THE MOSAIC LEDGER

Structural shift

PARTIAL · structural shift · tier RESOLVED

Registered claim

Indium phosphide (InP) substrate capacity emerges as an explicit named constraint on AI compute interconnect deployment by 2026-Q3, on par with CoWoS-as-named-constraint in the prior 2024-2025 cycle. Pattern is established by 5+ independent attestations (VECO + AXTI + GFS SCALE + GLW Springboard + IntelliEPI primary-supplier voice) and at least one hyperscaler will explicitly cite InP supply as a deployment constraint in Q2 or Q3 2026 commentary.

Registered 2026-05-10.

Resolution spec

Spec registered 2026-06-12.

Verdict rule: HIT if leg A and leg B. PARTIAL if not leg A and leg B. MISS if not leg B.

As frozen: HIT = A && B; PARTIAL = !A && B; MISS = !B

Tier history

Evaluated daily since registration: 70 evaluations. The 4 that moved the tier are listed above.

Resolution receipt

Verdict: PARTIAL. Resolved 2026-09-30.

Rule: HIT if leg A and leg B. PARTIAL if not leg A and leg B. MISS if not leg B.

As frozen: HIT = A && B; PARTIAL = !A && B; MISS = !B

Autopsy

Read the full autopsy
Resolved PARTIAL on 2026-09-30. The call said two things about indium phosphide: that the supply chain would keep naming it as a constraint, and that at least one hyperscaler would say so too by the end of the third quarter. The first happened twice over, and the second didn't happen at all.

Start with what cleared. The bar was five independent supply-chain voices and the record holds ten: Veeco, AXT, GlobalFoundries, Corning and IntelliEPI when the call was made, then Coherent's Sherman fab, AXT's long-term supply deal, Lumentum naming AXT and converting its Greensboro line, Broadcom tripling laser capacity, and IQE turning existing tools over to indium phosphide. These are the companies that buy and make the material, saying so in their own filings.

The condition that mattered most failed on silence. Alphabet, Microsoft, Meta and Amazon reported in July and Oracle in September, and none of them named indium phosphide as something limiting what they could deploy. They talked about power, chips and memory, and nobody on those calls went three tiers down the optical stack.

Why not is the part worth being careful about. One reading is that the shortage is real and the buyers at the top simply don't feel it yet, because their transceiver suppliers are absorbing it through prepayments and reserved capacity. Another is that it never reaches them as a named line item, the way a hyperscaler doesn't name the resin in a substrate. The record can't yet tell those apart, and they lead to different places: the first says the citation is late, the second says it was never coming. What can be said is that the call borrowed its template from the advanced-packaging shortage of the last cycle, and that shortage was probably named mostly by the foundry and the chip designer, not by the cloud buyers either. If that's right, the comparison argued against the condition from the start.

There's a timing wrinkle as well. The window closed on September 30, and the calls that report the third quarter come in late October. A hyperscaler naming the constraint three weeks from now would vindicate the claim and change nothing here. That's how it should work. The date was written down in advance.

One more thing, and it matters for how much weight this result deserves: the conditions and the scoring rule were written a month after the call itself, on June 12, with a month of evidence already visible. The supplier count stood at five that day, so the condition that passed was already met when it was drafted. All of the information in this result is in the condition that failed.

What it changed: the thesis no longer argues that indium phosphide binds deployment as the buyers experience it. The claim that the substrate and epiwafer makers are a narrow group whose output is being locked up under contract stands, and stands on more evidence than in May. The part that reached up to the hyperscalers was withdrawn the day this resolved.

The authoring lesson is about who you ask. A shortage gets named by whoever has to explain a number because of it. For a material this deep in the stack that's the laser maker and the module maker, and maybe the systems vendor. Writing the test around the party furthest from the problem made the test about disclosure habits, and the supply question it was meant to settle got answered by the leg ranked second.200200