The Unphotogenic Frontier

At the start of 2025, Jensen Huang stood in front of a room of analysts and said a thing that sent the publicly traded quantum companies tumbling, billions of dollars of market value gone in an afternoon. Useful quantum computers, the kind that do work you would pay real money for, were fifteen to thirty years away. He gave the room a number to hold onto. Fifteen was probably the early side. Thirty was probably the late side. Pick twenty, and a whole row of people who build this for a living would nod along.

Around the same time, the company he runs put four billion dollars into light. In March 2026 NVIDIA committed two billion each to Coherent and Lumentum, two names most people outside the supply chain have never heard, both of them working in photonics, the technology of carrying information with photons instead of charge. Huang said so plainly, framing it as the work of building the next generation of gigawatt-scale AI factories. The signal sits in the distance between the two gestures. Quantum gets a date two decades out, the safe horizon you are free to dream about. Photonics gets the cash, now, in the present tense.

The bottleneck moved

For a decade the story of AI was a story about compute. More GPUs, bigger models, larger clusters, and the wall everyone watched was the chip. Sometime in the last two years that wall moved, and a good deal of the public conversation did not move with it. The constraint now is energy, and underneath energy, the cost of moving data. Goldman Sachs has data center power demand climbing roughly 160 percent by 2030, toward a level that approaches the entire annual electricity consumption of Japan. The Electric Power Research Institute has American data centers rising toward nine percent of national generation by the end of the decade, up from around four percent in 2023.

Inside the machine, the expensive part is no longer the arithmetic. It is the moving. Shuttling bits between chips, between memory and processor, across the racks of a training run, costs more energy than the computation those bits feed. A modern model spends a startling share of its power budget simply carrying numbers from one place to another and keeping the heat from melting the building. That is the real frontier, and it does not photograph well. There is no countdown clock on a data interconnect.

The data was always the hard part

This is the point where quantum is supposed to ride in. The pitch, repeated at conferences and in fund decks, is that quantum computing will do for AI what GPUs did, another order of magnitude, a fresh substrate underneath the whole field. The pitch rarely survives contact with the question of data.

In 2018 an eighteen-year-old undergraduate named Ewin Tang showed why. Her advisor, the complexity theorist Scott Aaronson, had handed her a celebrated 2016 quantum algorithm for recommendation systems, the kind of engine that decides what you might want to watch next, and asked her to prove the expected thing, that no classical computer could match its speed. She spent months trying to prove it and could not, because it was not true. She built a classical algorithm that matched the quantum one, running in comparable time, doing the same work without a single qubit. The exponential speedup everyone had celebrated turned out to live in the assumptions about how the data was fed in, not in the quantum hardware. Give a classical algorithm the same kind of sampling access to the data, and the advantage dissolved.

The result was clean enough to start a small industry. Researchers went looking at other prized quantum machine-learning speedups and dequantized them one after another, finding the classical shadow hiding inside each quantum claim. What Tang had really exposed was structural. For the data-heavy problems that look most like machine learning, the quantum advantage was a story about the input model, and the input model was doing the work.

The input model is where quantum keeps breaking. To run these algorithms on real data you first have to get classical data into quantum states, the job of something called QRAM, a quantum memory that can hand the processor many addresses at once. For a wide class of algorithms, loading the data costs as much time as you hoped to save, so the speedup is paid back at the door before any computation happens. Large-scale QRAM remains largely theoretical. It is an open question among the people closest to it whether a practical version will ever be built. The promise rests on a component that does not yet exist, to clear a data problem that was the binding constraint all along.

None of this makes quantum useless. It makes quantum useful for a different set of problems. Last October Google reported what it called the first verifiable quantum advantage, its Willow chip running a computation roughly thirteen thousand times faster than Frontier, one of the fastest supercomputers on earth. The computation was a physics simulation. It says something real about quantum systems and nothing at all about training a language model. There is, as of today, no known quantum algorithm that meaningfully speeds up the training or the inference of the transformer models the entire AI economy runs on. Estimates for running even a small one climb into the thousands of error-corrected qubits, against machines that currently count their reliable qubits on a hand or two.

The same enemy, twice

Now look back at the four billion dollars NVIDIA committed to photonics. The field does an unglamorous thing. It replaces the electrons that carry data between chips with photons, light moving through a waveguide instead of charge moving through copper. Light runs cooler, carries more, and spends less energy per bit. The figures people cite cluster around a tenfold gain in energy efficiency and large multiples of bandwidth. The market is still small and climbing on the kind of curve that stays invisible until the year it isn’t.

Hold the two frontiers next to each other and the same two words sit underneath both. Data movement. The binding constraint on AI is moving data. Quantum, the frontier that collects the magazine covers and the keynote countdowns, is itself stopped cold by data movement, because you cannot get the data in without handing back the speedup at the threshold. Photonics, the frontier almost nobody outside the supply chain can name, exists to attack data movement head on. One frontier is blocked by the problem. The other is the answer to it. The capital has already sorted the two, whatever the covers and the keynote countdowns say.

Watch the feet

There is a reading skill buried in all of this, and it travels well past quantum. Glamour absorbs attention. Infrastructure absorbs value. The technologies that get photographed, given countdowns, assigned a date by a chief executive on a stage, tend to be the ones far enough away to be safe to dream about. The technologies that actually move the next two years are often boring enough that describing them clears a room. An operator trying to read where a field is really going learns to discount the mouth and watch the money’s feet, to ask not what a technology promises but which constraint it removes, and for whom, and by when.

This is the same logic that governs careers and businesses built on top of AI platforms. Value accrues to whoever sits closest to the constraint that actually binds, and the constraint that binds is rarely the one with the brightest marketing around it. The people who read the infrastructure layer, the routing, the memory, the movement, the parts of the stack that exist to carry the work rather than to perform it, end up positioned where the leverage compounds. The people who read the forecast layer end up holding a date that keeps sliding.

Which leaves a question worth carrying out of this. If the most important deep tech for AI right now is a thing nobody puts on a cover, what else in the stack is load-bearing and unphotographed, holding up the whole structure while the cameras point somewhere more flattering. The frontier is seldom where the light is brightest. It is usually one layer down, in the part of the machine that exists to move the light itself.