The silicon and the soil: Engineering Europe's sustainable AI ecosystem

On the 9^th of June, Lubomila Jordanova, CEO of Diginex Group, delivered the keynote address at the European Commission’s annual European Digital Innovation Hubs (EDIH) Summit. The address explored why the structural sustainability of rapidly growing AI infrastructure is the most critical issue facing deployment of these systems in Europe and, more importantly, how to build sustainability into our models from minute zero. Read the full text below.

9th of June 2026, Brussels, Belgium – As Europe transitions from setting landmark tech policies to operationalising large-scale artificial intelligence, a critical question echoes through the halls of the Square Meeting Centre: Is our innovation ecosystem actually functioning effectively?

With the AI Act fully enforced, the AI Continent Action Plan adopted, and the Apply AI Strategy actively guiding small and medium-sized enterprises (SMEs) and public administrations, the legal scaffolding is secure. Yet, a glaring reality remains unaddressed. An artificial intelligence ecosystem that is not sustainable will, eventually, cease to function entirely. True innovation cannot exist in a thermodynamic vacuum. We must bridge the gap between our boundless digital ambition, the silicon, and our physical, finite planetary boundaries, the soil.

‍

The illusion of weightless innovation

The tech sector has long hidden behind the comforting metaphor of the "cloud." It implies something weightless, immaterial, and floating harmlessly in the stratosphere. But the cloud does not live in the sky; it lives in our electrical grids, our rivers, and our soil. Every parameter update, every automated agentic workflow, and every single token generated triggers a very real, physical cascade of resource consumption.

Behind the sleek API endpoints sits a massive, resource-heavy infrastructure: high-voltage transmission lines, concrete data centres, thousands of rotating cooling fans, and intensive hardware supply chains. If European Digital Innovation Hubs (EDIHs) guide thousands of businesses toward AI adoption without accounting for this physical overhead, they are not deploying software assets; they are distributing resource liabilities.

Consider the sheer physical cost of our daily digital habits. Generating a single AI-driven image consumes as much raw electricity as running a standard home LED light bulb for 17 continuous minutes. Furthermore, a routine conversation consisting of just 20 to 25 prompts with a frontier Large Language Model effectively "drinks" a 500ml bottle of fresh water just for server-rack cooling. When we scale up to deeper tasks, executing complex agentic reasoning chains or generating a single high-definition AI video requires up to 4.1 litres of fresh water, which actually surpasses the recommended daily drinking requirements for an adult human.

We are actively building an artificial intelligence that drinks more water than the humans who created it.

‍

Macro critical data and the efficiency trap

This challenge has been brought into sharp focus by immediate global data. The United Nations University emergency report reveals a sobering truth: the global water footprint of data centres has tripled over the last 24 months. Crucially, the report dispels an industry myth: 90% of total AI energy consumption occurs during the inference phase, the daily execution of prompts by businesses and citizens, not the model training phase. As we scale AI across millions of European enterprise workflows, this cumulative inference footprint poses a structural threat to municipal grids and local aquifers.

When confronted with these numbers, the tech industry traditionally points to hardware efficiency as our ultimate saviour. We are told that as chips become faster and more optimised, the environmental problem will naturally dissolve. However, this relies on a dangerous misunderstanding of economic history. More than a century ago, the economist William Stanley Jevons observed that when you make a resource more efficient to use, you do not use less of it. Instead, because it becomes cheaper, faster, and more accessible, demand explodes exponentially. This is the Jevons paradox, and we are walking straight into its trap.

Moving to sub-2nm chip architectures makes each individual AI prompt incredibly cheap. But by making it cheaper, we invite businesses to run billions more prompts every single hour. Hardware efficiency alone does not lower our footprint; it merely accelerates our consumption. Efficiency is an engineering milestone, but sustainability must be an intentional ecosystem choice.

‍

The strategic shift: Brute force vs. green AI by design

Europe should not compete with global tech ecosystems by copying their unchecked, brute-force approach to scaling. Silicon Valley has long operated on a philosophy of moving fast and breaking physical grids; Europe must move smart and build what lasts. Our distinct competitive advantage lies in pioneering green AI by design, a philosophy that treats environmental impact as a core performance metric equal to latency, accuracy, and cost.

To achieve this, we need to fundamentally reshape how we deploy technology across four critical dimensions, moving away from a legacy model of unconstrained consumption and toward a framework of digital craftsmanship.

First, we must reconsider model architecture. The dominant industry standard relies on passive, generic, multi-billion parameter models. This is the computational equivalent of using a commercial rocket ship just to drive across the street. It is massive overkill for the vast majority of enterprise requirements. The green AI alternative champions small, highly specialised, domain-specific models. By deploying a compact model that is hyper-optimised for one precise business task, an enterprise can achieve identical or superior accuracy at a fraction of the environmental overhead.

Second, we must change where this intelligence physically lives. The status quo tethers European businesses to hyperscale, centralised cloud data centres, requiring immense amounts of data to travel thousands of miles across the globe. Green AI by design shifts this gravity toward localised, edge-based systems and regional green AI factories. By processing data locally and tying our infrastructure directly to regional renewable microgrids, we keep both the economic value and the resource consumption contained and sustainable.

Third, our environmental strategy must evolve from a retrospective accounting exercise into an active engineering discipline. For too long, corporate sustainability has meant reactive reporting and paper carbon offset purchases, a practice that amounts to little more than greenwashing. Green AI by design embeds thermodynamic efficiency directly into the codebase itself. Through advanced techniques like quantisation and lean algorithmic routing, European software can be engineered to require less memory and less mathematical computation from day one.

Finally, this shift fundamentally redefines the software, network, and edge cost profile for small businesses. Relying on massive external cloud infrastructure leaves European SMEs vulnerable to unpredictable, volatile monthly subscription and compute fees dictated by foreign monopolies. Conversely, utilising compact, localised models on regional infrastructure creates a highly stable, low-cost processing environment. Sustainability, in this light, ceases to be a regulatory burden and becomes an economic shield.

‍

An action plan for Europe’s tech gatekeepers

To turn this vision into a reality, our infrastructure must evolve. The European Digital Innovation Hub network, our Testing and Experimentation Facilities, and our AI Factories can no longer act as neutral technical advisors; they must become active green architects.

This transformation begins at the point of entry. Every single hub consultation with an SME or public administrator needs to lead with two non-negotiable questions: What exact business value does this model provide, and what is its expected inference footprint? We must actively steer businesses away from massive, bloated, generic models when a compact, highly specialised, 7-billion parameter alternative can deliver the exact same results at a fraction of the energy cost.

Furthermore, we must change how we test and host these systems. We need to leverage our testing facilities to run strict energy-efficiency stress tests on software before a single line of code is deployed to the public. Simultaneously, we must mandate that new European AI Factories operate entirely on localised renewable microgrids backed by closed-loop liquid cooling systems.

‍

Guarding the horizon

Choosing green AI is not a regulatory limitation or an environmental tax; it is an economic superpower. By adopting optimised, compact, and localised architectures, we do not just save carbon; we build a highly stable, low-cost, and entirely self-reliant enterprise tech ecosystem for Europe.

As we look toward the horizon, the ultimate legacy of this digital transition must be the deliberate binding of our digital future to our planetary survival. We have a profound responsibility to ensure that the intelligence we build does not destroy the world that hosts it. Let us never burn the soil just to scale the silicon.

‍