Quantum computing gets the headlines, but it offers essentially no advantage for the one computation that matters most in AI: matrix multiplication. Photonic computing — using light instead of electricity to perform math — could change the complexity class entirely, potentially displacing Nvidia's dominance. Here's a clear-eyed look at why quantum falls short, where photonics shows genuine promise, and what the remaining engineering obstacles actually are.
Why Quantum Computing Doesn't Help AI
The honest case against quantum computing as an AI accelerator comes down to one inconvenient fact: there is no quantum speedup for matrix multiplication. Whether you're on a GPU or a quantum processor, multiplying two matrices of size n remains roughly O(n³). Exotic algorithms like Grover's can get you to n^2.5, but that's the ceiling — and it applies to classical hardware too.
What quantum computing does offer is a spectacular reduction for a very narrow class of problems: going from exponential complexity (2^n) down to polynomial (n² or n³). This is mathematically elegant and genuinely exciting. But the only practical application is period-finding, which underpins breaking RSA and elliptic-curve cryptography. That's it. OpenAI doesn't need it. Nvidia doesn't need it. Microsoft, Meta, and Anthropic don't need it — and notably, none of them are doing deals with quantum companies. They're buying memory, TPUs, and traditional chips.
When quantum companies claim they can help with AI, listen carefully: they talk about optimizing hyperparameters, not running neural networks. That's a footnote, not a revolution. The other use case quantum vendors tout — optimization problems like the traveling salesman problem — represents a vanishingly small fraction of real-world compute. If 90% of compute is AI, and most of the rest is databases, sorting, and search, traveling salesman isn't even a rounding error.
There's also the hardware reality. IBM's best machines sit around 150 qubits. A working, useful quantum computer would require something like 10 million qubits. A 256-qubit machine, even a perfect one, can hold roughly 32 bytes of information — enough to store a short name. Claims about loading the entire bond market into a quantum computer to solve portfolio optimization are physically impossible: there's no RAM, no bandwidth, and no way to encode that volume of data into the machine.
What Photonic Computing Actually Does Differently
Light performs certain mathematical operations naturally. Destructive and constructive interference subtract and add. Fourier transforms happen physically as light propagates. These aren't approximations — they're exact analog computations that emerge from the physics of waves. This means a photonic processor can perform the matrix multiplications at the heart of neural networks in linear time: O(2n), compared to O(n³) for conventional hardware.
32:10
Diagram showing wave interference performing destructive subtraction — the physical basis for photonic matrix math
Watch at 32:10 →
This is a fundamentally different complexity class. If 90% of compute today is AI, and 90% of AI compute is matrix multiplication, then a processor that handles matrix math in linear time is not an incremental improvement — it's a potential regime change. Microsoft has estimated that AI accounts for 99% of cloud energy consumption. Solving matrix multiplication faster isn't a niche optimization; it's the whole game.
Unlike quantum computing, working photonic computers already exist. Researchers — including teams at Tsinghua University and Microsoft — have demonstrated functional optical neural networks running image classifiers like MNIST and Fashion-MNIST with strong accuracy. These aren't simulations. They are physical systems performing inference using light.
48:20
Microsoft photonic chip demonstrating MNIST and Fashion-MNIST classification results
Watch at 48:20 →
The Core Engineering Problem: Nonlinearity
Here's where photonics hits a genuine wall. Electricity gives rise to transistors, which are nonlinear. Nonlinearity is what makes logic gates possible — and it's what allows neural networks to apply activation functions like ReLU, sigmoid, and softmax between layers. Light doesn't naturally produce nonlinear behavior at useful scales.
This creates a fundamental pipeline problem. A photonic neural network looks roughly like this:
- An API call arrives; a digital-to-analog converter (DAC) encodes the input as light
- The light passes through the photonic chip, performing matrix multiplications at linear complexity
- After each layer, an activation function must be applied — which requires converting back to digital (ADC), running the activation, then converting back to analog
- This ADC/DAC round-trip happens between every single layer of the network
39:45
Diagram of a multi-layer neural network with ADC/DAC conversion steps between photonic layers
Watch at 39:45 →
Neural networks are inherently sequential — you cannot compute layer two without the outputs of layer one. Each conversion step introduces latency. If that latency is large enough, it erases the speed advantage gained from doing the matrix math in linear time. This is the central unsolved problem in photonic computing.
Some engineers working in photonics argue that with correct pipelining the ADC/DAC penalty is manageable. Others are less optimistic. What's clear is that solving this problem — ideally by finding optical equivalents of activation functions, so the signal never has to leave the photonic domain — would be a breakthrough comparable to early transistor scaling. Weight loading is a secondary challenge: getting the model parameters into the photonic hardware efficiently is nontrivial at scale.
Timeline, Stakes, and Honest Perspective
A practical, scalable photonic computer capable of running large neural networks end-to-end is probably 5 to 10 years away. That's not dismissive — it's a realistic read given where prototype hardware stands today. The physics works. The engineering is hard.
The commercial stakes are straightforward. Nvidia is worth roughly $5 trillion largely because it dominates matrix multiplication hardware. A photonic computer that performs that function faster and more efficiently doesn't just take market share — it redefines the architecture of AI infrastructure. That's why Microsoft, Taiwan Semiconductor, and academic labs in China are investing in silicon photonics research now.
Quantum computing, by contrast, is chasing a narrower prize. The mathematical miracle of going from exponential to polynomial complexity is real and beautiful — but it applies to cryptographic period-finding, not to the workloads that consume the vast majority of compute dollars today. A working, scalable quantum computer would be a scientific landmark. It would not replace your GPU.
The distinction matters for investors and technologists alike. Photonic computing has a clear path to relevance in AI infrastructure, demonstrated prototypes, and a solvable (if difficult) engineering roadmap. Quantum computing has extraordinary mathematics and genuinely limited near-term application to the problems that actually dominate compute demand. Following the money — which flows toward memory, traditional chips, and inference hardware — tells you where the industry currently places its bets.
