Hosting AI capacity, controlling it, and using it are three different things – and on the third, Bulgaria and Europe lag furthest behind.

Kristalina Georgieva, Managing Director of the International Monetary Fund, opened her speech to EU finance ministers in Dublin on 19 September with a personal recollection: her first purchase in hard currency, in a Bulgaria where membership of the European Union still seemed unthinkable, was a can of Guinness. She drew the point that Europe has repeatedly proved capable of transformations few anticipated.

Two days later, on 21 September, Prime Minister Rumen Radev addressed the AmCham business breakfast in Sofia and set out the value Bulgaria can bring to the AI economy. In the preceding week, Christine Lagarde, President of the European Central Bank, and Mario Draghi, former ECB President and former Prime Minister of Italy, had each published their own assessment of Europe’s position.

Read together, the four contributions address different parts of the same problem. Kristalina Georgieva focuses on where Europe should build, pointing to the cheapest energy. Christine Lagarde explains why some capacity needs to be European, citing data and the need for continuity of access. Mario Draghi sets out how the build-out might be financed and which uses require European control. Rumen Radev presents Bulgaria as a location for it. Taken together, they suggest that the question facing Bulgaria is less whether to host AI factories than on what terms, and for whose benefit.

Europe’s case: build capacity, protect data, pool demand

Kristalina Georgieva argued that Europe needs to be a provider, a builder, and an adopter of AI, since an economy of the EU’s size cannot realistically outsource all of it. On energy, she called for deeper integration of national grids through an energy union and, in the meantime, siting data centres where electricity is cheapest.

Christine Lagarde, speaking in Vienna, was careful not to argue against importing technology; she noted it is not inherently harmful, and Europe benefits from American investment in AI. She identified, however, three respects in which AI differs from earlier technologies. The first concerns data: a cloud service stores information, whereas a model reads it, and almost half of the firms that considered AI and decided against it cited data protection. The second concerns access: within a few years AI will screen goods at the border, select tax returns for audit, dispatch trains, monitor hospital patients, and clear payments, so any withdrawal of access, or change in its terms, would affect every sector at once. The third concerns the frontier, since in highly competitive sectors such as pharmaceuticals, finance, and defence a modest advantage in capability can translate into disproportionate gains.

Her proposed response proceeds in three steps: more European computing capacity, models that are ‘good enough’ for most purposes and run on European infrastructure, and continued access to frontier systems. She notes that the gap between European data centre demand and installed capacity already stands at around 3 GW and is expected to reach about 20 GW by 2036.

Mario Draghi’s article in the Financial Times is the most specific of the four, combining quantitative evidence, a principle for deciding which uses require European control, and a financing model already in operation. The EU hosts less than 5% of the world’s AI computing capacity, compared with some 75% in the United States. As sovereignty considerations gain weight, he argues, compute ceases to be fully interchangeable, and much of the infrastructure that stores and processes European data will need to be located in Europe. American operators can supply much of this through so-called sovereign cloud services, which remain under American ownership; the most sensitive uses, however, would need to run under European control.

This tiered approach underlies the Commission’s Cloud and AI Development Act, proposed on 3 June 2026 and currently under negotiation. The proposal sets out four sovereignty assurance levels for cloud and AI services. They progress from processing and storing data on infrastructure located in the EU, through independence from third countries and transparency over the software supply chain, to EU ownership and control, and finally full control of the supply chain without third-country interference. Public bodies would select the level appropriate to the risks of their workloads, and the Commission may recognise certain non-EU providers at the third level.

Mario Draghi also examines what is holding Europe back. Part of the problem lies on the supply side: owing to permitting and grid-connection delays, a data centre takes around 24 months to build in the United States and 42 months in Germany. In his view, however, the principal obstacle is demand. Investors require long-term offtake contracts before committing capital. At the same time, European demand remains dispersed across millions of firms, which helps explain why European AI data centre operators such as Nscale sell most of their capacity to American buyers.

His proposed remedy is demand pooling. A group of European companies, including ASML, Capgemini, and Amadeus, has committed to multi-year purchases of compute from the French AI company Mistral, aiming to underwrite 1 GW of capacity by 2030. He also sets conditions for host regions: clean energy, a fair contribution to the grid costs data centres create, and tangible benefits for local communities.

Bulgaria’s value: energy, storage, and fast connection

Rumen Radev told AmCham that companies from developed countries want to build large AI factories in Bulgaria because, in their home markets, they face waits of around five years for a grid connection. Bulgaria, he argued, can offer surplus energy, an increasingly green supply, a stable grid, and, above all, rapid connection.

The energy claim is at least partly borne out. According to data from ENTSO-E, the European network of electricity transmission system operators, reported by the Bulgarian business weekly Capital, Bulgaria’s operational battery storage reached 5.4 GW in mid-August, which strengthens the grid’s capacity to absorb variable renewable generation. Whether a particular site can offer a large data centre a rapid connection would, however, still require a site-specific assessment.

Rumen Radev explicitly invited American investment in AI factories, large data centres, robotics, and semiconductors. He also stressed that the government’s interest lies less in selling electricity than in the high-technology ecosystem such centres could stimulate, in education, space, defence, robotics, and semiconductors.

Where the value holds up, and where it falls short

On the supply side, the value Bulgaria brings responds directly to the problem the others describe. Mario Draghi’s 42 months in Germany and the five-year waits reported by Rumen Radev’s interlocutors reflect the same bottleneck: permitting and, above all, grid connection. Ireland illustrates where this can lead, having restricted new data centre connections in the Dublin area because the grid could not keep pace.

Mario Draghi’s answer is to build where construction is cheap and fast. He cites Sweden, where an AI data centre costs only about a tenth more than one in China. Bulgaria is positioning itself as a location of this kind, with connection speed as much a part of its value as the price of power.

The Commission’s own evidence qualifies this, however. The impact assessment accompanying the Cloud and AI Development Act notes that Bulgarian zoning regulations lack an adequate category for data centres. It records stakeholder evidence that this gap causes long delays and deters investment. The same assessment identifies Bulgaria as a developing region whose data centre market EU-wide fast-track measures could unlock. Bulgaria was also among the four Member States from which the Commission’s public consultation received no responses. Speed of grid connection may therefore be a genuine advantage, while permitting could be viewed as an obstacle.

Data centres can be permitted today, but only by interpretation. Bulgaria’s Spatial Development Act has no category for data centres: they are permitted as generic production or service buildings, graded mainly by headcount. The Act does place construction with an uninterruptible technological process in the first category, but the ordinance on the nomenclature of construction works illustrates this with metallurgy, chemicals, and machine-building and does not mention data centres. A large facility’s category, and therefore which authority permits it, is thus a matter of interpretation by the municipal chief architect. The existing Sofia sites avoided the question by building inside industrial zones. The Act does, however, let the Council of Ministers declare any project an object of national significance, which moves permitting to the Ministry of Regional Development and Public Works, and the Investment Promotion Act already treats data processing and hosting as a high-technology service, with administrative deadlines a third shorter, a single point of contact, and land without tender. The scarcer resource is grid access. By August 2026, requests to connect data centres to the grid, including investment intentions, totalled 9.3 GW, more than the country’s winter peak demand of around 7 GW. These are requests for access, not capacity under construction: only 3.8 GW had progressed to a connection opinion from the transmission operator, and those opinions are valid for one year.

On ownership, Mario Draghi’s framework is more useful than a simple opposition between European and American capacity. By his standard, American-owned capacity in Bulgaria could legitimately serve most of Europe’s workloads. The more pertinent question is whether the value Bulgaria brings is confined to that tier.

Under the Commission’s proposal, providers controlled from outside the EU can qualify for the first two sovereignty levels, with the second only if they have safeguards against third-country access and interference. In contrast, the third and fourth levels require EU ownership and control, subject to a narrow possibility for the Commission to recognise third-country safeguards at the third level. The Commission’s working assumption is that around 90% of public-sector use cases require Level 1 or 2, and around 10% Level 3 or 4. If all new AI factories in Bulgaria remain under American ownership and control, the country could host capacity for most of Europe’s public-sector workloads, but not the most sensitive tenth: public-sector records, health data, and industrial data of the kind Mario Draghi identifies as Europe’s remaining sovereign asset. Given that defence features on Rumen Radev’s own list of target sectors, that tenth is unlikely to remain theoretical.

An ecosystem already in place, and a gap in adoption

Bulgaria already has the foundations of a European AI ecosystem, much of it located at Sofia Tech Park. INSAIT, the Institute for Computer Science, Artificial Intelligence and Technology at Sofia University, develops the models, including BgGPT, an open family of Bulgarian-language models now in its third generation. The Discoverer supercomputer supplies the compute, including Discoverer+, a GPU partition of 32 NVIDIA H200 accelerators. BRAIN++, the €90 million AI Factory backed by the EuroHPC Joint Undertaking and developed with INSAIT, aims to extend these foundations with a new supercomputer (Discoverer++), a federated data lake, and services for start-ups, SMEs, and public bodies, some of which it already offers. InnovationAmp, a European Digital Innovation Hub, offers firms subsidised AI and HPC services.

This is precisely the foundation on which the ecosystem Rumen Radev envisages could be built. Rather than waiting for an ecosystem to grow around new AI factories, Bulgaria could require new capacity to connect to what already exists. New factories would then extend an ecosystem already in place, rather than being expected to create one.

The stack is under European control, although its foundations are not entirely European. BgGPT 3.0 is built on Google’s open Gemma 3 model. Its published weights can be retained and deployed independently. Still, the current development approach depends on externally developed foundation models and non-European computing hardware, which is the kind of dependency Christine Lagarde identifies for open models.

The stack is also modest in scale: the Discoverer+ partition supports research and smaller AI workloads, far below the large commercial data centres discussed at the AmCham event.

The more significant gap, however, concerns use. In 2025, 8.55% of Bulgarian enterprises with ten or more employees used AI, compared with 19.95% across the EU, the third-lowest share in the Union. The gap relative to the EU average has widened from 4.4 percentage points in 2021 to 11.4 percentage points. Among large enterprises, the figures are 26% in Bulgaria and 55% across the EU.

Kristalina Georgieva’s central point is that productivity gains come from changes in how firms operate rather than from the technology alone, and Mario Draghi’s virtuous circle depends on domestic demand financing domestic capacity. Gigawatt-scale factories serving foreign firms that cannot connect at home would not, in themselves, raise Bulgarian adoption. Kristalina Georgieva’s own suggestion is that the state lead by example: in Ireland, four in five firms report that they would be more likely to adopt AI if the state integrated it into its own systems.

What Bulgaria could negotiate

Mario Draghi’s tiered approach and his conditions for host regions, together with the European stack already in place, point towards a possible negotiating agenda for Bulgaria:

  • Equal access to fast connection for pooled European buyers and European AI data centre operators, not only for American investors, since Mario Draghi’s pooled buyers require sites that Bulgaria is in a position to offer. This requires queue discipline for large consumers, so that fast connection goes to projects that build rather than to the 9 GW of grid access requested so far on paper.
  • A share of capacity under European control, large enough to host the sensitive tier and linked to the existing European stack at BRAIN++ rather than built anew.
  • Access for Bulgarian firms, with part of the new capacity offered to Bulgarian start-ups and SMEs through BRAIN++ and InnovationAmp, not only to foreign tenants.
  • A fair contribution to grid costs from those who connect.
  • Ecosystem commitments set out in the agreements themselves, rather than in public statements.

Part of the work is domestic and does not depend on any investor. Bulgaria could add a data centre category to the Spatial Development Act and to the nomenclature of construction works, and designate acceleration zones in advance next to existing high-voltage substations. It would not need new instruments to do so, as priority-project status under the Investment Promotion Act could cover data centres if its jobs threshold were adjusted, and sites above 50 MW could be declared objects of national significance. Together, these steps would close the permitting gap the Commission has identified and put in place, ahead of time, the acceleration zones the Cloud and AI Development Act would require of every Member State.

Bulgaria’s bargaining position rests on what others currently lack: time to connect. Christine Lagarde’s advice to Europe as a whole, to become a link in the chain that others cannot afford to remove, would seem to apply here as well. Speed of connection gives Bulgaria that position for now; the terms it negotiates will largely determine whether it retains it. That advantage will not last on its own: with 9 GW of grid access already requested, more than the country’s peak demand, it holds only if connection goes to projects that actually build. Timing also matters: the Cloud and AI Development Act is not expected to enter into force before 2029, so the terms Bulgaria sets now will precede the EU rules rather than follow them.

The timing also points to a second task. The Act is still under negotiation, and the Commission’s public consultation received no responses at all from Bulgaria, whether from government, business, or academia. Securing good terms at home is only half of the work; the other half is taking an active part in shaping the European rules within which those terms will have to operate, from the sovereignty levels to joint procurement.

Resilience, autonomy, and use through the AI-SAF lens

The three questions running through this article – where AI capacity is located, who controls it, and who uses it – are easy to blur in public debate and harder to measure. My open assessment framework is designed to distinguish them systematically.

The AI Strategic Autonomy Framework (AI-SAF) reads these questions along two axes that it keeps separate. Resilience asks what still works if a supplier cuts off access tomorrow: fallbacks, reserves, diversification, and control over the infrastructure one relies on. Autonomy asks who decides the technology path: whether a country or firm can shape its own models and data, develop its own talent, and set its own rules. A subject can be strong on one and weak on the other, which is why the two are reported side by side rather than averaged. The framework also distinguishes capability, meaning what is in place today, from commitment, meaning what has been planned or announced.

On that reading, the AI factories discussed at AmCham are primarily a resilience question. Bulgaria’s commissioned battery storage is delivered capability that directly supports resilience. Computing capacity located in Bulgaria but controlled from another jurisdiction is different: that jurisdiction can restrict access or switch it off by decision. Whether such capacity strengthens Bulgaria’s resilience therefore depends on who controls it and under which law, which is the same question the Commission’s sovereignty levels ask. Until investment, construction, grid connection, and operation are demonstrated, the factories also remain commitments rather than capability.

Autonomy lies elsewhere, in the stack Bulgaria already has. INSAIT’s BgGPT models and the existing Discoverer infrastructure are delivered capability, and published model weights add resilience, since they cannot be withdrawn once downloaded. Their autonomy is nonetheless bounded: each new generation depends on externally developed foundation models and non-European chips. Discoverer++ and the full BRAIN++ offering should be assessed according to their actual implementation status rather than their announced scope.

The national variant of the framework does not measure how widely firms use AI; the corporate variant captures part of it through AI literacy and deployment speed. Control over AI is nonetheless of limited value unless the economy actually relies on it, and the Eurostat figures already answer that question: it is the dimension on which Bulgaria needs to improve.

Kristalina Georgieva is right that Europe has repeatedly proved capable of transformations few anticipated. Bulgaria has built a strong position in energy. What it gains in return will depend on the terms it sets now, and on whether Bulgarian and European firms more broadly begin to use what is already in place. With only one in five EU enterprises using AI, the adoption gap is a European problem, and arguably a larger one than the infrastructure gap that dominates the debate.

Sources