Morgan Stanley Updates CPO Supply Chain: TSMC PIC Capacity Is the Mass Production Gatekeeper
N.R. Finch
Morgan Stanley's latest report pushes the CPO discussion to the packaging-interface layer, identifying TSMC's photonic IC (PIC) capacity as the binding constraint on optical-engine shipments — capacity numbers look strong, but yield and test throughput decide how much actually ships.
How fast is TSMC ramping PIC capacity?
Current capacity sits at roughly 500 wafers/month. TSMC plans 10,000/month by Q2 2026, 15,000/month by Q4 2026, and at least 25,000/month by 2028.
This means → a 50× ramp in two years; TSMC is treating PIC as a standalone production line.
But capacity ≠ shipments. Each wafer yields about 648 dies; SoIC — a packaging process that stacks multiple chips together — runs at roughly 50% yield. If downstream assembly yield stays at 20%, 25,000 wafers per month cannot translate into mass optical-engine output.
In plain terms = yield is a funnel — many wafers go in, far fewer qualified optical engines come out. Only when assembly yield improves to 50% can actual shipments approach 48.6 million units.
Why is test throughput the underestimated bottleneck?
The report highlights the "Insertion 2" test step after SoIC packaging — covering electro-optical, opto-electronic, optical-optical, and high-speed S-parameter tests — as a hidden chokepoint for CPO mass production.
Test cycle time has improved from one wafer per day in H2 2025 to roughly six hours per wafer now. The target for the next six to twelve months is three to four hours per wafer.
This means → test speed directly determines whether TSMC's expanded capacity converts into deliverable optical engines. If testing lags, extra capacity is just a paper number.
Morgan Stanley argues that test-cycle improvement is a more convincing mass-production indicator than customer demos, and the core driver of order momentum for test-equipment firms such as Chroma, MPI, Ying-Zwei (YZTEK), and HongJin.
What does the CPO switch shipment curve look like?
Morgan Stanley projects CPO switch shipments at roughly 23,000 units in 2026 (mainly 100T, Nvidia Spectrum taking the largest share), 59,000 in 2027, and about 200,000 by 2030.
This reflects a shift from sample stage to a volume inflection point — but real scale arrives after 2028.
Supply-chain companies still stuck at NRE (non-recurring engineering fees) and small-batch samples cannot sustain high valuations; only entry into mainstream-platform mass production triggers a material valuation re-rate.
Is GlassBridge a disruptor or a long-dated option?
GlassBridge integrates glass waveguides, pluggable connectors, and passive alignment into a unified interface, aiming to improve manufacturability and serviceability for high-density fiber access — and to enable test-before-connect and field repair.
Morgan Stanley is measured: the technical roadmap is compelling, but GlassBridge currently aligns more with edge coupling and linear fiber layouts, making it hard to directly displace the grating-coupling path still pursued by TSMC COUPE, Nvidia, AMD, and Ayar Labs.
In plain terms = GlassBridge is a parallel lane — the direction may prove right, but the main-lane players are not switching yet.
Investment impact comes in two layers: first, it compresses the long-term valuation ceiling for traditional FAU suppliers; second, it signals that CPO is a system-wide migration across packaging, testing, optics, connectors, and architecture — not a single-component swap.
How much runway does the FAU supply chain still have?
Morgan Stanley expects the 2026 Nvidia Spectrum CPO switch to be the primary volume-revenue source for FAU (fiber array unit) supplier FOCI, which holds roughly 40% share in Nvidia's CPO FAU.
Nvidia's revenue contribution to FOCI is projected to rise from 18% in 2026 to 80% by 2028. This means → FOCI is concentrating its eggs in the Nvidia basket — customer concentration is a double-edged sword.
GlassBridge at this stage is more a long-dated risk item than a near-term displacement threat — the FAU supply chain still has a roughly two-year window.
How does Morgan Stanley rank CPO investments?
The report defines CPO investing as now in Phase 2: Phase 1 was buying the concept and optical-module beta; Phase 2 requires ranking by mass-production gates.
The ranking logic: first verify TSMC PIC expansion → Insertion test cycle time → Spectrum / Quantum / Rubin Ultra platform nodes, then check whether FOCI, O-Net (天孚通信), AllRing, MPI, and YZTEK can convert share into revenue and profit.
AllRing's CoWoS business remains its revenue base; CPO equipment revenue share is projected to rise from 11% in 2026 to 26% by 2028, with CPO equipment gross margins of 55–60%. This reflects a profit model built on packaging expansion and optical-engine coupling certainty, not a single-optics-path bet.
When GlassBridge moves from sample to customer-standard interface will be the key validation node for whether the broader CPO narrative holds.
Content is for reference only, not financial advice.