Why does indium phosphide (InP) matter for AI?

Silicon cannot efficiently emit light, but indium phosphide can, so it is the base material for the lasers in AI optical links. It is grown on small 2 to 4 inch wafers by only a few suppliers, and AI demand jumped from telecom scale to cloud scale almost overnight, which is why Lumentum's CEO warned the shortage could be worse than memory.

Who benefits from the AI optical-interconnect shift?

The scarce layers underneath the switch brand: the indium phosphide substrate makers such as Sumitomo, AXT and Resonac, the laser and optical component makers such as Lumentum and Coherent, and the advanced packaging that stitches optics to silicon, including TSMC. Whoever wins the switch fight, those inputs get bought.

The Lens · The Bottleneck After the Memory Wall

AI's Next Bottleneck Is the Wire Between the Chips.

Q: What is co-packaged optics (CPO)?

Co-packaged optics builds the parts that turn electricity into light into the same package as the chip, instead of in a separate pluggable module across the board. That nearly eliminates the slow electrical hop and lets data leave the chip as a laser beam. Nvidia's roadmap introduces it in its systems around 2027 to 2028.

We wrote earlier about the memory wall — how AI's bottleneck moved from the chip to the memory feeding it. It's moving again, one step further out: to the wires that connect the chips to each other. AI clusters have gotten so fast that ordinary copper can no longer carry the data between chips without garbling it. The fix is to send it as light — and the lasers that make the light need a rare crystal that only a handful of companies produce. The CEO of one of them says the coming shortage will be “worse than memory.” Hover or tap any underlined term.

Dragonfly Lens · August 2026 · Updated 25 Aug 2026 · The newest link in the bottleneck chain.

The short version

The problem: inside an AI cluster, chips constantly pass data to each other over copper wires. But network speed doubles every 2–3 years, and the faster copper runs, the shorter the distance it can carry a clean signal — now down to tens of centimeters. Copper has run out of road.
The fix: replace the copper with light. Move tiny optical engines right next to the chip so data leaves as a laser beam instead of an electrical pulse — called co-packaged optics. Nvidia's own roadmap puts this into its systems in 2027–2028.
The bottleneck: those lasers are built on indium phosphide (InP), a specialty crystal grown on tiny 2–4 inch wafers (vs. silicon's 12-inch) by a few suppliers. Demand just jumped from telecom-scale “hundreds” to AI-scale “hundreds of millions,” and supply is short by more than 30%. Lumentum's CEO calls it “worse than memory.”
The opportunity (the Lens angle): this is the memory-wall pattern again, one layer out. The value pools in the scarce physical inputs — the InP substrate makers (a concentrated group, much like the HBM oligopoly), the laser and optical-component makers, and the packaging that stitches optics to silicon — not just the switch or GPU brand on top. Bet the bottleneck, not the horse.
The honest caveat: the CPO shift is a 2027–2028 target, not shipping at scale today; one analyst house (LightCounting) expects the InP crunch to ease by end-2026; and co-packaged optics has an unsolved serviceability problem (if one laser dies, you may have to replace the whole expensive package). This is a real, developing constraint — not a done deal.

The chain so far — and the newest link

If you've followed the Lens, this is a familiar shape. The AI buildout keeps hitting a wall, we solve it, and the wall moves to the next-most-physical thing. Each link is a piece we've mapped:

Power. First the limit was electricity — enough always-on power to run the data centers at all. (See the nuclear pieces.)

Memory. Then the chip could compute faster than memory could feed it — the memory wall, and the sold-out HBM that eased it.

Packaging. Then even memory wasn't enough if you couldn't assemble logic and memory together at volume — advanced packaging (TSMC's CoWoS) became a rate-limiter.

The interconnect (this piece). Now the chips are so fast that the wires between them are the limit. The fix is optical — and the optics have their own scarce input. This is the newest link, and it's tightening right now.

The plain-English version: why copper hit a wall

An AI cluster isn't one chip — it's tens of thousands of them that have to talk to each other constantly, at once, as if they were a single giant brain. That conversation runs over wires. And here's the physics: every couple of years the required speed doubles, but the faster you push data through copper, the more the signal smears and fades — so the distance it can travel cleanly keeps shrinking. We've reached the point where copper can only carry the top speeds a few tens of centimeters — not even reliably across a rack.

The crowded party analogy. Copper is like talking across a loud room. At a whisper (slow data) your voice carries fine. But to pass more information you have to talk faster and faster — and the faster you talk, the more your words blur together and the less far they carry, until you can only be understood by the person right next to you. Light (optics) is like handing them a written note instead: it doesn't blur with speed, and it carries clean across the whole room. That's the switch the whole industry is now making — from shouting over copper to passing notes in light.

Sending data as light isn't new — that's what fiber-optic internet is. What's new is moving the light-making hardware right up against the chip (co-packaged optics), so the slow, lossy electrical hop is almost eliminated. And to make light on a chip, you need a material that can emit it. Silicon — the stuff of ordinary chips — is bad at that. The material that's good at it is indium phosphide. Which is where the bottleneck lives.

What actually happened (August 2026)

Two things brought this from an engineering footnote to a supply-chain story:

First, Lumentum — one of the main makers of the lasers these links need — had its CEO, Michael Hurlston, warn publicly that the InP laser shortage is set to be more severe than the memory-chip shortage, with demand exceeding supply by more than 30%. His framing: demand has leapt from telecom-scale “hundreds” of units to cloud-scale “hundreds of millions.” Coherent, another key supplier, is racing to double its internal InP output in 2026 and more than double it again by end-2027 — but Lumentum's own new US fab isn't expected to contribute much until 2028. Update (25 Aug 2026): Coherent’s own results confirmed it from the other side of the market. In its fiscal Q4 2026 report (12 Aug), revenue hit $2.05B (+34% year-on-year) with datacenter & communications at $1.62B (+58.6%) — and management told investors that “indium phosphide capacity continues to be our primary constraint,” naming the ramp of 6-inch InP wafers (up from today’s 2–4 inch) as the limiter on how fast transceiver revenue can grow. Two supplier CEOs, one bottleneck.

Update (25 Aug 2026) — one layer deeper. Lumentum and Coherent make the lasers; the lasers are grown on InP substrate wafers, and that layer is thinner still. On 29 July 2026 Lumentum reserved AXT’s substrate capacity through the end of 2031, prepaying $87 million in two $43.5M deposits (the second in 2028) — a laser maker paying years ahead to guarantee crystal. The price tells the same story: InP substrates are up more than 200% in a year (a 6-inch wafer from roughly $1,400 to about $5,000, four consecutive quarterly hikes). And the geopolitics sit on top: AXT’s main plant is in Beijing and has needed Chinese export permits since June 2025 (permit delays cut its Q4 2025 revenue), while China has capped exportable indium at 30% of output. So the “handful of firms” point gets sharper the further upstream you look — and the scarcity is partly political, which capacity adds alone can’t fix.

Second, the industry started moving from proprietary experiments toward a shared standard: Lightmatter and a coalition of major hardware firms (Corning, Dell, Foxconn, Qualcomm, Keysight and others) proposed an open reference architecture for co-packaged optics inside the Open Compute Project. When an industry standardizes something, it's usually the sign the thing is about to ship in volume — and that the fights over who supplies it are beginning. Nvidia, for its part, has already been locking up laser supply: it took a ~$2 billion equity stake in each of Coherent and Lumentum (early-to-mid 2026), plus multi-year supply deals running through the end of the decade — roughly $4B to secure the optical layer. [WATCH: OCP spec milestones + any new Nvidia optics commitments.]

Read the numbers with one eye on the source. “Worse than memory,” “30%+ short,” and the 2027–2028 rollout dates are a mix of a supplier CEO's framing and roadmap targets — directionally well-supported, but not settled fact. And at least one analyst house expects the crunch to ease by end-2026. Anchor on the structural point (optics is the next interconnect, and its inputs are scarce and concentrated); treat the exact figures and dates as live and subject to revision.

The opportunity: who gets paid no matter whose switch wins

The horse-race question is “does Nvidia or Broadcom win co-packaged optics?” The Lens question is the one that pays: if data movement is going optical, who owns the scarce parts every optical link needs?

The layer	Who	Why they win either way
The crystal (substrate)	Sumitomo Electric, AXT, Resonac (+ emerging China supply)	The real chokepoint. InP substrates come from a few firms on tiny 2–4 inch wafers that can't be scaled up like silicon overnight. Concentrated supply meeting vertical demand — the same setup that made HBM so powerful.
The lasers, components & modules	Lumentum, Coherent (lasers); AAOI (800G/1.6T transceivers)	Whoever's switch wins, it needs the lasers, modulators, transceivers and optical engines — and those makers are already sold past their capacity (30%+ short). Nvidia is pre-paying to lock supply — a ~$2B stake in each of Coherent and Lumentum — and AAOI is already volume-shipping 800G AI transceivers, scaling toward 1.6T. The AI-cluster optics market is set to jump ~60% to $26B in 2026 (LightCounting).
The integrators / packaging	CPO designers (Lightmatter, Nvidia, Broadcom) + the foundry (TSMC)	Someone has to stitch the optics onto the silicon at volume — the same advanced-packaging chokepoint from the memory-wall story, now with optical parts added. The brand on the switch changes; the packaging bottleneck underneath doesn't.
The moonshot 🌙	Silicon photonics that need fewer InP lasers	Clearly speculative. If a design lands that moves more of the work onto cheap silicon and leans less on scarce InP, it reshuffles the board. Early, unproven — we'd only act on data.

The insight that ties it together: this is the exact same move as the memory wall, one layer further out. In a gold rush where everyone suddenly needs to move data as light, own the light-making inputs — the crystal, the lasers, the packaging — rather than betting on which switch brand wins the loud fight on top. The bottleneck keeps moving; the picks-and-shovels logic doesn't.

The risks — named, not buried

The timeline is a target. Co-packaged optics ramps in 2027–2028 on Nvidia's roadmap — not at scale today. Roadmaps slip; copper will hobble along in parallel for a while. This is a multi-year re-plumbing, not an overnight flip.
The shortage may be shorter than the headline. At least one analyst house (LightCounting) expects the InP crunch to ease by end-2026 as capacity comes online. A supplier CEO calling it “worse than memory” has an incentive to stress scarcity. Watch the actual capacity adds.
The engineering is messier than the clean story. Many production designs (Nvidia's included) keep the lasers outside the chip package in external laser-source modules — lasers are heat-sensitive and fail faster next to a 1,000-watt chip — and attaching the fiber arrays needs sub-micron alignment that must survive thermal cycling. Thermal management and yield are first-order problems, not footnotes. The bottleneck also isn't waiting for 2027: today's 800G/1.6T pluggable modules are already consuming more InP lasers than the industry can make.
Co-packaged optics isn't a solved product. Its big unsolved problem is serviceability — if a laser fails, replacing it can mean pulling the whole expensive package. That could slow adoption or reshape who wins.
Concentrated supply cuts both ways. “Own the InP oligopoly” also means owning a small, specialized, cyclical layer — powerful when demand runs hot, brutal when it turns.
Not advice. This is how we read the chain, not a recommendation. Company names illustrate a structural shift, not buy calls — do your own work and size to your own risk.

The viral take and the true take are rarely the same trade

The switch war grabs the headlines. The crystal decides who ships.

Dragonfly Lens maps the AI buildout as one connected chain — and keeps finding the layer where the value actually pools as the bottleneck moves. Plain English, every claim sourced and flagged. When we're wrong, we say so.

Join the Lens →

Not ready to subscribe? Get the free weekly →

More: The memory wall · Big Tech goes nuclear · All explainers

Quick answers

Why is copper a problem for AI now? Inside an AI cluster, chips constantly exchange data over wires, and the required speed doubles every 2–3 years. The faster data moves through copper, the shorter the distance it can travel before the signal degrades — now just tens of centimeters at top speeds. Copper can no longer reliably carry the traffic even across a rack, so the industry is switching to light.

What is co-packaged optics (CPO)? It's building the parts that turn electricity into light into the same package as the chip, instead of in a separate module across the board. That nearly eliminates the slow electrical hop and lets data leave the chip as a laser beam. Nvidia's roadmap introduces it in its systems around 2027–2028.

Why does indium phosphide (InP) matter? Silicon can't efficiently emit light; indium phosphide can, so it's the base material for the lasers in AI optical links. It's grown on small 2–4 inch wafers by only a few suppliers, and AI demand has jumped from telecom-scale to cloud-scale almost overnight — which is why Lumentum's CEO warned the shortage could be “worse than memory.”

Who benefits from the optical-interconnect shift? The scarce layers underneath the switch brand: the InP substrate makers (Sumitomo, AXT, Resonac), the laser and optical-component makers (Lumentum, Coherent), and the packaging that stitches optics to silicon (including TSMC). Whoever wins the switch fight, those inputs get bought.

Sources: InP laser shortage "worse than memory," demand>supply by >30%, telecom-to-cloud scale jump; Coherent doubling InP output 2026 / >2x by 2027; Lumentum NC fab ~2028; 2-4in wafers vs 12in CMOS; Nvidia multi-billion supplier lock-ups; LightCounting sees easing by end-2026 — TrendForce, MLQ/Lumentum, Futurum. Copper reach shrinks as bandwidth doubles every 2-3 years; CPO ramps 2027-2028 (Nvidia GTC roadmap; Rubin Ultra late 2027); optical packaging becomes the bottleneck 2028-2030 "similar to the HBM shortage"; serviceability debate — Network World, SemiAnalysis, Semiconductor Engineering. Lightmatter OCP co-packaged-optics reference-architecture initiative (Corning, Dell, Foxconn, Qualcomm, Keysight, et al., Mar 2026) — Lightmatter, HPCwire. InP substrate makers (Sumitomo Electric dominant, AXT, Resonac; emerging China supply; Eindhoven 6-inch fab) — Mordor Intelligence. Nvidia ~$2B equity stake in EACH of Coherent + Lumentum (~$4B) + multi-year supply deals through end-decade; AAOI volume-shipping 800G AI transceivers, scaling to 1.6T; LightCounting: AI-cluster optics +60% to $26B in 2026 — Futurum, Yahoo Finance / Zacks (AAOI).

Educational research, not personalized investment advice. Dragonfly Lens is not a registered investment advisor. Figures are as reported by the sources above; several (shortage severity, rollout dates) are supplier framing or roadmap targets and carry revision risk — this is a developing story. Company names illustrate a structural shift in the AI compute supply chain, not buy recommendations — verify against primary filings before acting. Past performance does not guarantee future results.