloads were latency-sensitive. Banks, exchanges, consumer internet platforms and enterprise cloud customers wanted infrastructure close to users and commercial activity.

AI changes that logic. While some AI inference workloads will still need to be located close to users, large-scale training workloads are less dependent on immediate proximity to major urban centres. They are more dependent on the ability to secure very large blocks of reliable electricity at predictable cost.

This creates a new geography. Countries and regions that were previously peripheral to Europe’s digital economy may become strategically important if they can offer:

large-scale low-carbon electricity;
available grid capacity;
land for hyperscale campuses;
cooling water or favourable cooling conditions;
fibre connectivity into major European markets;
political and regulatory stability.

In other words, AI infrastructure shifts part of the competitive advantage away from expensive metropolitan centres and towards energy-rich regions. This is the central reason why northern Bulgaria deserves serious consideration. It may not have the technology ecosystem of London, Paris, Amsterdam, Frankfurt or Dublin, but it has the potential to offer something increasingly scarce: scalable energy-linked infrastructure inside the European Union.

Europe is trying to build sovereign AI capacity

The opportunity must also be understood in the context of European sovereignty. Europe does not want to be permanently dependent on US and Chinese technology platforms for AI infrastructure, models, chips and cloud capacity. This is why the European Commission has launched the AI Continent Action Plan, including the InvestAI initiative, which is intended to mobilise €200 billion for AI investment in Europe. The plan includes €20 billion for up to five AI gigafactories and support for AI factories serving startups, industry and research.

The EU is therefore not only regulating AI; it is trying to build the industrial base that supports AI. That creates an opening for countries that can credibly offer sites for large-scale AI infrastructure. Bulgaria should position itself within this wider European strategy, not as a low-cost outsourcing destination, but as a strategic host for sovereign European compute.

This is particularly important because the EU’s AI ambitions require more than research programmes and software companies. They require physical capacity: power, data centres, chips, networks and operators. If Europe wants AI sovereignty, it needs infrastructure sovereignty.

Low-carbon baseload power is becoming a strategic asset

AI infrastructure creates a tension between digital growth and energy policy. On one hand, Europe wants more AI infrastructure. On the other hand, Europe is trying to decarbonise its electricity system, manage grid constraints and avoid politically unpopular increases in energy costs.

This makes low-carbon baseload electricity particularly valuable. Intermittent renewables are essential, but hyperscale AI campuses require stable, high-volume electricity supply. Nuclear power, hydro power, grid-scale storage and flexible gas capacity therefore become strategically important in supporting AI growth.

The European Commission has itself highlighted the energy challenge of data centres, noting that global data centre electricity consumption was around 415 TWh and is projected to more than double towards 945 TWh by 2030, mainly due to energy-intensive accelerated computing used for AI. Recent EU proposals also point towards greater scrutiny of data centre energy efficiency, water consumption and clean energy use.

This is where Bulgaria’s position becomes interesting. Bulgaria already has operating nuclear capacity at Kozloduy and is pursuing further nuclear expansion. If that expansion proceeds, Bulgaria could offer a rare combination: EU jurisdiction, nuclear baseload power, Danube cooling water and low-cost land.

Smaller countries with surplus power can become compute exporters

The most important strategic point is that AI creates a new export model. Historically, countries exported energy-intensive goods such as aluminium, fertiliser, steel or chemicals. In the AI economy, countries with surplus power may be able to export compute instead. That does not necessarily mean selling electricity across a border. It means converting domestic electricity into digital services used by companies, researchers and governments elsewhere.

For Bulgaria, this is a powerful idea. The country does not need to become Europe’s largest software market to become strategically relevant. It needs to become one of Europe’s most attractive locations for power-intensive AI infrastructure. If successful, Bulgaria could export compute to the rest of Europe while using the associated infrastructure to attract companies, talent, research and industrial activity into northern Bulgaria.

However, this is also where the strategy must be careful. If Bulgaria simply provides cheap power and land to foreign hyperscalers, most of the economic value will be captured elsewhere. The value chain would look like this:

Energy → data centre → foreign AI platform → foreign customers

That model risks leaving Bulgaria with construction jobs, grid strain, modest tax receipts and limited local economic transformation. The better model is:

Energy → AI infrastructure → industrial cluster → companies, talent, research and capital

This is the key distinction. The objective should not be merely to host data centres. It should be to use AI infrastructure as the anchor asset for a new economic region.

Stategic Use Cases

Europe needs secure access to advanced AI capability and that access cannot depend entirely on foreign-controlled cloud platforms, foreign foundation models, foreign chips and foreign policy decisions. The issue is not that Europe should cut itself off from American technology. That would be unrealistic and economically damaging. The issue is that Europe needs enough independent capacity to remain operational if access to external systems is restricted, degraded, delayed or repriced.

Recent events have made this point less theoretical. In 2025, the United States paused intelligence sharing with Ukraine, including categories of satellite and targeting-related support that Ukraine had relied on heavily. Europe could provide some support, but not at the same level or scale. In 2026, European policymakers and companies were also openly discussing the risk that dependence on US-controlled AI models, cloud infrastructure and chips could leave Europe vulnerable to policy shifts outside its control. Furthermore, the Trump administration issued an export-control directive requiring Anthropic to suspend access to its most advanced models, Fable 5 and Mythos 5, by “any foreign national”, whether inside or outside the US.

The lesson is that dependency is a strategic weakness. If a capability is essential to defence, cyber security, industrial productivity, scientific research and public administration, then Europe needs the ability to operate it, govern it and sustain it under its own control.

Sovereign AI compute

Europe is already trying to increase its AI infrastructure base through the AI Continent Action Plan, InvestAI, AI factories and proposed AI gigafactories. The Commission’s stated objective includes large-scale AI computing infrastructure, up to five AI gigafactories and a material increase in European data centre capacity.

The Bulgarian corridor fits this policy direction. It would provide a lower-cost, energy-linked location for European AI compute inside the EU. The facility could support European companies, universities, governments and public-sector bodies that require access to AI infrastructure governed by EU law and operated within a European strategic framework.

This does not mean every workload needs to be sovereign. Many commercial workloads can continue to use global platforms. The relevant point is that Europe needs sovereign fallback and sovereign capacity for critical workloads. That includes defence, cyber security, public administration, critical infrastructure, healthcare research, industrial systems and sensitive datasets.

Defence and dual-use capability

Modern defence increasingly depends on data processing, satellite imagery, autonomous systems, signals analysis, logistics optimisation, cyber operations, drone coordination, electronic warfare and simulation. These are compute-intensive activities, developed further from a recruitment and land-use perspective. They also require secure infrastructure and trusted supply chains.

Europe’s experience with Ukraine has shown that military effectiveness can depend on external digital infrastructure, including intelligence flows, satellite connectivity, cloud systems and targeting data. If those systems are controlled outside Europe, then European security policy is partly dependent on decisions made elsewhere.

A Bulgarian AI corridor would not solve Europe’s defence dependency by itself. It would, however, contribute to a European defence compute layer. This is also relevant to the Black Sea region: Bulgaria sits on the EU’s south-eastern frontier, and a serious AI infrastructure project in this region would have strategic value beyond commercial cloud capacity.

Cyber secutiry

Cyber security should be treated as one of the corridor’s main commercial and strategic use cases. AI is already changing cyber security in both directions. It improves detection, incident response, code analysis and anomaly monitoring, but it also increases the capability of attackers. European governments, banks, utilities, telecoms operators, transport networks and defence contractors will need access to secure AI systems that can analyse large volumes of sensitive operational data without exporting that data into uncontrolled environments.

The corridor could support a cyber-security cluster anchored in Sofia and Ruse, with the main compute infrastructure in northern Bulgaria. The relevant customers would include public authorities, banks, insurers, energy networks, telecoms companies, defence suppliers and critical infrastructure operators.

Economic development and EU cohesion

The investment case will also support EU cohesion. This is important because public money should do more than solve one problem, and a Bulgarian AI corridor would address several problems at once: sovereign AI capacity, defence resilience, cyber security, low-carbon infrastructure, regional development and demographic decline.

Bulgaria is not starting from the same position as France, Germany or the Netherlands. It remains at the bottom of the EU GDP-per-capita range on a purchasing-power basis. This matters for investment efficiency. A euro of strategic infrastructure investment in northern Bulgaria is likely to have a larger regional development effect than the same euro spent in an already-saturated Western European technology centre. This is also where specific EU funding instruments beyond general cohesion language become relevant: the European Regional Development Fund, the Just Transition Fund (given Bulgaria’s coal-region transition profile), and the Recovery and Resilience Facility all have application criteria this project could credibly meet, and a formal submission should name the specific instrument being applied to rather than referring to “cohesion” generically.

The corridor should therefore be framed as dual-use in the broadest sense: it supports European strategic autonomy, and it supports economic convergence inside the EU.

The investments case

The investment case for the corridor is that it creates multiple layers of value from one infrastructure platform.

The first layer is infrastructure value: land, grid connection, data halls, substations, cooling, fibre and energy contracts.
The second layer is strategic value: sovereign AI compute, defence capability, cyber resilience and public-sector capacity.
The third layer is economic-development value: employment, suppliers, housing, technical training, university partnerships, land uplift and regional regeneration.
The fourth layer is industrial value: companies using the corridor to build AI products and services in energy, mining, defence, logistics, agriculture, manufacturing and critical infrastructure.
The fifth layer is geopolitical value: a stronger EU digital presence in the Black Sea region.

This is why the project is more attractive than a conventional data centre investment. A normal data centre investment mainly produces infrastructure returns. The Bulgarian AI corridor can produce infrastructure returns and strategic returns at the same time.

Why Bulgaria and Why This Region Should Be the EU’s Choice

The European Union does not lack candidates for AI gigafactory infrastructure. By the time the Commission’s informal call for expressions of interest closed in June 2025, it had received 76 proposals across 60 sites in 16 member states, implying more than €230 billion in proposed investment. Spain, Romania, Austria, Czechia, Germany, France, the Netherlands and others are all advancing credible, well-funded bids. Bulgaria’s case cannot simply rest on the fact that it has nuclear power, land and a river. It has to explain why, among a genuinely crowded field, the EU’s money and strategic attention should land specifically in northern Bulgaria.

This section makes that case directly. It also states plainly where Bulgaria is currently behind, because a case that only states strengths is not a serious one.

The case has to start with the human stakes, not the infrastructure

Most countries competing for gigafactory status are arguing for an economic upgrade. Bulgaria is arguing for something closer to a rescue.

Bulgaria’s population peaked at roughly 9 million in 1988, just before the end of the communist period. It now stands at approximately 6.4–6.5 million — a cumulative loss exceeding 2.5 million people, or close to 28%, in under four decades. No other EU member state has lost a comparable share of its population in peacetime in the modern era. Multiple independent sources — Bulgaria’s own Center for Demographic Policy, international demographic rankings and financial press coverage describing Bulgaria as the EU’s poorest member state — converge on the same conclusion: Bulgaria’s population decline is among the fastest in the world, not just in Europe. Left on its current trajectory, UN population projections point toward roughly 5.4 million people by 2050, and as low as half of today’s population by the end of the century.

This is not evenly distributed. It is concentrated precisely where this thesis proposes to build. Vidin Province — part of Severozapaden, and one of the provinces on the proposed corridor — has lost more than half its population since 1989, falling from approximately 170,000 to around 70,000 residents, making it one of the most depopulated administrative units anywhere in Europe. This sits alongside the figures already established elsewhere in this thesis: Severozapaden and Severen Tsentralen ranked among the EU’s five poorest regions in the European Commission’s own 2026 Bulgaria Country Report, and Bulgaria lost 19.1% of its working-age population between 2011 and 2021.

The point is not that infrastructure alone reverses demographic collapse, but that a gigafactory built in a wealthy or stable region is an economic add-on. A gigafactory built in Severozapaden is a test of whether Bulgaria can interrupt a 35-year decline curve in the one part of the EU where that curve is steepest. That is a materially different and more urgent proposition than any rival bid is making.

Where Bulgaria sits against the live competition

The most recent and current public information on rival bids is summarised below.

Four things stand out from this comparison.

No other contender combines an expanding nuclear baseload with a host region that is, by the EU’s own statistical measure, its poorest. Spain’s nearest equivalent site sits next to a nuclear plant that may be retired, not built up — the opposite energy trajectory to Kozloduy. Romania has the nuclear-expansion story but not the extreme-poverty claim. Vienna and Prague have neither.
Several rivals are already using regeneration language — Móra la Nova’s own mayor has explicitly framed the Spanish bid as an answer to rural depopulation — so “regional regeneration” alone will not differentiate Bulgaria. What differentiates it is the degree: Severozapaden’s GDP per capita, at roughly 42% of the EU average, is materially lower than that of Ribera d’Ebre’s wider region, Catalonia, which is itself classified as a “more developed” EU region overall. The Bulgarian case is not a milder version of the Spanish one; it sits in a different EU funding tier altogether.
On the specific point this thesis wants to emphasise — room to expand — northern Bulgaria compares well. Spain’s flagship site is reported at roughly 4 hectares: workable for an initial phase, but tightly bounded. The Kozloduy–Mizia–Oryahovo corridor, by contrast, was deliberately scoped earlier in this thesis as a land-rich triangle capable of phased growth from a 100–150MW first phase toward 500MW-plus, plus adjoining land for substations, storage, heat-reuse industries and future tenants. A campus that can credibly promise multi-phase growth on contiguous land is a different proposition from a single bounded plot, and the EU’s own gigafactory criteria explicitly reward long-term scalability, not just first-phase capacity.
On top of all of this, Bulgaria offers a cost base that no rival in this list can match. Eurostat’s 2025 figures put Bulgaria’s average hourly labour cost at €12.0 — the lowest of any EU member state, against an EU average of €34.9. Bulgaria also has some of the lowest agricultural land prices anywhere in the EU, with permanent grassland the cheapest in the Union outright and other land categories consistently among the three or four cheapest nationally. Decades of depopulation have left northern Bulgaria with exactly the kind of large, undeveloped, unencumbered land this project needs — not as a consolation for the demographic decline described earlier, but as a direct practical consequence of it: land that emptied out is land that is available now, at minimal cost, without the planning conflicts or land-assembly delays that constrain denser, more expensive sites elsewhere.

This matters strategically, not just financially. At a moment when European sovereignty over compute is being treated as urgent rather than aspirational, the binding constraint on most EU gigafactory bids will not be vision — it will be speed: how fast a site can be assembled, financed and broken ground on. A low cost base lowers the capital required to move first and reduces the execution risk that comes with waiting for slower, more expensive sites to clear planning and financing hurdles. Bulgaria’s advantage here is not that it is the most ambitious bid on the table. It is that it can plausibly be the fastest and cheapest to get moving — at precisely the moment the EU has signalled that speed, not just ambition, is what sovereignty now requires.

What this means for EU funding logic specifically

Under EU cohesion policy, regions below 75% of EU average GDP per capita are classified as “less developed regions” and qualify for the most favourable co-financing terms available under the EU budget. Severozapaden, at roughly 42% of the EU average, is not marginally inside that category — it sits at the extreme end of it, among the very lowest-ranked regions across the entire Union.

This matters financially, not just rhetorically. A euro of InvestAI or EU gigafactory capital deployed in Severozapaden can, in principle, be matched against cohesion-policy co-financing rates that none of the rival sites discussed above can draw on in the same way — not Catalonia, not Vienna, not Prague, and not most of the named German, French or Dutch locations. Bulgaria’s pitch to the Commission should make this explicit: this is not only the most strategically logical site on energy and land grounds, it is also the site where EU capital, euro for euro, is structurally positioned to do the most good under the Union’s own funding rules.

Limitations that have to be acknowledged

Bulgaria is currently behind procedurally. Spain has a funded, structured SPV with €719 million already approved and named private partners. Romania has a government mandate split across two ministries and a March 2026 submission target. Austria and Czechia have signed political applications. Bulgaria’s position, by contrast, is a single exploratory conversation between the Prime Minister and IBM from October 2025 — no consortium, no committed capital, no confirmed expression of interest in the original 76. Bulgaria’s one confirmed EU compute win to date, the BRAIN++ AI Factory, is real but smaller in scale than a gigafactory and is sited in Sofia, not the Kozloduy corridor, so it does not yet anchor the northern Bulgaria case at all.

The timing is now urgent rather than theoretical. The European Commission’s formal call for AI Gigafactory proposals is live. However strong the underlying energy, land and demographic case for northern Bulgaria is, it will not translate into a competitive submission without a delivery vehicle, a capital commitment and a single accountable leadership structure assembled on a similar timeline to Spain’s.

The nuclear-expansion upside also still carries genuine execution risk. Units 7 and 8 are not yet financed, built or connected, and the thesis should continue to rest its base case on what already exists — roughly 2GW of operating capacity — treating the expansion as upside rather than a precondition. Equally, the thinness of the local labour and supply ecosystem around Kozloduy and Oryahovo is real, which is precisely why the corridor model built around Pleven and Ruse.

The synthesis: why the answer should still be Bulgaria

Strip away the comparison table and one conclusion holds — every other serious contender in the current EU gigafactory process is, at its core, an infrastructure investment that happens to sit in a particular place. The Bulgarian bid, if assembled with the urgency this moment requires, is something different: a demographic and regional-survival project that happens to take the form of AI infrastructure.

No rival combines an expanding, rather than retiring, nuclear baseload; a host region that is, by Eurostat’s own measure, the poorest in the European Union; land with genuine multi-phase expansion capacity rather than a single bounded plot; and a national demographic emergency that the project is explicitly designed to help reverse. That combination is precisely the kind of dual-purpose case that EU cohesion policy and EU AI-sovereignty funding are designed, separately, to reward, and which almost no other bid on the table can claim to satisfy simultaneously.

Both the strategic and funding logic favours Bulgaria. The case that has not yet been made is the procedural one — and that is now the most urgent task in front of this thesis.

Why Northern Bulgaria, Not Sofia

Where this project sits is almost as consequential as whether it happens at all. A campus built around Sofia would be commercially the path of least resistance, and a campus built in the north carries real execution risk that a capital-city site would not. The case for northern Bulgaria has to be made on its own terms, not assumed.

The pull of the capital, and why this project should resist it

Sofia is Bulgaria's natural centre of gravity, and for most kinds of technology investment that would settle the question. It has the country's deepest talent pool, its government and regulatory institutions, its universities and its capital markets. A data centre built there would be the easiest version of this project to finance, staff and explain to investors in the short term.

That ease is exactly the problem. Sofia already captures a disproportionate share of Bulgaria's economic activity, and a Sofia-anchored AI strategy would not correct that imbalance — it would deepen it. Land and labour are already more expensive there than almost anywhere else in the country, which means a Sofia campus would compete with the rest of the economy for the same scarce inputs rather than putting underused ones to work. It would also sever the project from the one asset that gives it real strategic weight: proximity to Kozloduy. A data centre an hour from Sofia is a private infrastructure investment. A data centre on the Danube, beside the country's nuclear baseload, in the part of Bulgaria that most needs it, is a national one.

None of this means Sofia is irrelevant. A large AI training campus does not need to be surrounded by thousands of software engineers — those engineers can sit in Sofia and connect to the corridor remotely, much as Bulgaria's wider technology sector already operates today. Sofia's role should be the headquarters, finance, policy and institutional interface for the project. It should not be where the project physically lives.

What makes northern Bulgaria the right answer, not just the cheap one

It would be a mistake to rest the case for northern Bulgaria on land prices alone. Cheap land is not, by itself, a strategy — it is usually cheap for a reason, and an investment thesis built only on cost says nothing about why the asset will appreciate. The stronger argument is that several genuinely strategic assets happen to sit in the same underused geography: Kozloduy's nuclear baseload, direct Danube access, low congestion, and an existing, if faded, industrial base that predates the region's decline.

That decline is itself part of the case, not a separate issue from it. Severozapaden and Severen tsentralen rank among the poorest regions in the European Union — a fact explored in full in the sections that follow — and the land available in the corridor today is abundant and inexpensive precisely because people, industry and investment have been leaving this part of the country for three decades. Framed this way, the cost advantage and the regeneration argument are not two separate selling points. They are the same fact, read from two directions. A region that emptied out is a region with room to grow into, at minimal cost and with minimal land-assembly friction — and a data centre sited near Sofia could never make that claim, however cheap its electricity.

This is also what separates a technology project from a national one. A campus beside Sofia is an investment. A campus on the Danube, anchored to nuclear power, in the poorest part of the country, is a deliberate act of regional rebalancing — and that distinction matters as much to the European Commission evaluating this bid as it does to Bulgaria itself.

Selecting the corridor

Site selection should follow a clear hierarchy, in this order: power access, grid connection, water and cooling, land availability, fibre connectivity, the ability to stimulate local economic renewal, transport and labour access and expansion potential. On that basis, the strongest candidates are not Bulgaria's largest towns. They are the places that combine power, land, water and a credible regeneration story in the same location. Six corridors along the Danube were assessed against that hierarchy.

Kozloduy–Mizia–Oryahovo is the clear first choice, and not only because it sits closest to the power source. It is the only candidate on this list that can credibly tell all four parts of the story this thesis depends on at once: nuclear-anchored power, Danube cooling, a serious regeneration case, and room to expand far beyond a first phase. Every other corridor trades one of those four for something else — Ruse trades the nuclear story for connectivity, Vidin trades it for transport infrastructure, Pleven trades proximity to the plant for an easier labour market. None of those trades are fatal, which is exactly why each retains a defined supporting role rather than being discarded. But none of them is a substitute for the core site.

The honest weakness of Kozloduy–Mizia–Oryahovo is that the surrounding ecosystem is thin. Kozloduy and Oryahovo are not deep labour markets, and the project cannot be built as a standalone facility dropped into the landscape — it has to be designed from the outset as a campus plus a regeneration zone, with housing, training, schools and technical colleges built alongside it. That is not a reason to choose a different site. It is the reason this site matters: if the objective is to change northern Bulgaria's economic trajectory, the project needs to be large enough to build its own ecosystem rather than borrow one that already exists elsewhere.

Pleven and Ruse matter because the main campus cannot succeed in isolation. Pleven supplies the regional labour and supplier base that Kozloduy itself lacks; Ruse supplies connectivity and logistics depth that the corridor's western end does not yet have. Neither competes with the main site. Both make it workable.

The political and demographic case, and what follows from it

Bulgaria's demographic decline is not spread evenly across the country and that matters for where this project sits, not only why it exists. It is sharpest in precisely the region this thesis proposes to build in. The point to establish here is narrower: anchoring the project in Severozapaden targets the part of Bulgaria where the population and skills drain is most severe, rather than adding further investment to a capital that does not face the same emergency.

That is also what gives a Kozloduy-centred strategy a political logic a Sofia-centred one could never have. It aligns energy strategy, regional development, EU cohesion policy, nuclear expansion and demographic renewal in a single project, in the single part of the country where all five problems are most acute. A data centre beside Sofia is, at best, a successful private investment. A corridor anchored on the Danube, in the poorest region in the European Union, is a national project that happens to take the form of a data centre — and that is the case Bulgaria should be making to Brussels, not the inverse.

Keeping the Value in Bulgaria

The easiest version of this project is also the worst one for Bulgaria: lease the land, sell the power, and let a foreign hyperscaler capture everything above the utility margin. That outcome is the default, not a risk at the edges — it is what happens automatically unless Bulgaria designs against it. This section is about designing against it, and the argument turns on a single point of timing that most host countries miss.

The financing model creates one window — and only one

The instinct is to treat value capture as something to negotiate later: attract the investment first, extract local benefit afterwards through tax and goodwill. That instinct fails, because leverage disappears the moment the capital structure closes.

The EU's gigafactory financing model makes this concrete. The Union's direct contribution is capped at a minority of project cost; the majority is expected from private investors through an offtake structure, where a partner effectively pre-purchases compute capacity in exchange for funding the build. That structure is Bulgaria's opening. Because the capital stack has to be assembled through negotiation rather than simply granted, the terms that determine who captures the value — equity, local content, domestic access — are on the table at the same moment the money is. Written in then, they cost nothing. Sought afterwards, they read as Bulgaria trying to claw back margin from a partner who has already priced the deal without them. A host that understands this negotiates once, at the start. A host that doesn't spends a decade as a landlord.

What to write into the deal

Three mechanisms matter, and all belong in the founding agreement rather than a later annex.

The first is a direct equity or revenue-sharing position for the Bulgarian state — a stake in the project company, not merely lease income and corporation tax. Spain's Móra la Nova bid is structured this way, with the state holding a direct position rather than standing outside the vehicle it helped fund.

The second is a local-content floor: a defined share of construction, maintenance, security, facilities and energy work reserved for Bulgarian firms, because this, not the data halls themselves, is where durable domestic employment actually sits. The permanent headcount of the campus is small; the supplier ecosystem around it is not, but only if it is contracted into existence rather than hoped for.

The third is a domestic compute set-aside: a guaranteed slice of capacity reserved for Bulgarian universities, research institutions and public bodies, so the country is not in the position of hosting Europe's compute while its own institutions queue for it elsewhere.

None of these are exotic. What is unusual is insisting on all three before financial close, when they are nearly free, rather than after, when they are nearly impossible.

Bulgaria's own industries are the recruitment edge

The country already has industries whose economics change when large-scale compute sits next door and most rival sites do not. Bulgaria has a real mining and mineral-processing sector, an operating power system mid-transition and a substantial agricultural base in the north itself. These are precisely the sectors where cheap adjacent compute is not a luxury but a margin lever — geological modelling and resource optimisation for mining, grid forecasting and storage dispatch for energy, yield and precision systems for agriculture.

A recruitment pitch built on these lands differently from a generic one, because it offers incoming operators an existing local customer and supplier base, not a greenfield. This is the part of the industrial story Bulgaria can tell that Vienna or Prague cannot, and it should be led with when courting tenants. The domestic base is the differentiator, not the compute itself.

Waste heat: a physical differentiator, handled honestly

A hyperscale campus rejects enormous quantities of low-grade heat and Northern Europe has already turned that by-product into infrastructure:

Meta's Odense campus feeds Denmark's largest data-centre heat-pump installation, recovering around 215,000 MWh a year and heating more than 12,000 homes;
Stockholm's Data Parks programme warms roughly 30,000 apartments;
Google's Hamina site supplies about 80% of local heating demand; and
Dublin's Tallaght scheme, fed by an Amazon facility, cut an estimated 1,100 tonnes of CO₂ in its first year.

The honest complication is that these all sit beside dense urban heating networks and the Kozloduy corridor does not — for the same reason its land is cheap. Industry analysis is clear that heat reuse is hardest to justify where no heat network exists within a few kilometres. That dictates sequencing, not abandonment. In early phases the heat should go to uses that need no urban network and can be built on adjacent land from the outset — greenhouses, aquaculture, food drying and processing, light industry designed around the heat source. District heating into homes is a credible phase-two ambition for a denser node such as Pleven as the corridor grows, not a first-phase promise for a site whose surrounding population does not yet exist. Stating that openly is more persuasive than claiming otherwise, and it is the kind of honesty a Commission evaluator will notice by its absence in rival bids.

The wrapper: an AI Special Economic Zone aligned to the EU's own instrument

Individually, the mechanisms above are a series of side-deals. Bound together into a dedicated legal and administrative zone, they become something a state can legislate once and an investor can evaluate at a glance. More importantly, they line up with a policy instrument the EU is already building.

The Commission's Cloud and AI Development Act process is actively considering "Special Compute Zones": pre-cleared sites for AI buildout with single-window permitting, on-site generation rights and harmonised security tiers, modelled explicitly on the Chips Act "green lane" that unlocked investment such as TSMC's Dresden fab The problem they address is Bulgaria's problem — EU data-centre permitting routinely runs past four years against build timelines measured in months. A zone designed for AI from the outset closes that gap, and a Bulgarian proposal framed as an early, working instance of the EU's own emerging instrument reads very differently from a one-off national request.

There is also a clean legal path, not an exception to be argued for. Member states may already operate special economic zones with tax and procedural incentives, and the Commission treats this as legitimate regional aid precisely where it targets regions with unusually low living standards or serious job shortages. Severozapaden is not a marginal case for that rule — it is close to the textbook one, closer than almost any region in the Union. Poland's long-running SEZ programme is the direct precedent for how this operates in practice, tax treatment and compliance obligations included.

A Kozloduy–Mizia–Oryahovo AI Special Economic Zone should therefore be an explicit ask: single-window permitting run jointly with ESO rather than sequentially through separate authorities; pre-cleared land banks so tenants are not each negotiating access from scratch; the equity, local-content and set-aside terms written into the zone's standing rules rather than deal-by-deal; and formal alignment with the EU framework as it emerges. That is the difference between arriving at the Commission with a request for money and arriving with a finished answer.

Governance, Delivery Vehicle and Execution Leadership

Every section of this thesis so far has argued that Bulgaria’s underlying case — energy, land, demographic urgency, cost — is strong. None of that matters if the project cannot actually be delivered on the timeline the EU’s own gigafactory process now demands. This section addresses delivery directly: not whether Bulgaria should pursue this, but how it should be run so that it actually gets built quickly.

The problem with the conventional approach

The default model for a project like this is a long competitive process: publish a call, receive bids, evaluate proposals, negotiate with multiple winning consortia, then manage a sprawling set of contractors, financiers and operators for years afterward. Spain’s Móra la Nova bid illustrates the more conventional version of this — a public-private SPV with several large named partners (Telefónica, Santander, ACS) negotiating jointly. That model works, but it is built for a world where speed is not the binding constraint.

Speed is the binding constraint here. The European Commission’s own formal gigafactory call is expected to open in mid-July 2026, and the broader EU AI Continent Action Plan exists precisely because Europe has concluded that the slow, fragmented way infrastructure has historically been built in the EU — extended permitting, multiple overlapping authorities, drawn-out procurement — is no longer fast enough for what AI sovereignty requires. A Bulgarian bid that proposes to fix that problem with more of the same — a long bidding process, a wide consortium of counterparties each with their own veto points — would be solving speed with the exact mechanism that causes the lack of speed in the first place.

The recommended model: one decision, one leader, one mandate

The right model is closer to a ship than a committee. A ship has many crew, many specialists and many functions, but it has one captain, and that captain’s decisions are not subject to renegotiation by every department head every time a course correction is needed. This project needs the same structure. Concretely, that means:

One decision, taken once. Rather than running a multi-year competitive process to choose between candidate sites and operators, the Bulgarian state should take a single, formal decision — at Council of Ministers level — that the Kozloduy–Mizia–Oryahovo corridor is the national site for sovereign AI infrastructure. This removes years of internal uncertainty and signals to the European Commission and to private capital that Bulgaria has already done its own site-selection work, rather than asking the EU to wait for it.
One delivery vehicle, with real authority. Bulgaria already has a working precedent for this kind of structure: Kozloduy NPP – New Builds PLC is a dedicated, single-purpose project company set up specifically to deliver Units 7 and 8, rather than running the expansion through the existing utility’s general management structure. The same model should be used here — a single, purpose-built national delivery vehicle for the AI corridor, with statutory authority over land assembly, permitting coordination and grid-connection sequencing, so that a tenant or investor deals with one counterparty for those functions rather than several ministries and agencies in sequence.
One accountable leader. The delivery vehicle should have a single named programme director, reporting directly to the Council of Ministers or the Prime Minister’s office, who is personally accountable for the delivery timeline. Diffuse committee leadership is the most common reason large infrastructure projects lose months to indecision; a named leader with a clear mandate and a clear reporting line removes that failure mode by design.
Minimise the number of counterparties Bulgaria has to manage directly. The EU’s own gigafactory financing model already requires a private capital partner to fund the majority of project cost through an offtake structure. Bulgaria cannot avoid having private partners. It can avoid having to manage dozens of them. The state’s commercial strategy should be to negotiate with one lead private partner — a single hyperscaler, AI lab or infrastructure investor willing to take the anchor role — rather than running a multi-bidder tender that produces a fragmented set of relationships none of which can be steered quickly. Spain’s consortium model, with several large co-equal partners, is the structure to avoid, not copy.

What this looks like in practice

A practical sequence, designed for speed rather than process completeness, would look like this:

A Council of Ministers decision formally designating the Kozloduy–Mizia–Oryahovo corridor as the national site, removing further internal site competition.
Establishment of a single delivery vehicle, modelled on the Kozloduy NPP – New Builds precedent, with a named programme director and statutory coordination authority over ESO, environmental permitting bodies and land assembly.
Direct, exclusive-track negotiation with one anchor private partner, rather than an open multi-bidder tender, structured around the EU’s offtake financing model.
A single, consolidated submission to the EU’s formal gigafactory call built around that one decision, one vehicle and one partner — arriving at the Commission as a finished proposal rather than an open competition for the Commission to adjudicate internally.

The honest caveat

This model cannot remove every procedural constraint. EU public procurement and state aid rules still apply, and some degree of competitive process or notification will be unavoidable at points in the sequence, particularly around the private capital partner and any state aid granted under the Special Economic Zone framework discussed earlier in this document. The recommendation here is not to evade those rules, but to ensure that wherever discretion exists, Bulgaria uses it to concentrate decision-making in one place rather than diffuse it. The honest comparison is not “fast versus slow” in absolute terms — it is “one set of hands on the wheel versus many,” and the latter is the more common reason large infrastructure projects in Europe lose years, not bad luck or bad sites.

This is also the argument Bulgaria should make to the European Commission directly: that the EU’s gigafactory initiative will move faster, and produce fewer stranded or half-finished projects, if member states arrive with a single accountable delivery structure already in place, rather than asking Brussels to manage the coordination risk across dozens of fragmented national bids after the fact. A Union that wants sovereign compute quickly should reward member states that can demonstrate this kind of leadership model, not merely the ones with the largest consortium.

Infrastructure: What the Site Actually Needs

Two infrastructure questions decide whether this project is real: can the corridor be powered, and can it be connected. They look similar and behave completely differently. Power is the hard constraint, and it is hard everywhere in the world right now, for reasons that have nothing to do with Bulgaria. Fibre is the easy one, and Bulgaria's position on it is better than the corridor's rural character suggests. Everything else — land, cooling, transport, workforce — is a matter of diligence rather than doubt.

Kozloduy's grid infrastructure was built for six reactors

The strongest and least appreciated fact about the site is that Kozloduy's transmission infrastructure was not sized for the plant that operates there today. The site was developed with six reactor positions across roughly 1,000 hectares; four of those units were shut down under Bulgaria's EU accession commitments, leaving two 1,000 MW reactors running into a switchyard and transmission system originally built to evacuate substantially more power than currently flows through it.

This is a materially different starting point from a greenfield site. A new campus elsewhere begins by asking whether a transmission corridor can be built at all. Here the corridor exists, the substations exist, and the high-voltage lines exist — the question is narrower and far more answerable: how much of that headroom is available to a new, non-nuclear industrial load without reinforcement, and on what timeline.

That figure is not established in this paper, and it should not be guessed. It is the single most important number in the project, and obtaining it — from ESO's network development plan and Kozloduy's own grid documentation, naming the specific 400 kV and 110 kV lines and substations and their firm transfer capacity — is the first technical task the delivery vehicle should commission. Everything downstream depends on it. If the answer is favourable, Bulgaria has a grid advantage most competing sites cannot manufacture at any price. If it is not, the reinforcement plan becomes the critical path and the timeline changes accordingly.

The design brief that follows is straightforward: a first phase of 100–150 MW, expandable to 250 MW and then beyond 500 MW. The campus should sit in the wider triangle between Kozloduy, Mizia and Oryahovo — close enough to transmission and the Danube to be practical, on higher ground away from floodplain and protected riverfront habitat, and deliberately not adjacent to the nuclear plant itself or inside Kozloduy town, where security, permitting and public perception all become needlessly harder.

The real constraint is transformers, not permits

This paper has argued throughout that Bulgaria's advantage is speed. That argument holds for everything Bulgaria controls (site designation, permitting, land assembly) where the single-decision governance model can compress timelines that take years elsewhere. It does not hold for the hardware, and it is important to say so plainly.

Global lead times for large power transformers and generator step-up units have moved from roughly 7–14 months before 2020 to commonly 24 months, and 36–48 months for the largest high-voltage units, driven by a worldwide shortage of transformer manufacturing capacity, grain-oriented electrical steel and skilled fabrication. No country is exempt from this. It is not a Bulgarian weakness; it is the physics of the current market, and it constrains Spain, Romania and everyone else in the gigafactory race equally.

What differs is the response. If equipment procurement waits until permitting concludes (i.e. the conventional sequence) the queue starts late and the delay compounds. The recommendation here is the opposite: place long-lead orders for transformers and switchgear on the day the site decision is taken, in parallel with permitting rather than after it. This is precisely the kind of decision that a single delivery vehicle with real authority can make in a week and a fragmented consortium cannot make in a year, and it is where Bulgaria's governance model converts directly into months saved.

There is a second lever worth evaluating. Data centre developers facing the same constraint elsewhere are increasingly turning to on-site generation — "bring your own power" — to energise an initial load while the grid-connected build completes. For a corridor sitting beside a nuclear plant, the option deserves serious assessment rather than reflexive dismissal: it could bring a first 100–150 MW phase online materially earlier than the grid path alone permits.

Fibre is the easy problem, and Bulgaria is better placed than it looks

The corridor's telecoms position appears weak and is not. Bulgaria has a genuine national backbone to connect into: GCN alone operates more than 3,100 km of owned fibre with over 30 points of presence and international interconnections to Romania, Turkey, Greece, North Macedonia and Georgia, alongside A1, Vivacom, Yettel and carrier-neutral facilities including SDC Ruse. The task is to extend a spur from an existing backbone to a greenfield campus, not to build a backbone from nothing.

That distinction matters because, unlike transformers, fibre is not gated by a global manufacturing shortage. Trenching, ducting and splicing a dedicated route is civil works and permitting: months, not years, provided wayleaves and road-crossing permissions are secured early. This is the one part of the infrastructure case where Bulgaria's speed claim is unambiguously sound, and where the single-window permitting proposed under the Special Economic Zone would shorten the timeline further still.

The design requirement is redundancy: two physically diverse routes, not one. A route back toward Sofia and a route toward Ruse, where established carrier-neutral infrastructure and pan-European fibre already terminate. Both should be commissioned in parallel. A single-route campus is not a serious proposition for sovereign compute, and treating the second route as a later upgrade would be a false economy.

Cooling, land and the things that require diligence rather than argument

The Danube strengthens the site, but the paper should not pretend a river solves cooling on its own. High-density AI infrastructure is moving toward liquid and closed-loop cooling regardless of what water sits nearby, and the Danube's own hydrology, including the low-flow years seen across Central and Eastern European river systems, makes over-reliance on open-loop abstraction unwise. The right design is a higher-ground campus within reach of the river, engineered around closed-loop cooling, treating river water as supplementary rather than foundational, with the heat-use zone described in Section 5 planned into the surrounding land from the outset rather than retrofitted. This is a risk to be engineered around, and it is treated as such in the Risk Register.

Land is the least contentious element. What the campus needs is unremarkable: a large, flat or gently sloping site clear of settlement, floodplain and Natura 2000 constraints, close to transmission, with room for substations, storage, cooling plant, construction staging, security buffers and the heat-use industries around it. Northern Bulgaria has an abundance of exactly this, for the reasons set out earlier in this paper, and land price should be the last filter applied rather than the first.

Transport follows the same logic. Ruse offers the corridor's strongest existing infrastructure — road and rail, industrial parks with their own substation and utilities. It should serve as the equipment-staging and logistics node during construction. The campus site itself needs adequate road and rail access into the national network via Vratsa, Pleven and Sofia; this determines construction logistics, not strategy.

Workforce and housing are infrastructure, not social policy

The corridor cannot supply its own labour immediately, and pretending otherwise would be the fastest route to a fenced industrial asset with no regional impact. The local area can provide construction, security, facilities and logistics work. The wider Bulgarian base (Pleven, Vratsa, Montana, Vidin and Sofia) supplies the supplier ecosystem and mid-tier technical labour. Higher-end software, AI and data engineering capability has to come from Sofia Tech Park, INSAIT, the technical universities and deliberate diaspora-return recruitment, and it will not arrive by accident.

Which means housing, schools, healthcare, training and transport belong in the infrastructure plan, on the same page as the substation and the fibre route. If the objective is to reverse emigration rather than merely to host a facility, engineers and their families have to be able to move to this region and stay. Pleven is the natural place for that weight to land offering a real city with an existing labour market, close enough to serve the campus and large enough to absorb a population, which is why it is designated the operational and skills hub rather than an afterthought. The campus is the anchor; Pleven is where the people live.

The corridor's viability reduces to one question, and it can be answered in months rather than years: can a 100–150 MW first phase be connected, cooled, permitted and serviced in the Kozloduy–Mizia–Oryahovo triangle?

Answering it requires four pieces of work, in this order.

The first and most important is the firm grid capacity figure, and it can only come from ESO, the Bulgarian transmission system operator, which manages the national high-voltage network of more than 15,000 km of line and 297 substations. ESO's network development plan, read alongside Kozloduy's own grid documentation, is the only source that can establish what the plant's existing switchyard can carry to a new industrial load without reinforcement. The publicly available maps, e.g. ENTSO-E's grid map and OpenInfraMap, are useful for orienting the search and identifying which lines and substations to ask about, but both are schematic and neither is a substitute for a connection study. They tell you where to look; ESO tells you the answer.

The second is a land screen across the triangle, filtered on flood risk, protected habitat, settlement conflict, road access and distance to transmission. The InvestBulgaria investment map is the right first pass, since it consolidates industrial and logistics zones, technology parks and research institutions in one place, though it is more likely to surface support sites than the main campus, which will probably need a bespoke strategic site rather than an existing industrial park. Land price should be the last filter applied, not the first.

The third is environmental and water-use screening, which must be run before any site is promoted publicly rather than after. The Natura 2000 viewer and associated protected-site data are the starting point, and a Danube-facing project will need careful assessment of protected habitats, river-basin constraints, flood exposure and abstraction limits. Getting ahead of this is not merely good practice — a site announced and then withdrawn on environmental grounds would do more damage to Bulgaria's credibility with the Commission than a slower, quieter search.

The fourth is a fibre and transport route survey: confirming two physically diverse fibre routes as set out above, and assessing road and rail access into the national and European networks via the TEN-T corridor maps.

None of this is speculative work. None of it requires EU funding to begin. All of it can start immediately, and most of it can be completed within the window before the Commission's formal call closes.

If the answer is yes, Bulgaria has an anchor that no rival bid can replicate — nuclear-adjacent power on oversized transmission infrastructure, cheap expandable land, and a river, in the region of Europe that most needs the investment. If the answer is no, the fallback is a matter of diagnosis rather than despair: Pleven if the constraint is labour and support infrastructure, Ruse if it is connectivity or logistics.

The point is that this is a knowable question with a near-term answer, and nobody has yet commissioned the work to answer it. That, more than any argument in this paper, is what the Bulgarian government should do first.

The test that decides everything

Risk register

The hedges in this thesis have, until now, been distributed across the sections where they arose: nuclear timeline risk, ecosystem-thinness risk, procedural risk, heat-reuse sequencing risk and power-equipment lead-time risk. This section consolidates them into a single register, in plain terms, so that the case for this project does not depend on a reader assembling the caveats themselves.

No single item in this table is disqualifying on its own. Taken together, they describe what “behind procedurally” actually means in practice, and they are the reason the Governance section of this document argues for speed and single-point accountability rather than a longer, more conventional process: every additional month of indecision is a month in which several of these risks get harder rather than easier to manage.

Conclusion

Europe has already done the hard part. It has accepted that dependence on foreign compute is a strategic vulnerability, not a commercial inconvenience. It has written the strategy, passed the plan, and committed the money — €20 billion for up to five AI gigafactories inside a €200 billion programme. Every argument about whether Europe should build sovereign AI infrastructure has already been had, and settled. What is left is the easy part, and Europe is somehow making it the hard one: deciding where.

This paper has argued that one of those answers is obvious, and that its obviousness is precisely what makes the current inertia so difficult to justify.

Consider what the Kozloduy corridor solves simultaneously. Europe needs sovereign compute; Bulgaria has operating low-carbon nuclear baseload sitting on transmission infrastructure built for three times the reactors now running through it. Europe needs to build faster than its permitting regimes allow; Bulgaria has the cheapest land and cheapest labour in the Union, in a region so empty that land assembly presents almost no friction. Europe needs its cohesion policy to mean something; Bulgaria offers the poorest region in the entire Union, one that has lost half its population in places, where a single project could alter a thirty-five year trajectory of decline. Europe needs its AI strategy and its regional policy to stop being separate conversations; here, they are the same conversation.

Not one of these is a weak argument dressed up. Each would be sufficient on its own to justify serious examination. Together they describe a location where European strategic autonomy, European industrial policy, European cohesion funding and European demographic anxiety all point at the same patch of ground on the Danube and where, for now, nothing is happening.

The risks are real and they are small. The nuclear expansion may slip; the base case does not depend on it. Transformers take two to four years; so do everyone else's, everywhere in the world, and the answer is to order them on day one rather than after permitting. The local skills base is thin; that is what Pleven, Sofia's institutions and a serious diaspora-return effort are for. These are engineering and sequencing problems with known solutions. They are not reasons to hesitate. They are the ordinary difficulty of building anything significant, and they are dwarfed by the cost of the alternative, which is that Europe spends its sovereignty budget in the places where infrastructure is most expensive, land is most contested, power is most constrained, and the cohesion benefit is precisely zero.

Two things now need to happen and neither can wait for the other.

The Bulgarian government must stop treating this as a possibility and start treating it as a national project. That means a single decision at Council of Ministers level designating the Kozloduy–Mizia–Oryahovo corridor as the national site; a single delivery vehicle with real statutory authority, modelled on the structure Bulgaria already built for its nuclear new-build programme; one named person accountable for the timeline; and the commissioning, immediately, of the grid capacity study from ESO that no one has yet asked for. None of this requires EU money. None of it requires permission. It requires a decision, and the decision has been available for some time.

The European Union must fund this with its own members' money, and stop waiting for someone else to underwrite Europe's strategic autonomy. There is a contradiction at the centre of the current approach: the Union has declared that it cannot depend on foreign platforms for critical compute, and has then designed a funding model that depends on private capital, much of it foreign, to decide whether that compute gets built at all. Sovereignty that is contingent on a hyperscaler's investment committee is not sovereignty. If the Union means what it wrote in the AI Continent Action Plan, then the capital should follow the strategy directly, publicly, and at risk, because that is what strategic infrastructure has always required, and because the entire lesson of the past two years is that waiting to see whether someone else will provide is how dependence begins.

The window is measured in weeks, not years. The Commission's formal call is imminent. Spain, Romania, Austria and Czechia have delivery vehicles, committed capital and submissions in motion. Bulgaria has a case that beats all of them and a process that beats none of them. That asymmetry is entirely fixable, and it is fixable now, but it will not be fixable indefinitely, as gigafactories, once sited, are sited for a generation.

The strategic logic points here. The funding logic points here. The demographic and cohesion logic points here, more sharply than anywhere else in the Union. The engineering is tractable and the land is waiting. What is missing is not evidence, and it is not a business case.

What is missing is a decision. Europe should make it, and it should make it now.