Discussions of superintelligence — AI systems that substantially outperform humans across a wide range of cognitive tasks — tend to focus on algorithms, data, and training methods. The implicit assumption is that hardware will keep up: that faster, more efficient chips will be available when they are needed, and that scaling laws will continue to hold because compute will continue to get cheaper. That assumption deserves scrutiny. Progress in AI has always been gated by hardware, from the GPUs that enabled deep learning to the tensor accelerators that enabled large language models. If hardware progress slows, AI progress slows with it, no matter how clever the algorithms.

This article argues that hardware is a bottleneck for superintelligence, not a solved problem, and that light-matter computing — including excitonic and polariton devices — is one of several attempts to move that bottleneck. It does not claim that excitonic hardware will enable superintelligence, or that superintelligence is imminent. It claims that the hardware question is central, that the current trajectory faces physical and economic limits, and that new computing paradigms deserve serious attention precisely because the incumbent trajectory is uncertain.

What superintelligence would require from hardware

Superintelligence, if it is achieved, will not be a single model running on a single chip. It will be a system: many models, many agents, vast memory, and extensive interaction with the world through sensors and actuators. It will need to learn continuously, to reason over long horizons, to plan, and to act. Each of those capabilities has hardware implications.

Training: Training a frontier model today requires massive compute, often described as requiring thousands of accelerators running for weeks or months, consuming large amounts of energy. Training a substantially more capable system — with more parameters, more data, more modalities, and more sophisticated training objectives — would require still more compute, unless algorithms become dramatically more efficient. The energy and cost of training are already constraints; they will become more binding as models grow.

Inference: If superintelligence is to be widely deployed, inference must be efficient. A system that can reason at human level but costs orders of magnitude more per query than a human will be limited to niche applications. Inference energy, latency, and throughput will determine whether advanced AI can be deployed at scale, in robots, in vehicles, in scientific instruments, and in everyday devices.

Memory: Intelligence requires memory — not just parameters, but working memory, episodic memory, and long-term knowledge. Current AI systems are limited by memory bandwidth: the rate at which data can be moved between memory and compute. The von Neumann bottleneck — the separation of memory and processing — is a fundamental constraint in conventional architectures. Systems that can integrate memory and compute more tightly, or that can move data with less energy, will have an advantage.

Communication: A superintelligent system will likely be distributed: many chips, many nodes, communicating over networks. The bandwidth and latency of those interconnects, and the energy per bit communicated, will limit how effectively a distributed system can coordinate. Optical interconnects already play a role here, and photonic computing could extend that role from communication to computation.

Adaptation: A system that must learn continuously, rather than being trained once and frozen, needs hardware that can update its parameters efficiently and that can handle the plasticity of learning without catastrophic forgetting. Current hardware is optimized for static inference and for batched training, not for continuous, online learning.

These requirements are not just about raw speed. They are about energy, about data movement, about memory integration, and about the ability to learn and adapt. They are the dimensions on which the hardware bottleneck will be felt.

Why the incumbent trajectory is uncertain

The conventional response to hardware demands is to scale the incumbent: build larger chips, pack more transistors, use more advanced process nodes, and connect more accelerators. This approach has worked for decades, but it faces growing challenges.

Moore’s law — the observation that transistor density doubles roughly every two years — has slowed. The cost per transistor is no longer falling as rapidly as it once did, and the engineering challenges of smaller nodes are increasing. Dennard scaling — the principle that smaller transistors use proportionally less power — ended years ago, which is why chips now face power density limits and why dark silicon — the fraction of a chip that must be powered off to stay within thermal limits — is a concern.

The energy cost of AI is rising. Training a large model consumes energy on a scale that is significant for a datacenter and, in aggregate, for the grid. Inference at scale, multiplied across billions of queries, also consumes substantial energy. The energy trajectory of AI is not sustainable indefinitely on the current hardware path without improvements in efficiency, in energy supply, or in both. This is not a claim that AI will run out of energy tomorrow; it is a claim that energy is a binding constraint that will shape what is feasible.

Memory bandwidth is also a bottleneck. The gap between compute throughput and memory bandwidth has widened over time. Accelerators can perform arithmetic faster than they can fetch the data to operate on. This is why much of the energy in AI hardware is spent moving data, not computing on it. Architectural innovations — such as larger on-chip memory, 3D stacking, and near-memory computing — help, but they do not eliminate the bottleneck.

These challenges do not mean that conventional hardware progress has stopped. GPUs and tensor accelerators continue to improve, through architectural innovation, through advanced packaging, and through specialization. But the rate of improvement, and the cost of achieving it, are changing. The assumption that hardware will automatically keep up with algorithmic ambition is less safe than it once was.

The case for new computing paradigms

When an incumbent trajectory faces physical limits, the field explores alternatives. For AI hardware, several paradigms are being explored in parallel.

Neuromorphic computing attempts to borrow the brain’s event-driven, sparse, memory-integrated architecture. Quantum computing attempts to exploit quantum parallelism for specific problems. Analog computing attempts to perform computation in the analog domain, where a single physical operation can replace many digital ones. In-memory computing attempts to perform computation where data is stored, reducing data movement. Photonic computing attempts to use light for computation, exploiting its speed, parallelism, and low-loss propagation.

Light-matter computing — including excitonic and polariton devices — sits within the photonic paradigm but adds a distinct capability: strong, fast nonlinearity from matter. As discussed in earlier articles in this series, photons are excellent for linear operations but poor for nonlinear ones. Excitons and polaritons offer a way to add nonlinearity without sacrificing the speed and parallelism of photonics.

The case for exploring these paradigms is not that any one of them is guaranteed to succeed. It is that the incumbent trajectory is uncertain enough that alternatives are worth pursuing, and that the potential payoff — a hardware substrate that is significantly more energy-efficient, faster, or more scalable for AI — is large enough to justify the investment.

What light-matter computing could offer superintelligence

If light-matter computing matures, what would it offer for the path to superintelligence? Several possibilities are worth considering, with appropriate caution.

Energy efficiency: If photonic and excitonic accelerators can perform matrix multiplication and nonlinear activation with lower system-level energy per operation than electronics, they could reduce the energy cost of training and inference. This would not by itself enable superintelligence, but it would expand the feasible scale of models and deployments. Energy efficiency is a multiplier: it makes every other advance more affordable.

Speed and latency: Photonic propagation is fast, and polariton dynamics are fast. A system that can perform inference with lower latency could enable real-time applications — robotics, autonomous systems, interactive agents — that are difficult with high-latency hardware. Speed also matters for training, where faster iteration enables more rapid experimentation.

Bandwidth and interconnects: Photonic systems naturally handle high-bandwidth communication. A computing system that integrates computation and communication in the optical domain could reduce the cost of moving data between chips and between nodes, which is a major bottleneck for distributed AI.

Memory integration: Some photonic and excitonic proposals integrate memory and computation more tightly than conventional architectures, for example by using phase-change materials or polariton lattices that store and process information in the same physical structure. Tighter integration could reduce data movement energy, which is a dominant cost.

New computational primitives: Beyond accelerating existing neural networks, light-matter systems could enable new primitives — such as reservoir computing, oscillator-based computing, or quantum simulation — that are difficult to implement efficiently in electronics. These primitives might support forms of computation that are not well captured by current AI models, such as continuous-time dynamics, synchronization, or collective behavior.

Each of these possibilities is speculative. None has been demonstrated at a scale relevant to superintelligence. They are reasons to explore, not reasons to conclude.

What light-matter computing cannot do — yet

It is equally important to be clear about what light-matter computing cannot do.

It cannot replace the software and algorithmic advances that superintelligence will require. Hardware is a bottleneck, but it is not the only bottleneck. Algorithms for reasoning, planning, learning, and alignment — the problem of ensuring that an AI system pursues the goals its creators intend — are equally important and equally unsolved. A faster chip does not solve alignment.

It cannot solve the data problem. Training advanced AI requires vast, high-quality data, and the availability of such data is a constraint. Hardware cannot create data that does not exist, though it can enable more efficient use of data through better training methods.

It cannot solve the manufacturing problem overnight. As discussed in the datacenter article in this series, turning a laboratory polariton device into a manufacturable chip is a long, difficult process. Superintelligence will not be enabled by a single device demonstration; it will require a mature, reliable, economical hardware ecosystem.

It cannot guarantee safety. A more powerful hardware substrate, if it enables more powerful AI, also increases the stakes of getting alignment and governance right. Hardware progress without corresponding progress in safety, evaluation, and oversight could increase risk. The hardware discussion should not be separated from the safety discussion.

These caveats are not arguments against researching light-matter computing. They are arguments for keeping hardware in perspective: it is a necessary condition for superintelligence, not a sufficient one.

The bottleneck framing: useful but incomplete

Framing hardware as “the bottleneck” for superintelligence is useful because it highlights a constraint that is often overlooked. But it can also be misleading if it suggests that removing that bottleneck would automatically lead to superintelligence. In reality, there are multiple bottlenecks — hardware, algorithms, data, safety, energy, and governance — and progress requires advancing on all of them.

A more accurate framing is that hardware is one of several coupled bottlenecks. Advances in hardware enable advances in algorithms, by allowing researchers to test larger models and more ambitious ideas. Advances in algorithms reduce hardware requirements, by making models more efficient. The two co-evolve. The history of AI is a history of this co-evolution: GPUs enabled deep learning, deep learning created demand for better GPUs, and the cycle continued.

Light-matter computing should be understood in this co-evolutionary context. It is not a standalone solution to superintelligence. It is a candidate for the next step in the hardware-algorithm cycle — a substrate that, if it matures, could enable new algorithms and new scales of computation that are difficult on current hardware.

What to watch

For readers interested in the intersection of superintelligence and hardware, several signals are worth tracking.

First, system-level energy and performance of photonic and excitonic accelerators on real AI workloads, measured transparently. Device-level demonstrations are not sufficient; the question is whether a system can beat the incumbent on energy per inference, latency, and accuracy for a relevant task.

Second, manufacturing progress: whether any excitonic or photonic platform achieves wafer-scale reproducibility, high yield, and integration with silicon photonics. Without manufacturing, the technology will remain a laboratory curiosity.

Third, algorithmic co-design: whether researchers develop algorithms and training methods that are specifically designed for photonic and excitonic hardware, rather than simply porting electronic algorithms. Hardware that requires new algorithms will succeed only if those algorithms are developed and adopted.

Fourth, energy and infrastructure: whether the energy supply and datacenter infrastructure can support the scale of computation that superintelligence would require, on any hardware substrate. The energy question is not just about chips; it is about the grid, cooling, and the economics of large-scale deployment.

Fifth, safety and governance: whether progress in AI capability is matched by progress in evaluation, alignment, and oversight. Hardware that enables more capable AI without corresponding safety progress increases risk.

These signals are not just about excitonic computing. They are about the broader question of whether humanity can build the hardware, the algorithms, and the institutions needed to develop advanced AI safely and beneficially.

Conclusion

Superintelligence is often discussed as if it were purely a software problem. It is not. Hardware is a binding constraint — on training, on inference, on memory, on communication, and on energy. The incumbent hardware trajectory faces physical and economic limits that make its continued progress uncertain. New paradigms, including light-matter computing, are worth exploring because they offer a different set of tradeoffs: speed, parallelism, and potentially lower energy, with strong nonlinearity from excitons and polaritons.

Light-matter computing will not by itself produce superintelligence. It will not solve alignment, data, or governance. And it faces formidable engineering challenges of its own. But it is a serious attempt to address a real bottleneck, and it deserves to be evaluated on system-level evidence: energy, latency, accuracy, manufacturability, and cost.

The hardware bottleneck is not the whole story of superintelligence, but it is a chapter that cannot be skipped. Whether light-matter computing helps to move it will be decided not by speculation but by the hard work of building devices, integrating them into systems, and measuring them honestly.

Sources and further reading

This article synthesizes literature on AI hardware, on the limits of conventional scaling, and on photonic and excitonic computing. For AI hardware and its energy and scaling challenges, see reviews in Nature Electronics, Nature Reviews Electrical Engineering, and Proceedings of the IEEE, and perspective articles on AI energy in journals such as Nature and Joule. For Moore’s law, Dennard scaling, and the memory bottleneck, see review articles on semiconductor scaling and computer architecture. For photonic and excitonic computing, see reviews in Nature Photonics and Nature Reviews Physics, as cited in earlier articles in this series. For superintelligence, alignment, and AI safety, see the scholarly and policy literature on AI alignment and governance, including work published by AI safety research organizations and in journals covering AI policy. No specific timelines, energy figures, or capability forecasts are asserted here, because those depend on uncertain algorithmic, hardware, and economic developments; readers should treat any such forecasts in other sources with skepticism and should verify the assumptions behind them.