Future Computing May Need New Physical Assumptions

Abstract

For decades, progress in computing has been driven by extraordinary advances in semiconductor manufacturing, device density, architecture, software, and systems engineering. That progression has produced computing capabilities that would have been difficult to imagine only a generation ago.

But continued advancement is placing increasing pressure on the physical systems beneath computation.

Energy must be delivered. Heat is generated. Information must move between increasingly dense components. Memory must remain sufficiently close and sufficiently fast. Materials must operate under increasingly demanding electrical and thermal conditions. And as artificial intelligence and other high-intensity computational workloads continue to expand, the infrastructure required to support computation can become nearly as consequential as the processors performing it.

These challenges are often approached as optimization problems within an established computing model. That approach may continue to produce important gains.

But another possibility deserves consideration:

What if some emerging limitations are not simply engineering problems inside the existing architecture, but signals that some of its underlying physical assumptions should themselves be reconsidered?

This paper considers how future computing may increasingly depend on new relationships among materials, energy, thermal behavior, interconnects, memory, and system structure.

Its central proposition is simple:

Future computing may require not only better components, but different assumptions about what a computing system physically is.

Keywords

Advanced computing; physical computing architecture; thermal behavior; energy efficiency; silicon carbide; graphene; silicon photonics; interconnects; memory wall; heterogeneous computing.

Method, Scope, and Limitations

This paper presents a forward-looking analytical framework drawing on peer-reviewed materials and device research, U.S. Department of Energy material on wide-bandgap power electronics, and current industry evidence from advanced computing infrastructure.

It examines a focused question: whether the physical foundations of advanced computing are becoming increasingly important to system design, and whether current developments already demonstrate a broader range of physical approaches.

Examples such as silicon carbide, graphene, in-memory computing, silicon photonics, high-density liquid-cooled infrastructure, and advanced electrical interconnects illustrate an expanding physical design space. They serve different functions, exist at different levels of maturity, and should not be treated as one technological category.

Future-computing research should also distinguish demonstrated performance from projected performance. Material properties, device-level results, prototype systems, and production infrastructure represent different levels of evidence and should not be treated as interchangeable.

1. The Architecture Beneath the Architecture

When people speak about computing architecture, they usually mean processors, memory, accelerators, software, instruction sets, or networks.

But beneath all of those systems is another architecture.

A physical one.

Electric charge moves through conductors. Transistors change state. Signals travel across interconnects. Materials absorb and release energy. Thermal energy accumulates and spreads. Components interact under electrical and thermal load. Packaging holds increasingly dense systems together. Power-delivery systems feed them. Supporting infrastructure exists because computation cannot be separated from the physical system in which it occurs.

Every computational act therefore exists inside a physical environment.

For much of computing history, extraordinary manufacturing progress allowed many of those physical constraints to remain secondary to the dominant story of rapidly rising computational capability.

That relationship is changing.

A 2026 Nature Reviews Materials perspective notes that modern systems consume energy not only through computation, but also through data movement between logic and memory, resistive losses in interconnects, and thermal-management demands; it further argues that as integration density rises, interfaces and material boundaries increasingly influence efficiency and scalability. [1]

As systems become more powerful, denser, and more energy-intensive, physical behavior increasingly becomes part of the architecture rather than merely a constraint surrounding it.

If a problem is confined to one component, the solution may be a better component. But when power, thermal behavior, memory movement, interconnects, packaging, and materials interact at system scale, the question becomes broader.

As these physical constraints increasingly interact, how should the architecture evolve alongside continued improvements in its individual components?

2. Computation Has Physical Costs

Information may be abstract.

Computation is not.

Every calculation performed by a physical computer requires physical events to occur. Charge moves. Transistors switch. Capacitances charge and discharge. Signals travel. Information moves between memory and processing elements.

Those events consume energy, and some of that energy is dissipated as heat.

At the microscopic level, electrical resistance reflects interactions experienced by charge carriers as they move through a material. In digital systems, additional energy is consumed through switching, leakage, signaling, and repeated movement of data. The resulting thermal burden is therefore not one single phenomenon, but the aggregate consequence of multiple physical processes.

As workloads intensify, those costs become more consequential.

More activity can require more energy. Greater density can concentrate thermal effects within smaller physical areas. More communication among components can increase both energy use and latency.

The deeper issue is not simply that a machine becomes warm. It is that the ability to perform additional computation becomes increasingly tied to the ability of the physical system to sustain that computation.

Heat is not merely a byproduct surrounding computation. At sufficient scale, it becomes one of the conditions governing how much computation is physically possible.

3. Optimization Within Established Architectures

Modern computing has advanced through extraordinary optimization.

Smaller transistors. Faster switching. Better lithography. Improved architectures. More efficient software. Advanced packaging. Faster memory. Specialized accelerators. Better interconnects.

Established computing architectures have been remarkably productive, and substantial progress continues within them.

At the same time, workloads now place pressure on power density, memory movement, interconnect bandwidth, thermal behavior, packaging, and physical scale simultaneously. That makes the design problem broader than whether any single device can be improved.

For many years, the processor could be treated as the center of the computing system while power, cooling, packaging, and networking were primarily enabling infrastructure.

That hierarchy is becoming less distinct.

When those surrounding systems increasingly determine sustained performance, efficiency, deployment density, reliability, and scale, they become performance-relevant parts of the computing architecture itself.

4. When Infrastructure Becomes Part of the Computer

In earlier eras, infrastructure often appeared separable from computation.

A computer computed. Power arrived. Cooling occurred around it. Networking connected it.

That separation becomes harder to maintain as system intensity grows.

Cooling is no longer merely an operational detail. Power delivery is no longer merely electrical housekeeping. Packaging is no longer only a protective layer around the chip. Interconnect strategy is no longer merely a pathway between components. Memory placement is no longer merely a convenience.

These have become performance-relevant variables.

Current AI infrastructure makes that shift unusually visible. NVIDIA states that its Rubin-generation systems are designed around fully liquid-cooled infrastructure and coolant temperatures up to 45°C, while its DSX reference design treats the broader AI-factory infrastructure stack as a coordinated system. [9]

The significance is not one particular cooling implementation. It is the architectural fact that thermal and facility conditions are now important enough to shape the computing platform itself.

Advanced computing is therefore increasingly understood less as an isolated processor and more as an integrated physical system.

5. Materials Expand the Physical Design Space

For most of modern computing history, silicon has provided an extraordinarily successful foundation.

At the same time, research and industry increasingly use materials with physical properties suited to different electrical, thermal, optical, and power functions.

One example is silicon carbide.

Silicon carbide is a wide-bandgap semiconductor used principally in power electronics rather than as a general-purpose logic material. U.S. Department of Energy research describes silicon-carbide and gallium-nitride devices as capable of operating at higher temperatures, voltages, and switching frequencies, with the potential for higher power-conversion efficiency than conventional silicon power-electronics components. [2]

Another example is graphene.

Graphene has attracted sustained research interest because of unusual electronic behavior and exceptionally high thermal conductivity. Peer-reviewed work has examined its potential in electronic devices and thermal-management applications, while more recent research continues to investigate graphene-based structures for next-generation nanoelectronics. [3][4]

These materials do not perform the same job.

That is precisely the point.

Advanced computing increasingly draws on different materials according to the physical functions they perform best.

A material may be valuable because it supports efficient power conversion. Another may be valuable because it operates under demanding thermal conditions. Another may be useful because of optical or electronic behavior that enables a different device class.

The question may increasingly be which combination of physical properties a computing system requires.

6. The Cost of Moving Information

Modern computing is not limited only by the act of computation itself.

It is also affected by the cost of moving information.

Data must travel between processors and memory, among accelerators, between chips and boards, and across larger systems.

That movement consumes energy and takes time.

The separation of memory and processing in conventional computing has long been associated with the memory wall. Peer-reviewed research on in-memory computing describes repeated data transfer between memory and processors as both energy-intensive and time-consuming, and more recent work continues to identify data movement as a major efficiency constraint in data-intensive workloads. [5][6]

This raises a broader architectural question:

How much of future computing performance will depend not simply on processing faster, but on reducing the physical burden of moving information?

That question is already influencing real systems.

Memory and processing are being brought into new relationships. Near-memory and in-memory approaches are being developed to reduce repeated data transfer. Optical interconnects are moving into production infrastructure. Electrical interconnects are simultaneously being pushed to high levels of sophistication.

The implication is not that one physical method has won.

It is that information movement has become too important to remain a secondary design consideration.

7. Efficiency Under Increasing Physical Demands

As computational demand rises, the challenge is not simply to build larger systems.

It is to increase useful computation without allowing the physical burden surrounding that computation to grow without restraint.

That means reducing unnecessary energy dissipation. It means improving power conversion. It means reducing avoidable data movement. It means using materials and devices that remain effective under increasingly demanding operating conditions. It means treating efficiency as a system-wide property rather than only a device-level metric.

This is where research directions that once appeared specialized begin to converge around a common problem.

Wide-bandgap power electronics.

Advanced packaging.

New memory relationships.

Optical networking.

High-efficiency electrical interconnects.

Material and interface engineering.

These approaches differ substantially, but each addresses some portion of the physical cost of scaling computation.

The next stage of computing may therefore be shaped not only by increasing raw computational capability, but by reducing how much physical burden must accompany each increase in capability.

8. Computing Is Becoming More Physically Diverse

Future computing is already being shaped by a broader range of materials, interconnect technologies, power systems, memory architectures, packaging approaches, and physical mechanisms.

Silicon carbide is expanding the operating possibilities of power electronics. Graphene and other advanced materials continue to be investigated for properties not available in conventional material systems. Memory and processing are being brought into new relationships. Optical interconnects are moving into advanced computing infrastructure. Electrical interconnects are simultaneously being pushed to increasingly sophisticated limits.

NVIDIA's current networking portfolio illustrates this diversity within a single computing ecosystem. Its Spectrum-X platform includes both a power-efficient copper-based cable cartridge for rack-scale connectivity and silicon-photonics-based co-packaged optical networking for larger-scale systems. [8]

Its silicon-photonics systems integrate optics directly with networking ASICs and are explicitly designed to reduce network power while scaling AI infrastructure. [7]

These developments are not competing predictions about one future technology.

They are evidence that computing itself is becoming more physically diverse.

The important question is no longer whether advanced computing will explore different physical approaches. It already is. The question is what those approaches will ultimately make possible.

9. Physical Constraints as First-Class Design Variables

As computational systems become denser and more demanding, physical constraints increasingly need to enter the architectural conversation at the same level as computational goals.

Power availability matters before a system is deployed.

Thermal conditions matter before sustained performance can be assumed.

Memory placement and information movement matter before bandwidth requirements can be treated as abstract numbers.

Packaging, material behavior, electrical characteristics, and optical pathways can all affect the practical limits of a system.

Treating these as first-class design variables does not prescribe one architecture.

It means that advanced computing design increasingly has to ask physical questions early:

What operating conditions can be sustained?

What energy costs are acceptable?

What information movement is necessary?

Which material and interconnect properties fit the intended workload?

Which constraints become dominant as density and scale increase?

As computation approaches physical constraints, physics increasingly becomes part of the computing architecture itself.

10. Evidence Before Adoption

New physical approaches should not be accepted because they are novel.

They should be evaluated because they can be tested.

Simulation.

Comparison.

Materials characterization.

Controlled testing.

Fabrication.

Experimental validation.

Reproducibility.

A serious future-computing research program must distinguish possibility from proof.

A material property is not a complete computer architecture.

A device result is not a production system.

A simulation is not independent experimental validation.

A prototype is not automatically scalable manufacturing.

The path from a physical idea to deployable computing infrastructure requires disciplined evidence at each stage.

That is not an argument against ambitious research.

It is the condition that allows ambitious research to become engineering.

11. The Physical Architecture of Future Computing

The developments examined here point in a common direction.

Advanced computing is becoming more dependent on the interaction among computation, energy, materials, memory, interconnects, packaging, and infrastructure.

That does not imply one universal future architecture.

It suggests a broader physical design space in which different mechanisms can be used where their properties make sense.

Some advances will come from devices. Others will come from memory organization, power systems, interconnects, packaging, materials, or infrastructure. Many will emerge from the way those elements are combined.

The significance is that the physical system beneath computation is becoming increasingly consequential to what computation can achieve.

The future of computing is therefore not only a question of what machines will calculate. It is also a question of what physical systems will make that computation practical.

Conclusion

Computing has advanced through relentless improvement of an extraordinarily successful technological foundation.

At the same time, the growing demands of advanced computation are making its physical foundations increasingly visible.

Energy matters.

Thermal behavior matters.

Materials matter.

Information movement matters.

Power delivery matters.

Packaging matters.

Infrastructure matters.

These are not peripheral issues. They increasingly shape what levels of sustained computation are physically and economically practical.

And the industry is already responding through a wider range of physical approaches: wide-bandgap power devices, advanced materials, memory-centered architectures, optical interconnects, increasingly sophisticated electrical connectivity, advanced packaging, and thermal infrastructure designed as part of the computing platform.

The future of computing will not be defined by one of these developments alone.

But together they indicate that the physical design space of computing is expanding.

Future computing may require not only better components, but different assumptions about what a computing system physically is.

If that is true, then some of the most important advances in computing may emerge not only from what we ask machines to calculate, but from how we choose to physically build the systems that make those calculations possible.

References

1. Nature Reviews Materials. “Engineered interfaces in electronic materials for energy-efficient computing.” 2026. https://www.nature.com/articles/s41578-026-00949-9
2. U.S. Department of Energy. “Can You Dig It? Heavy-Duty Hybrid Electric Vehicles Benefit from Lab’s Wide-Bandgap Expertise.” Discussion of silicon-carbide and gallium-nitride power electronics. https://www.energy.gov/cmei/articles/can-you-dig-it-heavy-duty-hybrid-electric-vehicles-benefit-labs-wide-bandgap
3. Balandin, A. A. “Thermal properties of graphene and nanostructured carbon materials.” Nature Materials 10, 569–581 (2011). https://www.nature.com/articles/nmat3064
4. Zhang, J. et al. “Bottom–up-synthesized graphene nanoribbons for nanoelectronics.” Nature Reviews Materials 11, 194–212 (2026). https://www.nature.com/articles/s41578-025-00880-5
5. Ielmini, D. and Wong, H.-S. P. “In-memory computing with resistive switching devices.” Nature Electronics 1, 333–343 (2018). https://www.nature.com/articles/s41928-018-0092-2
6. Sebastian, A., Le Gallo, M., Khaddam-Aljameh, R., and Eleftheriou, E. “Memory devices and applications for in-memory computing.” Nature Nanotechnology 15, 529–544 (2020). https://www.nature.com/articles/s41565-020-0655-z
7. NVIDIA. “Silicon Photonics Networking for Agentic AI.” Current product material describing co-packaged silicon photonics integrated with networking ASICs. https://www.nvidia.com/en-us/networking/products/silicon-photonics/
8. NVIDIA. “Spectrum-X Ethernet Platform for AI Networking.” Current product material describing copper-based rack-scale connectivity and optical scale-out. https://www.nvidia.com/en-us/networking/spectrumx/
9. NVIDIA. “Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI’s Biggest Machines.” June 21, 2026. https://blogs.nvidia.com/blog/liquid-cooling-ai-factories/