Analog Computing Returns: Why the Future of AI May Not Be Digital

01 What Analog Computing Actually Means

An analog computer solves problems by using physical quantities to represent mathematical variables. Voltage, shaft position, fluid pressure, or current can each stand in for a number in a computation. The machine performs operations by manipulating those physical quantities directly: adding voltages with resistors, integrating with capacitors, multiplying with potentiometers. The result is not a sequence of binary digits but a continuous signal whose value is the answer.

This stands in fundamental contrast to digital computing, where every quantity is discretized into bits and every operation is a logical step performed on those bits. Digital computers are exact but serial: they process one operation at a time per core, even if that core runs billions of operations per second. Analog computers are approximate but parallel: every component operates simultaneously, and the computation happens in the physics of the device rather than in a sequence of logical gates.

The distinction matters because neural network computation is fundamentally about matrix multiplication and addition, operations that analog circuits can perform in a single step by exploiting Kirchhoff's laws. When current flows through a resistor network, the physics of the circuit naturally computes the weighted sum. No clock cycles, no instruction decoding, no memory fetches, just electrons doing arithmetic at the speed of light.

02 The Golden Age of Analog Computers

From the 1930s through the 1960s, analog computers were serious tools for engineering and science. Vannevar Bush's Differential Analyzer, built at MIT in 1931, used shafts, gears, and wheel-and-disc integrators to solve differential equations that described mechanical systems, electrical circuits, and ballistic trajectories. It was used for power grid analysis, fire control, and early atomic weapons research. By the 1950s, electronic analog computers built from operational amplifiers became standard equipment in aerospace laboratories, where they simulated missile guidance systems and aircraft dynamics in real time.

The advantage of these machines was speed and interactivity. An engineer could turn a knob to adjust a parameter and immediately see how the solution changed, because the computation happened continuously rather than in discrete steps. For problems expressed as differential equations, analog computers could produce solutions as fast as the physical components responded. The disadvantage was precision: analog computers were limited by component tolerances, noise, and drift, typically achieving accuracy of one part in a thousand or so.

03 Why Digital Won the 20th Century

The digital computer overtook the analog computer in the 1960s and 1970s for several converging reasons. Transistors replaced vacuum tubes, making digital logic cheap, reliable, and compact. Integrated circuits scaled exponentially, following what became known as Moore's Law, doubling transistor counts every two years. Digital computers could represent numbers with arbitrary precision, store programs in memory, and run different programs without rewiring. Analog computers, by contrast, were purpose-built: each machine solved a specific class of equations, and reconfiguring one meant physically rearranging components.

General-purpose digital computing won on flexibility and accuracy. A single machine could run payroll, simulate physics, compose music, and play chess. The same architecture that solved one problem could solve any problem, given enough memory and time. Analog computing retreated into niches where real-time control mattered more than precision: process control in chemical plants, flight simulators, and specialized signal processing equipment.

04 The Energy Wall in Digital AI

Digital computing has been hitting physical limits for over a decade. Transistor scaling has slowed as feature sizes approach atomic dimensions, and the power density of modern chips has become a first-order constraint. A single NVIDIA H100 GPU draws 700 watts. Training a frontier language model requires thousands of these GPUs running for weeks, consuming enough energy to power a small town. The cost is not just financial; it is a physical limitation on how much AI the world can deploy.

The problem is fundamental to the digital paradigm. Every digital operation requires switching transistors between voltage levels, charging and discharging capacitors, and moving data between memory and processing units. The energy cost per operation has been decreasing but is bounded below by the physics of switching. Moreover, the von Neumann architecture, which separates memory from processing, requires shuttling data back and forth, consuming energy that grows with model size. For a 175-billion-parameter language model, each inference requires loading gigabytes of weight data from memory, and that data movement dominates the energy budget.

Analog computing sidesteps this problem entirely. In an analog neural network chip, the weights are stored as physical properties of the circuit components, such as the conductance of memristive devices. The matrix multiplication happens in the physics of current flowing through the circuit, not in a sequence of digital operations. The energy cost is proportional to the physical work done by the current, which can be orders of magnitude smaller than the energy needed to perform the equivalent digital computation.

05 How Analog Neural Computing Works

The core operation in a neural network layer is a matrix-vector multiplication: multiply each input by a weight, sum the products, and pass the result through an activation function. In an analog implementation, this entire operation happens in a single step using a crossbar array of programmable resistors or memristors.

Input values are converted to voltages and applied to the rows of the crossbar. Each column of the crossbar collects the current from all rows, and by Kirchhoff's current law, the total current in each column is the sum of the input voltage times the conductance of each intersection point. That sum is exactly the matrix-vector product. The conductance of each memristor is the weight, the input voltage is the input value, and the column current is the output. No clock cycles, no memory access, no instruction pipeline. The computation is the physics.

The activation function is applied by an analog circuit at the bottom of each column, typically a simple thresholding or transimpedance amplifier. The output is then either fed to the next layer as an analog signal or converted back to digital for storage or further processing. The entire layer processes all inputs and all outputs simultaneously, achieving a level of parallelism that digital systems cannot match.

06 Real-World Analog AI Chips and Prototypes

Several companies and research groups are building analog AI processors today. IBM Research has demonstrated analog chips using phase-change memory devices that store weights as conductance values, achieving energy efficiency orders of magnitude beyond digital GPUs. Mythic, a startup, built analog AI chips using flash memory cells as programmable resistors, running neural network inference at a fraction of the power of equivalent digital hardware. researchers at Stanford, MIT, and Tsinghua University have published prototypes using memristor crossbar arrays that perform matrix multiplication in a single analog step.

These chips are not general-purpose computers. They are specialized accelerators that excel at the one operation neural networks need most: multiply-accumulate. They cannot run an operating system, execute arbitrary code, or perform the vast range of tasks a digital processor handles. But within their domain, they offer a combination of speed, energy efficiency, and parallelism that digital hardware cannot match. The likely future is hybrid: digital processors handle control flow, memory management, and complex logic, while analog accelerators handle the dense matrix operations that dominate neural network inference.

07 What Analog Computing Means for the Future of AI Hardware

The return of analog computing is not nostalgia. It is a response to a concrete problem: the energy cost of AI is growing faster than the efficiency of digital hardware, and at some point the physics of digital switching becomes the binding constraint. Analog computing offers a fundamentally different physical basis for computation, one where the operation happens in the continuous behavior of physical systems rather than in discrete logical steps.

The challenges are real. Analog systems are sensitive to noise, temperature, and component variation, which means they produce approximate results rather than exact ones. For tasks that require precise arithmetic, this is a dealbreaker. But neural networks are already approximate: they are trained to be robust to noise in their inputs, and their outputs are statistical rather than deterministic. The noise inherent in analog computation may be acceptable, even beneficial, acting as a form of regularization that prevents overfitting.

If analog AI chips mature to commercial viability, the implications are significant. Inference could move to edge devices with microwatt power budgets, enabling always-on AI in phones, sensors, and implants. Data centers could reduce their energy consumption by orders of magnitude, easing the strain on power grids. And the frontier of what models can be deployed would expand, no longer limited by the energy cost of running them. The future of AI hardware may look less like a digital processor and more like the machines that preceded it: physical systems that compute by being, not by counting.