For half a century, making AI faster meant one thing: more transistors on a chip. That era is quietly ending — not because we ran out of ideas, but because the bottleneck moved. The thing slowing AI down now isn't the math. It's moving the data between chips, and the staggering amount of energy that takes. The fix is the oldest trick in physics: send it as light. Hover or tap any underlined term.
Here's the thing almost no one outside the industry has internalized: Moore's Law — the engine that made chips relentlessly faster — has slowed to a crawl. Shrinking transistors keeps getting harder and costlier. So the gains now come from a different place. And when engineers measure where the energy actually goes on a modern AI chip, the answer is surprising: far more is burned shuttling data — from memory to processor, and from chip to chip — than doing the actual calculations.
That flips the whole problem. The question stopped being “how do we compute faster?” and became “how do we move data faster and cheaper?” There are two walls:
| The wall | What it is | Why it hurts |
|---|---|---|
| The memory wall | Getting data from memory into the processor fast enough | The chip sits idle waiting for data — addressed by HBM stacked memory and in-memory compute |
| The interconnect wall | Moving data between chips, and across the cluster | Copper wires lose speed and waste energy over distance — the wall light is built to smash |
Today, chips mostly talk to each other over copper. Copper is fine for short hops, but it has two brutal limits: the faster you push data through it, the more signal it loses to heat, and the farther you send it, the worse it gets. As AI clusters grow to tens of thousands of chips that must act as one machine, copper becomes the chokepoint. Light doesn't have those limits in the same way:
This is where the Lens earns its keep — separating the shipping-now layer from the science project. Co-packaged optics and silicon photonics are real and ramping, and the supply chain is identifiable:
| Layer | Who plays | What they sell |
|---|---|---|
| Optical switching silicon | NVIDIA, Broadcom, Marvell | The switch + DSP chips that route light through the cluster — NVIDIA is building co-packaged optical switches directly into its networking |
| The lasers & optical components | Coherent, Lumentum, Fabrinet | The actual light sources, modulators, and assembly — the “arms dealers” of photonics, paid no matter whose switch wins |
| Photonics pure-plays | Lightmatter, Ayar Labs, Celestial AI, POET | Startups built entirely around moving (or computing with) light — higher risk, higher torque to the theme |
| Advanced packaging | TSMC (CoWoS), Intel, Corning | The packaging — including glass substrates — that lets optics and chips sit together precisely |
Beyond moving data with light sits a row of ideas that range from “promising” to “don't bet the thesis on it.” We rank them so the hype can't:
Optical computing emerging — doing the matrix math in light itself, not just moving it. Real physics, real companies (Lightmatter, Lightelligence), unproven at scale. The step beyond optical interconnect.
Neuromorphic & in-memory compute emerging — brain-like, ultra-low-power chips, and doing math inside the memory array. Promising, repeatedly hyped, still mostly niche.
Room-temperature superconductors moonshot — would change everything (grid, magnets, compute) by moving electricity with zero loss. Also a graveyard of false claims (remember LK-99). Treat any announcement as guilty until peer-reviewed.
2D-material & carbon-nanotube transistors moonshot — the post-silicon transistor itself. A decade-plus horizon.
Quantum “for AI” category error — quantum computers help optimization, simulation, and cryptography. They do not train large language models. Anyone selling quantum as an AI-compute disruptor is confusing two different machines.
Step back and every one of these — optics, packaging, memory, even the CPU comeback — does the same thing: it removes a bottleneck, which lets the system get bigger or cheaper, which expands demand to fill the new room. A breakthrough doesn't shrink the AI buildout; it makes an already-exponential thing grow faster. That's why the durable bet never changes shape.
Dragonfly Lens maps where the real bottleneck sits — and who gets paid to break it — while labeling the moonshots as moonshots. Plain English, every claim sourced.
Join the Lens →Why is “light” the next bottleneck in computing? Because the bottleneck moved. On a modern AI chip, more energy is spent moving data between chips and memory than doing the actual math. Copper wires waste power and lose signal over distance, so the industry is switching to optics — sending data as light — which carries more data using far less energy.
What is co-packaged optics? Putting the optical engine (the part that turns electrical signals into light) directly onto the chip package, instead of off to the side. It lets chips communicate with light right where the data is, cutting the energy and distance penalty of copper — the most important near-term hardware shift in AI.
Will optical or quantum computing replace today's chips? Optical interconnect (moving data with light) is real and ramping now. Optical computing (doing math in light) is emerging but unproven at scale. Quantum is a different machine entirely — useful for optimization and simulation, not for training AI models. Don't conflate them.
Sources: Co-packaged optics / silicon photonics as the interconnect-and-energy fix; data movement dominating chip energy budgets — industry technical reporting and vendor disclosures (NVIDIA co-packaged optical networking, Broadcom/Marvell optical DSP & switch silicon, Coherent/Lumentum/Fabrinet optical components, Lightmatter/Ayar Labs/Celestial AI/POET photonics). Microsoft Project Silica glass storage (~2M books on borosilicate glass, ~10,000-year durability, femtosecond-laser write) — Nature, Live Science.
Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Technologies described are at varying maturity, clearly labeled; emerging and moonshot items are unproven and may not commercialize. Company names illustrate the supply chain, not buy recommendations. Verify against primary sources before acting. Past performance does not guarantee future results.