Press enter or click to view image in full size
9 min read Just now
--
As someone who writes about AI, I practice what I preach — AI tools were used to assist in the writing and editing of this article.
NVIDIA’s H100 GPU costs around $30,000. It runs so hot that it needs industrial cooling. Right now, thousands of them are stacked in data centers around the world, running continuously — every hour, every day — just to keep up with demand for AI tools that did not exist five years ago.
That is not a sustainable model. The people building the next generation of hardware understand this better than anyone.
Here is what most AI coverage misses: the bottleneck is not the algorithms. It is not the data. It is physics. Transistors have been shrinking for sixty years — but the smaller they get, the more heat they produce relative to the work they do. At some point, you are not just engineering a chip; you are fighting thermodynamics. Moreover, thermodynamics does not negotiate.
So where does AI go when silicon stops cooperating? Three technologies are fighting for that answer: photonic computing, neuromorphic chips, and quantum processors. None of them will dethrone NVIDIA next year. However, at least one is already inside a working supercomputer — and that changes the conversation from theory to fact.
The GPU was never designed for AI — it was designed to render video games. Its parallel architecture turned out to be a surprisingly good fit for the matrix math used in neural networks, so the AI industry borrowed it and never gave it back. NVIDIA turned that historical accident into a trillion-dollar business.
However, the costs are staggering. According to Patterson et al. (arXiv:2104.10350), based on hardware measurements provided directly by OpenAI and Microsoft, training GPT-3 consumed approximately 1,287 megawatt-hours of electricity. The carbon footprint came to roughly 552 metric tons of CO₂ equivalent, calculated using the US average grid carbon intensity for Azure. That is a one-time cost. The permanent cost is inference.
Training is a one-time cost. What never stops is inference — queries, generated images, real-time summaries, round the clock, at scale. That is where the electricity really goes. Meta’s internal data (Wu et al., 2022) indicates that inference accounts for up to 70% of their AI’s power consumption. Google’s figures (Patterson et al., 2022) landed at 60% of their ML energy. Neither number is a universal law — both are specific to each company’s workload — but the pattern is consistent: once a model is deployed and running, inference owns the bill.
Moreover, that scale is only growing. The International Energy Agency projects global data center electricity consumption will roughly double by 2030. AI is the primary driver.
Something has to change. Here are the leading candidates.
Forget electrons. What if you moved information using photons instead?
That is the core proposition of photonic computing, and the physics are genuinely compelling. Electrons traveling through silicon generate heat — they push against the material, lose energy to resistance, and that energy has to go somewhere. Photons in a glass waveguide do not work that way. There is no resistance, no capacitor charging, no heat overhead — the light moves.
A 2025 paper in Nature Communications Physics studying electro-photonic accelerators running BERT-large inferences found they used roughly one order of magnitude less energy per inference than an NVIDIA A100 GPU. A separate paper (arXiv:2401.05121, Photonics for Sustainable Computing) put the advantage at multiple orders of magnitude for deep neural network inference versus CMOS systems. Read those numbers carefully — both involve experimental systems running specific workloads at analog precision, not chips you can reprogram for arbitrary tasks. The efficiency gains are genuine, but so is the flexibility cost. For high-volume, fixed-inference work in data centers, though, that trade-off starts to look worth it.
Photonic systems also gain a parallelism advantage through wavelength-division multiplexing (WDM). Instead of one signal on one fiber, you send dozens of signals simultaneously — each at a different wavelength of light. Think of it as running 64 conversations on the same wire, simultaneously, without any of them interfering.
Q.ANT, a Stuttgart-based company spun out of Trumpf (one of Germany’s major industrial groups) in 2018, is the clearest real-world proof point that photonic computing has left the lab.
In July 2025, Q.ANT deployed the world’s first analog photonic co-processor in a live, operational HPC environment — Germany’s Leibniz Supercomputing Center (LRZ). Not a pilot program. Not a controlled demo. A working research supercomputer. They followed that in March 2026 with a second-generation deployment at the same facility, funded by the German Federal Ministry of Research. Benchmark results showed their Gen 2 processors delivering more than 50x higher throughput for matrix multiplication and 25x faster inference on a ResNet-18 than their first-generation processors. Bob Sorensen, Senior VP for Research at Hyperion Research, said the deployment “moves photonic co-processing beyond proof-of-concept and into production HPC environments.”
EE Times independently reported in November 2025 that Q.ANT’s Gen 2 chip runs at 8 GOps. CEO Michael Förtsch told the publication directly: “These are complex analog nonlinear operations, not the standard multiply-accumulate calculations GPUs are benchmarked against.” You cannot line them up on the same scale. The generational jump is real — but it measures Q.ANT’s own internally defined operational units improving, not a head-to-head gain over a digital system. Q.ANT’s roadmap targets 100,000 GOps by 2028, which no independent analyst has validated, and the gap from 8 GOps today is very large.
What is verified: Q.ANT raised a €62 million Series A in 2025 (co-led by Cherry Ventures, UVC Partners, and imec.xpand, per SPIE Optics) and has live hardware running inside an operational supercomputer. For a technology that was lab-only five years ago, that is a significant line crossed.
The second approach is harder to summarize — and in some ways the most intellectually interesting of the three.
Join Medium for free to get updates from this writer.
Your brain runs on roughly 20 watts. For context: a single H100 GPU draws 700. The brain processes input continuously, learns from experience, handles noise and ambiguity — and does all of it without the architecture that every computer you have ever used relies on. Von Neumann systems keep memory and processing separate, shuttling data back and forth between the two. That back-and-forth has an energy cost baked into every cycle. The brain sidesteps it: memory and computation happen at the same place — the synapse — and the whole system only activates when there’s something worth responding to.
That is what neuromorphic chips try to build in silicon.
IBM’s TrueNorth chip was one of the first serious hardware demonstrations of this principle. Its specs: 4,096 cores, 1 million programmable neurons, 256 million synapses, fabricated on a 28nm CMOS process, running at 70 milliwatts — IBM claimed a power density 1/10,000th of conventional microprocessors. (Note: IBM has since commercially discontinued TrueNorth, shifting its hardware focus toward quantum computing and hybrid cloud.)
Intel followed with Loihi, then scaled the concept further than anyone had. Its Hala Point system — announced in 2024 and at the time the world’s largest neuromorphic system — stacks 1,152 Loihi 2 processors into a single six-rack-unit chassis. That brings it to 1.15 billion neurons and 128 billion synapses, distributed across 140,544 neuromorphic cores.
Honesty requires acknowledging the gap. As a December 2025 Nature Computational Science review put it, strong barriers remain between neuromorphic engineering and the large language models that dominate today’s AI landscape — particularly transformers. Spiking neural networks (the computational model that neuromorphic hardware runs) are harder to train and generally less accurate on standard benchmarks than dense floating-point math, which GPUs excel at. Bridging that gap is an active area of research, not a solved problem.
Where neuromorphic chips are already compelling: edge applications. Robotics, autonomous drones, hearing aids, industrial sensors — anywhere you need always-on intelligence in a battery-powered device. The physics make a strong argument. The brain runs on 20 watts. That is hard to argue with.
Quantum computing gets the most hype and generates the most confusion, so let us be precise.
A quantum computer does not run calculations faster. It runs different ones — built on behaviors that classical physics simply does not allow. Superposition allows a qubit to exist in multiple states at once rather than being fixed to 0 or 1. Entanglement links two qubits so tightly that measuring one tells you the state of the other instantly, no matter the distance between them. Those two properties together let quantum systems cut through certain problem types that would take conventional machines effectively forever to solve. For specific problem classes — such as large-number factoring, molecular simulation, and combinatorial optimization — the theoretical advantage is exponential.
For running a large language model? Quantum computers are not the right tool, at least not with architectures that exist today. The honest position is that quantum’s near-term impact on AI workloads is indirect and limited.
What is more interesting is a specific material approach that sidesteps quantum computing’s biggest practical obstacle: the need for cryogenic cooling. Most quantum computers today must be cooled to fractions of a degree above absolute zero — an enormous engineering burden and ongoing energy cost.
Ephos, an Italian-American startup founded by physicist Andrea Rocchetto, is building photonic chips from glass rather than silicon and using them as the substrate for quantum photonic circuits. Why glass specifically? According to Ephos’s technical documentation and independent coverage in Silicon Republic, Photonics Spectra, and PIC Magazine, glass carries optical signal losses less than one-tenth those of silicon — meaning light travels farther through it without degrading. Moreover, unlike silicon-based systems that require cryogenic cooling, glass-based photonic circuits operate entirely at room temperature: no liquid helium, no dilution refrigerators, no cooling infrastructure that draws its own power budget.
The fabrication process is unusual enough to warrant description: Ephos uses a femtosecond laser to write waveguides directly inside the glass, in three dimensions — not etched onto a surface, but buried within the material itself. Silicon’s 2D planar process cannot do that. In September 2024, Ephos raised $8.5 million in seed funding led by Starlight Ventures. NATO’s Defense Innovation Accelerator and the European Innovation Council separately backed the company — Ephos was chosen from over 1,300 applicants for both. Then in July 2025, the European Commission formally approved a €41.5 million EU Chips Act grant to fund Fab-2, Ephos’s manufacturing facility in Milan. That approval, confirmed under EU state aid rules, made Ephos the first startup to receive direct Chips Act funding, alongside recipients such as TSMC, Infineon, and STMicroelectronics.
Quantum computing’s timeline for practical AI impact remains genuinely contested among researchers. However, the material science Ephos is building — low-loss, room-temperature photonic circuits — does not care whether you call it quantum or classical. The physics work either way.
Not soon. Not completely. Moreover, almost certainly not in the way the question implies.
NVIDIA’s real moat is not the hardware — it is CUDA. That software ecosystem took twenty years to build. Every AI researcher, every ML engineer, every graduate student running experiments writes code for it. Switching away from CUDA is not a procurement decision. It is a question of whether you are willing to rewrite millions of lines of code and retrain the institutional muscle memory of an entire industry. Nobody does that lightly.
However, “replace” is probably the wrong frame anyway. The more honest picture is specialization — each technology finding the slice of the problem it was actually built for:
The energy problem in AI is not coming — it is already here. Electricity contracts are getting harder to sign. Grid capacity limits where data centers can be built. Companies that crack the hardware problem are not just winning a market; they are winning the future. They are answering a question the whole industry needs answered: Can AI actually scale without the grid collapsing under it?
This article originally appeared in my Substack newsletter, [https://substack.com/@imranvaliani]. Subscribe for new posts in this series.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。