Casting Intelligence in Silicon

Why on-device AI will define what 6G networks can—and cannot—understand

While much of the public conversation around 6G still revolves around spectrum, speed and new antennas, a far more consequential shift is happening quietly beneath the surface. The real battle is no longer in the air interface, but inside the chip. And in that domain, one company plays a far more decisive role than is often acknowledged: Qualcomm.

As operators such as BT and Telefónica focus on capacity and operating costs, and vendors like Nokia and Ericsson design the physical infrastructure, Qualcomm is grappling with a different question altogether: how can intelligence scale without becoming energetically unsustainable?

From Cloud Intelligence to On-Device Understanding

The traditional AI model — collect data, send it to the cloud, process it in centralized data centers — is running into hard limits. Not only in terms of latency and privacy, but more fundamentally in terms of energy. The global power grid simply cannot support an endlessly expanding cloud-AI paradigm.

Qualcomm’s response is clear and deliberate: on-device AI. Intelligence that runs directly on the device itself, embedded in the chipset.

“One of the biggest bottlenecks for large-scale AI deployment is energy. If we want AI everywhere, it cannot all run in the cloud.”
Cristiano Amon
CEO — Qualcomm

This shift is not ideological. It is structural. If AI is to become ubiquitous, it must move closer to the source — and become radically more efficient.

Efficiency as a Prerequisite, Not an Optimization

At the heart of Qualcomm’s strategy lies its chipset architecture. Modern Snapdragon platforms integrate multiple specialized processing units — CPU, GPU, NPU and sensing hubs — each optimized for specific workloads. Through what Qualcomm refers to as heterogeneous computing, tasks are dynamically routed to the most energy-efficient unit available.

The result is not just performance, but feasibility. Multiple international studies suggest that local AI inference can reduce energy consumption per task by a factor of 100 to 1,000 compared to cloud-based processing. For operators and policymakers, this is not a marginal gain — it is the economic foundation of 6G.

“Energy efficiency is no longer an optimization goal. It is the condition for scalability.”
Senior research statement
World Economic Forum

Beyond the Smartphone

Crucially, Qualcomm does not frame on-device AI as a smartphone feature. It sees it as the backbone of a broader ecosystem:

• XR and AR devices capable of real-time interpretation and translation
• Always-on sensing for context-aware services
• Edge devices that can act autonomously, even without cloud connectivity

In this model, devices cease to be passive endpoints. They become active participants — local AI agents that interpret, filter and decide.

Yet this is where the deeper tension begins.

The Semantic Boundary

On-device AI is fast, private and energy-efficient. But it is also constrained. Models running on devices must be smaller, more selective and purpose-built. That means making choices — about language, context, relevance and meaning.

“If we embed intelligence into devices, we also embed assumptions. Those assumptions become invisible once they are cast in hardware.”
Kees Hoogervorst
Independent Analyst — Altair Media

The core challenge therefore shifts from encryption to interpretation. Not only is the data secure?
But also: does the system understand what it is processing?

When models are primarily trained on dominant languages and cultural frameworks, the risk of semantic blindness emerges — systems that function flawlessly in technical terms, yet fail to grasp context, nuance or cultural meaning.

Hardware as a Silent Standard-Setter

This is where Qualcomm’s role becomes uniquely influential. As a chipmaker, the company does not merely determine performance and efficiency. It implicitly defines which forms of understanding can exist locally — and which cannot.

“The most powerful technologies are not the ones we see, but the ones that quietly define what is possible.”
Evgeny Morozov
Technology critic and sociologist

Conclusion: Efficiency Is Not Neutral

On-device AI is neither a gimmick nor a marketing trend. It is a necessary re-architecture of the digital foundations beneath 6G. Qualcomm has demonstrated that intelligence can scale efficiently and sustainably — positioning itself as a cornerstone of next-generation networks.

But efficiency alone does not confer neutrality. What is embedded in silicon is difficult to revise.

“We risk building networks that are technically secure, energetically efficient — and semantically fragile.”
Kees Hoogervorst
Independent Analyst — Altair Media

The future of 6G will therefore not be decided solely by standards bodies or spectrum auctions, but by a deeper question: who determines what our devices are able to understand?

Why AI Needs Formula One Power — and When It Doesn’t

What toddlers, museums and modern computing reveal about the real purpose of compute power

A toddler sees a giraffe once. The next day, walking through a museum, the same child looks at a skeleton and immediately recognizes the animal again. No manual. No training cycle. No second explanation required.

Artificial intelligence approaches the same task very differently. It may need millions — sometimes billions — of examples to achieve a comparable result. Entire data centers operate day and night to detect patterns that children seem to grasp intuitively.

This contrast, recently highlighted in a column by physicist and science thinker Robbert Dijkgraaf, captures a growing paradox in modern technology. At the very moment computing power is expanding at unprecedented speed, understanding itself remains elusive.

If toddlers can learn with so little information, why do machines require so much?

The answer lies not in intelligence levels, but in the nature of learning itself.

Learning Without Datasets

Human learning is not based on isolated data points. Children learn through interaction — through tone of voice, facial expressions, repetition, correction and encouragement. Meaning is shaped socially, not statistically.

“Children don’t learn from isolated data points; they learn through emotional resonance and social feedback. A toddler doesn’t need a billion parameters because they have a social compass — a parent or a peer — who provides instant context”,
— Ruben Fukkink, Professor of Pedagogy and Child Development, University of Amsterdam

Artificial intelligence lacks such a compass. It does not receive meaning through interaction but infers patterns through correlation. Where a child builds an internal model of the world, AI optimizes probabilities.

In that sense, modern AI systems are extraordinarily capable calculators — yet socially and contextually deprived. To compensate, they rely on scale.

Power as a Substitute for Intuition

The consequence of this design choice is visible in the infrastructure behind AI. Models grow larger, data centers expand and energy demand increases sharply.

“We are currently building Formula One cars just to do the groceries. If we want AI to reason more efficiently, we shouldn’t only build bigger data centers — we need to rethink how information moves at the physical level,”
— Martijn Heck, Professor of Electrical Engineering, Eindhoven University of Technology

The contrast with human cognition is striking. The human brain operates on roughly twenty watts — less energy than a household light bulb. Large AI models, by comparison, may require megawatts of continuous power to function at scale.

This does not indicate failure. It reveals a difference in architecture. Machines compensate for missing intuition with computation.

Where Compute Truly Matters

In many domains, this brute-force approach is not excessive — it is essential.

Climate modeling depends on high-performance computing to simulate planetary systems that no human mind could fully grasp. Drug discovery, protein folding and materials science rely on vast computational exploration to uncover patterns invisible to traditional research.

In such contexts, compute power functions as a scientific accelerator. It compresses decades of trial and error into months or even weeks.

The same applies to real-time systems such as autonomous vehicles, robotics, aviation and industrial automation. Here, milliseconds matter. Failure is not theoretical but physical. Unlike toddlers, machines operating in the real world cannot afford playful experimentation.

In these environments, scale is not indulgence. It is safety.

When More Power Adds Little Understanding

Yet many everyday AI applications do not operate under such constraints. Tasks such as summarization, assistance or recommendation often rely on cloud-scale processing even when the cognitive challenge itself is limited.

As energy costs rise and sustainability concerns deepen, this imbalance has become harder to ignore.

“A toddler sees a giraffe once and understands the essence of ‘giraffeness’ — even in a sketch or a skeleton. AI sees millions of pixels and calculates a probability. The paradox of modern computing is that we are scaling power when we should be scaling insight,”
— Inspired by Robbert Dijkgraaf, Physicist and President-Elect, International Science Council

The question is no longer whether AI can become faster. It already is. The more pressing issue is whether speed alone leads to better intelligence.

The Shift Toward Intelligent Placement

This realization is driving a structural change in computing.

Rather than sending every task to centralized cloud infrastructure, engineers are increasingly focused on running AI closer to where data originates. Specialized chips now allow inference to happen directly on devices, reducing latency, energy consumption and dependence on continuous connectivity.

This approach — often referred to as edge intelligence — represents a move from unlimited scale toward proportional power.

Instead of asking how much compute is possible, the industry is beginning to ask where compute is truly necessary.

Rethinking Intelligence Itself

The toddler in the museum offers more than a charming anecdote. It exposes a deeper truth about intelligence.

Children do not learn faster because they process more information. They learn because their brains organize experience into meaning. Curiosity guides attention. Feedback shapes interpretation. Context provides coherence.

Artificial intelligence excels at calculation. Humans excel at sense-making.

The future of AI will not emerge from choosing one over the other, but from understanding the difference.

Compute power remains one of humanity’s most powerful technological tools. Used wisely, it accelerates science, medicine and global coordination. Used indiscriminately, it becomes expensive noise.

The lesson may be simpler than expected: intelligence is not about having the biggest engine — but about knowing when to press the accelerator, and when to slow down.

Altair Media US — exploring technology, strategy and society at the frontier of innovation.

America’s AI Plumbing

How Broadcom, Palantir and Arista underpin America’s AI infrastructure

While headlines remain dominated by the “Magnificent Seven” a different group of American technology companies is quietly reshaping the foundations of the AI economy. These firms rarely feature in consumer narratives, yet their market capitalisation, strategic relevance and structural importance now rival — and in some cases surpass — far more visible names such as Tesla.

Broadcom, Palantir and Arista Networks operate largely behind the scenes. Together, they form what analysts increasingly describe as “AI plumbing”: the invisible hardware, software and network infrastructure required to make advanced AI systems actually work at scale. Without them, even the most powerful chips from Nvidia would remain underutilised.

This is a closer look at America’s silent AI powerhouses — and the vision of the leaders behind them.

Read More

Infineon and Ion Traps

Why Europe’s Quantum Future May Be Built on Industrial Discipline

Quantum computing is often presented as a race between exotic physics concepts and dazzling promises of exponential speed-ups. In practice, however, the decisive question is far more down to earth: which technologies can actually be engineered, manufactured and maintained at scale?

This is where Infineon Technologies enters the quantum conversation — not as a software visionary or a quantum algorithm startup, but as an industrial hardware company with decades of experience in controlling electrons, power flows and complex physical systems.

In recent years, Infineon has made a deliberate move into ion trap quantum computing, positioning itself at the intersection of precision electronics, photonics and atomic-scale control. The choice is telling. It reveals much about where quantum computing is heading — and about Europe’s potential role in turning quantum from laboratory science into infrastructure.

From Power Electronics to Atomic Control

Infineon is best known for its dominance in power semiconductors, automotive electronics and secure systems. These are not fast-moving consumer markets, but environments where reliability, reproducibility and long lifetimes matter more than raw performance.

At first glance, quantum computing may seem worlds apart from electric drivetrains or industrial inverters. Yet the underlying mindset is surprisingly similar. Ion trap systems demand extreme precision, stability and control over physical processes that are highly sensitive to noise and environmental variation.

In that sense, Infineon’s step from controlling electron flows to controlling individual ions is less a leap than a continuation — moving from classical precision engineering to its quantum equivalent.

Why Ion Traps?

Among the many competing quantum hardware approaches, ion traps occupy a distinctive position. Instead of fabricating artificial qubits, ion trap systems use individual atoms, suspended and controlled by electromagnetic fields. These atoms act as qubits with exceptional coherence and uniformity — nature provides perfect copies by default.

The advantages are compelling:

• long coherence times
• high-fidelity operations
• intrinsically identical qubits

But these strengths come with a price. Ion traps require:

• ultra-stable electromagnetic control
• extremely precise timing
• sophisticated laser systems for manipulation and read-out

This is not a domain for improvisation or artisanal lab setups. It is a system-engineering challenge — and precisely the kind of challenge where industrial semiconductor expertise becomes relevant.

Photonics as the Nervous System of Quantum Hardware

In ion trap quantum computers, photonics is not an auxiliary technology; it is the nervous system of the machine.

Lasers are used to:

• initialize qubit states
• manipulate quantum operations
• entangle ions
• read out results

Each of these steps requires light sources with extraordinary stability, precision and reproducibility. As systems scale from tens to hundreds or thousands of qubits, the problem shifts from “can we control a single ion?” to “can we control many ions in a repeatable, manufacturable way?”

This is where photonics transitions from experimental optics to industrial infrastructure. Integrating optical control with semiconductor electronics — and doing so reliably — is one of the central challenges of ion trap scaling.

Infineon’s interest here aligns closely with broader European strengths in photonics, semiconductor manufacturing and high-precision systems engineering.

The Quantinuum Partnership: Dividing the Quantum Stack

Infineon’s collaboration with Quantinuum illustrates a pragmatic division of labour within the quantum ecosystem.

Quantinuum focuses on:

• quantum system architecture
• software stacks and algorithms
• integrated quantum platforms

Infineon contributes:

• ion trap chip design
• control electronics
• manufacturing know-how and scalability

Rather than attempting to “own” the entire quantum stack, the partnership reflects an industrial logic: quantum computing will only mature if its hardware components can be produced, integrated and maintained within a real supply chain.

This is a subtle but important shift — away from demonstration systems toward industrial readiness.

From Laboratory Success to Industrial Reality

The hardest problems in quantum computing today are no longer purely quantum-mechanical. They are engineering problems.

How do you ensure consistent yields in ion trap fabrication?
How do you package delicate quantum structures for long-term operation?
How do you manage heat, noise and interference in increasingly dense systems?

These questions are familiar territory for semiconductor companies — even if the physics involved is new. Infineon’s quantum work is therefore less about chasing quantum advantage headlines and more about laying the groundwork for systems that can survive outside the lab.

What This Means for Europe

Europe is unlikely to dominate quantum computing through consumer platforms or hyperscale cloud services. But it holds deep strengths in:

• semiconductor manufacturing
• photonics
• precision engineering
• industrial systems integration

Ion trap quantum computing plays directly to those strengths.

Infineon’s strategy suggests that Europe’s opportunity may lie not in writing the most quantum algorithms, but in building the infrastructure that makes quantum computing viable, reliable and scalable.

In that sense, quantum computing begins to look less like a moonshot — and more like what Europe does best: turning complex physics into dependable technology.

At Altair Media, we will continue to follow how quantum technologies evolve as industrial systems — not just scientific breakthroughs.