Interview — Professor Martijn Heck (2)
The Architecture of Intelligence: Why Better Chips Matter More Than Bigger Datacenters
I travelled to Eindhoven again, this time not to speak about geopolitics or sovereignty, but about something far less visible — architecture. Not the architecture of buildings, but of systems. Because in the age of artificial intelligence, power no longer resides primarily in steel, land or energy. It resides in the way information moves.
Professor Martijn Heck received me in the same understated manner as before: no grand statements, no rehearsed narrative, only a willingness to follow the logic wherever it led. Our previous conversation had revolved around photonics and Europe’s position in the world. This time, I asked a simpler question:
What, in his view, is the core of the story?
He did not hesitate.
“Heterogeneous integration.”
Prof. Dr. Martijn Heck
Professor of Photonic Integration — Eindhoven University of Technology
The term sounds technical, almost administrative. But as he spoke, it became clear that he was describing something far more consequential: the merging of fundamentally different technologies — electronic chips, photonic components, radio-frequency systems, memory and packaging — into a single functioning whole.
Not a better component, but a different kind of system.
The AI Factory Risk
Public debate increasingly revolves around “AI factories” — massive data centres designed to train and run large models. Governments announce investments in the hundreds of millions, sometimes billions, as if scale alone guarantees relevance.
Heck is sceptical.
“The lifetime of an AI factory is extremely short — perhaps three years. If construction is delayed by one year, you may already have lost half the economic value. If it is delayed by two years, you risk building something obsolete on the day it opens.”
Prof. Dr. Martijn Heck
The problem is not the building itself, but the pace of technological change. Hardware generations move faster than infrastructure projects.
A provocative alternative emerged during the conversation: acquiring existing facilities from technology giants once they are depreciated. It is not glamorous, but it reflects how the capital cycle in high-tech actually works.
Prestige projects appeal to politics. Efficiency does not.
From More Power to Better Architecture
At one point, the discussion turned to energy — an issue dominating European policy. Data centres consume vast amounts of electricity and AI is expected to multiply that demand.
Yet Heck reframed the dilemma entirely.
“You can double capacity by building more data centres or you can achieve a similar effect by improving chip efficiency. The second option is usually far more effective.”
Prof. Dr. Martijn Heck
This observation exposes a deeper cultural pattern. Industrial societies historically solved problems through scale: more factories, more fuel, more infrastructure. But digital systems behave differently. Small architectural improvements at the chip level propagate through the entire stack.
In that sense, the real bottleneck of AI is not energy supply — it is the efficiency of computation and communication at microscopic scales.
It is a shift from nineteenth-century thinking to twenty-first-century physics.
Light Meets Logic
The core of heterogeneous integration lies in recognising that no single technology excels at everything.
Electronic chips are unparalleled for logic and control. Photonic systems excel at transmitting information rapidly and with minimal energy loss. Radio-frequency components handle wireless communication. Memory technologies store state.
The future lies in combining them.
“Electronics for thinking, photonics for transport. Each technology does what it is best at.”
Prof. Dr. Martijn Heck
Traditional chips rely on dense networks of copper interconnects. As complexity increases, these become sources of heat, delay and energy loss. Light, by contrast, can move data without electrical resistance and with far greater bandwidth.
Modern data centres already use optical connections extensively between racks and servers. The next step is bringing that optical capability directly onto or into the chip package itself.
Not replacing electronics — augmenting it.
The Quiet Priority Gap
One of the more striking themes was the mismatch between political investment priorities and technological realities.
Quantum computing receives enormous funding across Europe. While its long-term potential is undeniable, its practical impact remains uncertain and distant for most applications.
Meanwhile, advances in semiconductor architecture determine the performance of AI systems today.
“AI runs on chips. If you want progress in AI, you need progress in chips.”
Prof. Dr. Martijn Heck
Companies such as NVIDIA, TSMC and Samsung dominate this landscape not because of isolated breakthroughs, but because of sustained system-level innovation — design, manufacturing, packaging and software ecosystems evolving together.
Europe possesses significant expertise, particularly in equipment and research, yet risks fragmentation if investments are not aligned with where value is actually created.
Ecosystems, Not Isolated Breakthroughs
Institutions such as IMEC, ASML and the Brainport Eindhoven cluster illustrate what coordinated innovation can achieve. Universities, industry and suppliers form a tightly coupled network.
Students trained in this environment rarely struggle to find positions — not because of shortages alone, but because the ecosystem itself generates opportunity.
Still, global competition remains intense. Manufacturing leadership resides largely in Asia, financing power in the United States and regulatory frameworks in Europe.
“Innovation is not just about invention. It is about the entire chain from concept to production.”
Prof. Dr. Martijn Heck
Beyond the Strategies of the Past
Technological strategy often lags behind technological reality. Policies are shaped by conditions that existed ten or twenty years earlier — the telecom boom, earlier semiconductor cycles, different geopolitical assumptions.
But the integration wave now underway demands new thinking.
Co-integration of photonics and electronics, advanced packaging, three-dimensional stacking and specialised accelerators are redefining what a “chip” even is. It is increasingly a system in miniature.
In that sense, heterogeneous integration is less a single technology than a new paradigm of engineering.
A Physics of Intelligence
Listening to Heck, one realises that artificial intelligence is not only a software phenomenon. It is constrained — and enabled — by physical processes: heat dissipation, signal delay, energy conversion and material properties.
The efficiency of intelligence ultimately depends on the efficiency of moving information.
As I argued elsewhere, the next phase of computing may be defined less by raw processing power than by the physics of communication — how quickly and cleanly information can traverse a system. The convergence of light and electronics is a tangible manifestation of that shift.
Heck did not frame it philosophically. He simply described what works.
Engineers rarely speak in metaphors. Nature itself is precise enough.
The Horizon Ahead
As the conversation drew to a close, the concept of heterogeneous integration no longer felt like a technical detail. It resembled a strategic compass.
Not a promise of instant transformation, but a direction in which multiple constraints — energy, performance, cost and scalability — can be addressed simultaneously.
If artificial intelligence is to continue expanding without overwhelming infrastructure, environment and budgets, progress must occur at the level of architecture rather than sheer scale.
Better chips, not merely bigger facilities.
In the quiet corridors of the university, students moved between labs carrying notebooks and laptops, unaware that the systems they will design may determine the trajectory of global technology for decades.
No announcements. No spectacle. Only work.
Perhaps that is where the real future of AI is being built — not in monumental factories, but in the subtle integration of light and logic on pieces of silicon measured in millimetres.
And perhaps the most important question is not how powerful our machines will become, but how intelligently we design the foundations on which they run.
Questions for Further Study
- How can heterogeneous integration reduce the energy footprint of large-scale AI systems compared to expanding data-centre infrastructure?
- What economic models best capture the rapid depreciation cycles of AI hardware and infrastructure?
- Which applications will benefit most from on-chip photonic communication and where do electronic solutions remain superior?
- How can Europe leverage existing research strengths to gain strategic influence within the global semiconductor ecosystem?
- What investment strategies balance long-term exploratory technologies (such as quantum computing) with immediate system-level innovation in chips?
