The Shortage Map of the "Engine of Light": The Wafer Even NVIDIA Is Scrambling For Is AI Optical Interconnect's Deepest Bottleneck
Everyone is watching GPUs and CPO, but the deepest bottleneck in AI optical interconnect is actually a 6-inch wafer — indium phosphide (InP). It is the "engine" of high-speed optical modules: EMLs and CW lasers are grown on it, and even silicon photonics and TFLN need an external InP source to emit light. Supply falls more than 70% short, over 90% of capacity sits with three companies, and volume relief won't arrive until 2028. NVIDIA has put $2 billion straight into Coherent as a 6-inch InP fab breaks ground — that's not an investment, it's a capacity grab.
1. Why now: NVIDIA turned a wafer into headline news with $2 billion
On June 16, 2026, construction began on an expansion of a 6-inch InP wafer fab in Sherman, Texas. It made headlines because the site belongs to Coherent, and that same week NVIDIA announced a $2 billion equity investment plus multi-year laser and optics purchase commitments (Source: NVIDIA Blog). The keywords for this fab are "6-inch" and "4x" — Coherent plans to quadruple its InP capacity to supply the lasers in NVIDIA's photonic networking switches (Source: Construction Review).
In plain terms: if you expect the scarcest thing over the next three years to be "enough, cheap enough InP lasers," the most rational move isn't to queue up and place orders — it's to take an equity stake and lock in capacity. What NVIDIA is buying isn't an investment; it's supply insurance.
2. Clearing up a misconception: silicon photonics and TFLN won't "replace" InP — they make it even scarcer
The market's biggest misreading is treating InP as a legacy material about to be replaced by silicon photonics. Think of optical communications as a relay race: InP runs the first leg and generates the light, TFLN runs the middle leg and handles modulation, and silicon photonics handles integration. The key point — silicon photonics solutions don't emit light themselves; they need an external InP CW laser. So the higher silicon photonics penetration goes, the more InP is needed. In 2026, silicon photonics is expected to exceed 50% share at 800G and reach 70% to 80% at 1.6T, and every one of those needs an InP laser to light it up. NVIDIA even projects InP wafer demand to surge roughly 20x from 2026 to 2030.
The shift to silicon photonics isn't InP's death knell — it's its amplifier. So just how short is that "engine" right now?
──────────────────────────────
That's the summary of this article.
STT's full analysis — the price and supplier map of the entire chain from phosphorus to laser, the real choke point behind the 70%+ gap in high-grade, large-diameter wafers, the rare-earth political card and the physical time lock of 2028, plus Taiwan's awkward position in this chain and three signals to track — is available in full in the paid section.
Related Reading
NVIDIA Buys Into Optics: When Optical Communications Goes from Growth Stock to Index Heavyweight: the full context of NVIDIA's investment in Coherent and the shift in capital structure
TFLN Breaks Another Record, but Don't Misread the Signal: Thin-Film Lithium Niobate's Opportunities and Three Bottlenecks in Optical Communications / CPO: another key material line complementary to InP
Comments