🤖 machine learning

Physical Foundation Models: Fixed hardware implementations of large-scale neural networks

This paper proposes "Physical Foundation Models" (PFMs), a paradigm where large-scale neural networks are directly realized as fixed hardware implementations leveraging natural physical dynamics to achieve unprecedented energy efficiency, speed, and parameter density for trillion-scale AI models.

Original authors: Logan G Wright, Tianyu Wang, Tatsuhiro Onodera, Peter L. McMahon

Published 2026-05-01

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Logan G Wright, Tianyu Wang, Tatsuhiro Onodera, Peter L. McMahon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: From a Swiss Army Knife to a Custom Tool

Imagine you have a Swiss Army Knife. It's incredibly versatile; it can cut, screw, open bottles, and saw. But because it has to do everything, it's not the absolute best at any single task. It's also heavy and takes up space in your pocket.

Current AI models (like the ones powering ChatGPT or Gemini) are like that Swiss Army Knife. They are massive, general-purpose "Foundation Models" trained to do almost anything. To run them, we use powerful, programmable computer chips (GPUs) that act like a workshop full of tools. These chips are flexible, but they are also energy-hungry and slow because they have to constantly fetch instructions and data, moving them back and forth like a chef running between the fridge and the stove.

The authors propose a radical new idea: Instead of building a flexible workshop, why not build a custom-made, single-purpose machine for every specific AI model?

They call these Physical Foundation Models (PFMs).

The Analogy: The "Frozen" Sculpture

Think of a standard computer chip as a clay sculpture. You can mold it, reshape it, and change it whenever you want. This is great for flexibility, but it takes time and effort to reshape it every time you want to do something new.

The authors suggest we stop using clay and start using frozen glass.

The Old Way: We train an AI model in software, then try to make a chip that can simulate that model. The chip has to be programmed to act like the model.
The New Way (PFM): We take the final, trained AI model and literally carve it into the hardware itself before the chip is even made. The "brain" of the AI isn't a set of instructions; it is the physical shape of the material.

Once the chip is manufactured, the AI is "hard-wired" into the physics of the device. You can't change the model without melting the glass and making a new one. But because the model is the hardware, it doesn't need to "think" or "fetch" data. It just lets the input flow through, and the physics of the material naturally produces the answer.

How It Works: The Light Pipe

The paper uses light as a primary example. Imagine a long, clear tube made of special glass.

The Input: You shine a pattern of light into one end of the tube.
The "Brain": Inside the tube, the glass isn't smooth. It has tiny, microscopic bumps and curves (nanoscale structures) that were carved exactly to match a specific AI model.
The Process: As the light travels through the tube, it bounces, bends, and interferes with itself based on those tiny bumps. This isn't a computer calculating numbers; it's light doing what light naturally does.
The Output: When the light hits the other end, the pattern has changed. That new pattern is the answer to your question.

Because the light is moving at the speed of light and the "calculation" happens just by the light traveling, this process is incredibly fast and uses almost no electricity compared to a digital computer.

Why Do This? (The Three Big Wins)

The authors argue that if we can build these "frozen" machines, we get three massive benefits:

Energy Efficiency: A digital computer is like a person running back and forth to a library to read a book, write a note, and run back. A PFM is like having the book printed directly on the wall in front of you. You just look at it. This saves a huge amount of energy.
Speed: Since the "computation" is just the physical movement of light or electrons through a material, it happens almost instantly, limited only by the speed of the signal, not by the speed of a processor.
Size and Scale: Digital computers need a lot of space to store data and move it around. PFMs can pack a massive amount of "brainpower" into a tiny space because the data is stored in the 3D structure of the material itself.

The "Forever" Problem

There is a catch, which the paper openly admits. Because these machines are "frozen" into the hardware, they cannot learn new things.

If you want to update the AI to know about a new event that happened yesterday, you can't just download a software update. You have to throw away the old glass tube and manufacture a brand new one with the new design carved into it.

The authors suggest this is actually okay because:

Foundation models (like GPT-5 or Llama 4) are updated only once a year or so.
We could build factories that churn out these "AI tubes" on a yearly schedule, just like we release new phone models.
For the specific tasks they are built for, they will be vastly superior to anything we have today.

The Future Vision

The paper imagines a future where:

Data Centers use these massive, fixed hardware tubes to run AI for the whole world, using a fraction of the current electricity.
Edge Devices (like your phone, a robot, or a car) could have their own tiny, custom AI tubes built right in, allowing them to be incredibly smart without needing to connect to the internet.
Huge Models: We might be able to build AI models with trillions or even quadrillions of parameters (connections), which are currently impossible to run on digital computers because they would require too much power and memory.

Summary

The paper proposes a shift from programmable computers (flexible but inefficient) to physical machines (rigid but incredibly efficient). By carving the AI directly into the physics of materials like glass or silicon, we could create "Physical Foundation Models" that are faster, smaller, and use less energy than anything we have today, potentially unlocking the ability to build AI systems of a scale we can barely imagine right now.

1. Problem Statement

The paper addresses the escalating computational, energy, and capital costs associated with training and deploying large-scale AI foundation models (e.g., GPT-4, Gemini, Llama 4).

Energy and Cost Crisis: Current datacenter AI workloads consume ~44 GW of power, projected to exceed 150 GW by 2030, with capital costs exceeding $5 trillion.
The Von Neumann Bottleneck: Traditional digital architectures separate memory and processing, leading to massive energy and latency penalties due to data movement (moving weights and activations).
Scalability Limits: Current digital hardware struggles to support models significantly larger than $10^{12}$ parameters. A hypothetical $10^{18}$ -parameter model would require exabytes of memory and decades to read via current memory bandwidths, making inference on separated memory architectures impractical.
Edge Constraints: High energy demands prevent deploying large foundation models on power-constrained edge devices (robots, sensors, mobile).

2. Methodology and Core Concept

The authors propose Physical Foundation Models (PFMs), a radical rethinking of AI hardware implementation.

Definition: PFMs are fixed, single-purpose hardware devices where the neural network architecture and parameters are "hard-wired" directly into the physical substrate (e.g., nanostructured glass, semiconductors) at the time of fabrication.
Key Distinction: Unlike programmable digital chips (GPUs) or even fixed-weight digital ASICs (like Taalas's HC1), PFMs perform computation using the natural physical dynamics of the hardware (e.g., light propagation, electron flow) rather than simulating mathematical operations via digital logic.
Philosophy: The approach merges Feynman's "plenty of room at the bottom" (nanoscale fabrication) with Leiserson's "plenty of room at the top" (eliminating abstraction layers). It treats the hardware itself as the software, optimizing the physical system directly for the inference task.
Trade-off: PFMs sacrifice programmability (the ability to update weights post-fabrication) to achieve massive gains in parameter density, energy efficiency, and speed. They are designed to be manufactured on a ~1-year cadence, matching the release cycle of new foundation models.

3. Key Contributions

Conceptual Framework: Introduction of the PFM paradigm, arguing that for foundation models (which are static after training), fixed hardware is the optimal solution.
Back-of-the-Envelope Calculations: The authors perform scaling analysis for a specific Optical PFM (a 3D nanostructured glass medium) and compare it against state-of-the-art digital GPUs (e.g., NVIDIA B200 class).
Platform Exploration: Discussion of potential realizations beyond optics, including nanoelectronics (crossbar arrays, inverse-designed resistive wires, semiconductors) and mechanics.
Research Roadmap: Identification of critical challenges required to realize PFMs, including large-scale physical inverse design, training strategies for fixed hardware, and handling fabrication variability.

4. Results and Performance Estimates

Using a 3D optical PFM example (nanostructured glass with refractive index variations encoding parameters), the authors estimate performance for models ranging from $10^{12}$ to $10^{18}$ parameters:

Metric	Optical PFM ( $10^{12}$ params)	Optical PFM ( $10^{18}$ params)	Digital GPU Baseline ( $10^{12}$ params)
Volume	$\sim (1 \text{ mm})^2 \times 1 \text{ m}$	$\sim (10 \text{ cm})^2 \times 100 \text{ m}$	$\sim (17 \text{ cm})^3$
Energy/Inference	$\sim 100 \, \mu\text{J}$	$\sim 10 \text{ mJ}$	$\sim 1 \text{ J}$
Time/Inference	$\sim 100 \text{ ns}$	$\sim 10 \, \mu\text{s}$	$\sim 25 \text{ ms}$

Energy Efficiency: PFMs offer $10^4$ to $10^8$ times lower energy per inference compared to digital hardware. Even a $10^{18}$ -parameter PFM would consume roughly 100x less energy than a $10^{12}$ -parameter digital system.
Speed: Inference latency is reduced by $10^3$ to $10^5$ times (nanoseconds/microseconds vs. milliseconds).
Scaling: Unlike digital systems where energy scales linearly with parameters ( $E \propto P$ ), the optical PFM energy scales as $E \propto P^{1/3}$ due to the 3D volumetric nature of light propagation and the ability to increase input/output dimensions without increasing the number of parameters.
Feasibility: The calculations suggest that $10^{18}$ -parameter models could physically fit in manageable volumes (e.g., a 100m long tube) and be powered by portable energy sources, enabling "trillion-parameter" AI on edge devices.

5. Significance and Future Outlook

Enabling Ultra-Large Models: PFMs provide a plausible path to building AI systems with $10^{15}$ to $10^{21}$ parameters, potentially approaching the complexity of the human brain ( $10^{14}$ synapses) or exceeding it, which is currently impossible with digital hardware.
Edge AI Revolution: By drastically reducing power consumption, PFMs could allow foundation models to run locally on robots, cars, and sensors, eliminating latency and privacy concerns associated with cloud inference.
New Research Agenda: The paper outlines a massive research challenge: Physical Inverse Design. This involves training neural networks directly on the physical parameters of the hardware (e.g., refractive index patterns) rather than abstract weights.
Challenges Identified:
- Training: How to train $10^{18}$ -parameter physical systems without simulating them digitally (which is too expensive). Proposed solutions include "networks-of-networks" (Mixture of PFMs) and using digital "manager" models.
- Fabrication Variability: Since every manufactured PFM will have slight physical differences, the system must be robust to defects or use post-fabrication tuning (e.g., small digital compensators).
- Nonlinearity: Implementing necessary nonlinear activation functions in physical media (e.g., via optical Kerr effects) without catastrophic signal degradation.
- I/O Bottlenecks: The speed of converting digital inputs to analog physical signals (and back) remains a potential bottleneck.

Conclusion:
The paper argues that the era of foundation models necessitates a shift from general-purpose programmable hardware to specialized, fixed physical hardware. By leveraging the natural physics of materials (optics, electronics) to perform computation, PFMs could overcome the fundamental limits of digital computing, enabling a new generation of AI that is faster, more energy-efficient, and capable of scales previously thought impossible.