When AI Moves From Software Into the Physical World

Artificial intelligence has taken technology to a different level. Even though the first recognizable breakthrough of AI occurred with Alan Turing in the 1950s with his paper “Computing Machinery and Intelligence”, asking for the first time whether machines could think, it is only in recent years that AI has become part of common discourse. From a discipline contained within computer science departments or software companies, it is now becoming the operating system of physical reality, embedded in sensors present in machines and infrastructure that perceive, decide and act in the real world.

Artificial intelligence has spent the last decade becoming weightless. AI progress was measured by what models could do in software: classify images, generate text, win at games (remember the time Alpha Go defeated the world champion of a highly complex, creative game called Go?). As AI moves into machines, the decisive question in 2026 is whether intelligence can function reliably in constrained physical environments.

This marks a fundamental convergence between AI and next-generation computing. Physical AI cannot be separated from the hardware, sensors, chips, and edge infrastructure that allow it to function in real time. A robot on a factory floor, an autonomous drone, a medical device, or a brain-computer interface does not simply “use AI”; it depends on a full computing stack that can perceive, decide, and act locally, safely, and continuously. In that sense, the next generation of AI companies will be defined by their ability to make intelligence operational in constrained physical environments where latency, energy, safety, and reliability matter as much as accuracy.

What experts are telling founders in 2026 is therefore different from what they were saying two years ago: “do not build a model, build the system around the model”. The most compelling ventures are those that understand edge deployment, telemetry, fail-safes, privacy-by-design, and regulatory accountability from day one. This is also where Europe can turn regulation into strategy. The EU AI Act, Digital Services Act (DSA), and Data Act are not merely compliance frameworks; for Physical AI ventures, they can become a trust architecture that travels well beyond Europe. In markets where industrial customers, hospitals, governments, and infrastructure operators need proof that AI systems are safe, auditable, sovereign, and privacy-preserving, European regulatory fluency becomes a commercial advantage rather than a constraint.

Alumni companies of the deep tech program Creative Destruction Lab AI stream, such as Cogitat and Lymbic AI  illustrate this shift. Cogitat, which decodes brain activity for movement control, points toward a world in which human cognition and machine action become more tightly coupled, while Lymbic AI approaches the same biological frontier from the security side. Its platform generates neuroprints, unique biometric signatures derived from brainwave patterns captured by EEG headsets, enabling authentication that cannot be stolen, replicated, or phished. Where Cogitat decodes intention, Lymbic verifies identity and together, they outline what a physical AI stack that is genuinely human-centred begins to look like: intelligence that reads the body, responds to its signals, and does so in a way that is both clinically useful and cryptographically accountable. The same underlying neurotechnology that enables these applications is also informing how researchers think about biological interfaces in drug discovery and synthetic biology.

The physical foundation AI requires

When intelligence becomes embedded in every machine, factory, grid, hospital, and defence system, the question of who controls the infrastructure that intelligence runs on becomes inseparable from industrial competitiveness and strategic autonomy. A continent that processes sensitive industrial data through external cloud infrastructure, raw materials to build the backbone of the technology it requires, has outsourced more than computation, because it has outsourced part of the operating layer of its economy. The implications of that dependency become harder to reverse as the systems built on top of it grow more and more complex. 

Physical AI at scale runs directly into a hard constraint: the infrastructure that has powered the current wave of AI progress was designed for centralized data centres running at scale on copper interconnects, and it was not built for the distributed, energy-constrained, latency-critical demands that Physical AI places on it. Next Gen Computing should not be considered a parallel track alongside AI, it is the physical foundation AI requires to fulfil its potential, although the relationship needs to be treated carefully. Physical AI creates demand for better computing infrastructure, but it does not automatically define the first market for every Next Gen Computing venture.

The more precise 2026 thesis is that AI is exposing the limits of today’s computing architecture while new computing substrates are searching for commercially urgent use cases. Some ventures will serve edge AI directly through sensors, low-power computing, encryption, data transmission, or photonic interconnects, while others may find their beachhead markets in semiconductor packaging, secure communications, advanced sensing, quantum software, materials discovery, or industrial optimization. The question is not whether AI and Next Gen Computing are connected in theory, but where customer demand is strong enough to pull the technology into the market.

Photonics is an example that illustrates this distinction quite clearly. The bottleneck used to sit heavily in process manufacturing, but that is beginning to change as larger foundries scale production for higher volumes. The harder question is now increasingly applicative: whether a venture can rely on a manufacturing process already available at scale, and whether its use case is aligned with a market need urgent enough to justify adoption. LumiSync has patented the world's first fully photonic oscillator, achieving performance one thousand times faster and one thousand times more energy-efficient than current electronic equivalents, progressing through both HEC Challenge+ and the CDL Paris Next Gen Computing stream. Moon Photonics, Lightspring, and Flatlight have all advanced through the same programme, supported by LuminEdge, the incubation platform co-founded by iXcampus and the Institut d'Optique Graduate School with HEC Paris as a partner, and provides mutualized infrastructure access, such as cryostats, cleanrooms, or advanced metrology equipment. 

Quantum computing will face a version of the same commercialization discipline as it matures. Scientific promise is not sufficient: the question is where the use case is precise enough, and the customer's pain acute enough, to justify adoption before the technology reaches full maturity. Quantum algorithms and software solutions are already advancing around specific applications in chemistry, materials discovery, logistics, and industrial optimization, and ventures such as Quantistry illustrate why this targeting matters. The strongest companies in this space are likely to be those that translate technical advantage into a problem customers can understand, test, and eventually procure, rather than those that lead with the architecture and hope the market catches up.

Underlying all of this is a political and industrial reality that founders in the Next Gen Computing space cannot afford to treat as background. The semiconductor industry has been built around CMOS (Complementary Metal-Oxide-Semiconductor) economics over decades, and some of its most powerful incumbents may prefer to slow, absorb, or redirect competing technologies rather than reorganize their strategies around them. Technical superiority has never been sufficient on its own: founders must also understand the incumbent incentive structures, the supply chain dependencies, the qualification standards required for medical, aerospace, and defence markets, and the true cost of asking industrial customers to change an architecture they have spent years optimizing. Europe's regulatory and industrial policy environment, including the Chips Act and France 2030's investment in photonics and quantum infrastructure, is attempting to reshape some of these incentive structures, but the window it creates will not remain open forever.

Physical AI brings the AI and Next Gen Computing argument back to a more fundamental ecosystem challenge. AI ventures are deploying edge intelligence in factories, hospitals, robots, defence systems, or critical infrastructure that need better hardware, sensing, energy efficiency, privacy, and certification far sooner than many Next Gen Computing technologies can naturally reach commercial maturity. The answer is not to force photonics, quantum, or advanced semiconductor ventures into premature markets, but to break the silos that keep AI founders, hardware founders, industrial customers, foundries, regulators, and investors working on different clocks. Physical AI can become the place where these clocks begin to align, because it gives Next Gen Computing founders a clearer view of real deployment pain points, while giving AI founders access to the infrastructure constraints that will define whether their systems can scale. The role of a deep tech ecosystem is precisely to create those encounters early: to help founders test use cases before the market is fully formed, design for the standards customers will eventually require, and build companies around the industrial bottlenecks that software alone cannot remove. 

Authors: Livia Kalossaka, Alex Charbonné, Abdelalim El Hamichi