Everyone wants to find the next Nvidia. At a $5 trillion market cap, Nvidia has proved that compute is now the most important input in the global economy — more central than oil, more contested than any other resource. But the race to find what comes after the GPU requires answering three deceptively simple questions: What do we need to compute? How do we compute it? And with what kind of machine?
What AI Actually Computes
Roughly 98–99% of modern compute workloads come down to one operation: matrix multiplication, sometimes called a MAC (multiply-accumulate). Every neural network, every large language model, every image classifier runs on this primitive. The standard algorithm operates at O(n³) complexity — meaning if your matrix has size n, it takes n-cubed steps to complete. GPUs handle this today with around 15,000–16,000 cores running at 2 GHz, drawing roughly 1 kilowatt of power.
12:30
Whiteboard comparison of GPU specs: ~16,000 cores, 2 GHz clock speed, ~1 kW power draw
Watch at 12:30 →
This is the benchmark everything else must beat. So the question becomes: can a different kind of computer do matrix multiplication faster, cheaper, or at lower energy cost?
Why Quantum Computing Doesn't Solve the Problem
The appeal of quantum computing stocks like D-Wave or IonQ is straightforward — investors don't need to understand the technology. They just need to believe that Nvidia-scale returns are possible from a $5 billion starting point. That logic isn't irrational. But the underlying technology case is weaker than most people realize.
Quantum computers offer no known general algorithmic speedup for the dense matrix multiplications that power AI. The HHL algorithm (a quantum approach to linear algebra) requires sparse, well-conditioned matrices — conditions that don't hold in practical neural network workloads. Grover's algorithm applies to unstructured search, not matrix math. ChatGPT, asked directly, confirms: "No known general quantum algorithm provides a substantial exponential or even strong polynomial asymptotic improvement for arbitrary dense matrix multiplications."
And even setting the math aside, the hardware comparison is brutal. Where a GPU runs at 2 GHz with 16,000 parallel cores, today's quantum computers operate at roughly 1 MHz with a single effective core — and require cryogenic cooling that consumes enormous power. On every axis — speed, parallelism, energy efficiency — quantum loses to the GPU for AI workloads.
18:45
Side-by-side comparison table: GPU vs quantum computer specs — clock speed, cores, power consumption
Watch at 18:45 →
The Case for Photonic Computing
Photonic computing — using light instead of electrons to perform calculations — is where things get genuinely interesting. In theory, a photonic system can perform matrix multiplication in linear time rather than O(n³). Some architectures suggest it could approach O(1) for the core computation, with setup costs making it more accurately O(2n+1). The clock speed ceiling is the speed of light itself, putting a theoretical upper bound somewhere near 100 terahertz. And the energy footprint is dramatically lower than both GPUs and quantum systems.
The basic architecture works like this: a laser feeds into an optical processing unit containing mirrors, resonators, and lenses arranged in what's called a 4F configuration. This system can perform Fourier transforms and convolutions — and with the right setup, two input matrices go in and the result comes out at the speed of light.
25:10
Diagram of photonic computing pipeline: laser input, optical black box with 4F lens system, DAC and ADC conversion steps on either side
Watch at 25:10 →
The problem is everything around that core computation.
The Unsolved Problems in Photonic Computing
1. DAC/ADC Conversion
To feed data into a photonic processor, you must convert digital bits into light wavelengths (digital-to-analog conversion, DAC). To read the result, you convert back from light to digital (analog-to-digital conversion, ADC). These two steps are slow — slow enough to potentially negate the speed advantage of the optical core. Some experts in the field dismiss this as a secondary concern, but it deserves serious attention. Any end-to-end benchmark has to include conversion time, not just core compute time.
2. Nonlinearity
Matrix multiplications are linear operations — and light handles linear math naturally. But neural networks require nonlinear activation functions: ReLU, GELU, softmax, sigmoid, tanh, and others. These functions are what give neural networks their expressive power. Implementing them optically is genuinely hard. There are roughly five approaches being explored, each with its own tradeoffs, and none is clearly dominant yet.
3. Data Loading and Weight Adjustment
Even in a fully optical system with no DAC/ADC bottleneck, you still need to load matrix weights and update them during training. This is a non-trivial data movement problem that doesn't disappear just because the core compute is fast.
4. Manufacturing Cost
Any photonic chip needs to be manufacturable at a cost that can compete with Nvidia's economies of scale. That's a high bar — Nvidia has decades of process optimization and a deeply entrenched supply chain through TSMC.
Other Contenders and the Bigger Picture
Photonic computing isn't the only alternative architecture being pursued. Groq (recently acquired by Nvidia for approximately $20 billion) bet on SRAM-based computing — placing memory directly on-chip to eliminate the von Neumann bottleneck of shuttling data between processor and memory. A recent TSMC paper explores a hybrid approach combining SRAM buffers with memristor-based in-memory computing, potentially offering the speed of on-chip memory with greater capacity. Thermodynamic and analog computing represent a fifth category, with companies like Extropic raising $500 million at a $5 billion valuation on a largely pre-product thesis.
The most important competitive threat to all of these, however, is simply more Nvidia. Nvidia is not standing still. Their partnership with TSMC, their software moat through CUDA, and their ability to absorb promising startups means that any alternative architecture needs to deliver a step-change improvement — not just incremental gains.
The honest framing is this: unlike the early days of Nvidia's GPU dominance, we now know exactly what the market wants. It wants faster, cheaper matrix multiplication at lower energy cost. The demand signal is clear. What remains unknown is which physical substrate — photons, memristors, analog circuits, or something not yet named — will deliver it. That uncertainty is exactly what makes early-stage positions in companies like Lightmatter or Light Intelligence either very expensive mistakes or the best trades of the decade.
