Optical Chip Scarcity and OCS Outsourcing Reshape AI Networking Profit Pools

核心结论

Global AI capex continues to drive optical transceiver unit growth, but the investment case is bifurcating. Transceiver ASPs are falling 10% annually while upstream 100G EML chip prices have risen from $5 to $7, squeezing module-maker margins. Simultaneously, Google’s OCS demand is set to more than double to ~30,000 units in 2027E, with the incremental volume flowing almost entirely to outsourced suppliers Coherent and Lumentum. The market’s focus on transceiver unit counts misses this profit migration toward component suppliers and OCS contract manufacturers. Conversely, XPO—a high-density pluggable concept—is losing the architectural debate to CPO/NPO, making XPO-linked valuation narratives vulnerable.

市场可能低估了什么

The consensus view treats optical transceivers as a volume-growth story. Two critical dynamics are underappreciated. First, structural shortage in EML chips, isolators, and filters means 1.6T shipments in 2026E will reach only 15 million units against demand of 25 million-plus, and 800G shipments will undershoot demand by up to 5 million. The bottleneck is not transceiver assembly capacity but upstream optical chip supply—a condition that directly lifts pricing power for chip vendors. Second, the explosion in OCS unit demand—from ~18,000 in 2026E to ~40,000 in 2027E—creates a concentrated revenue tailwind for the two named contract manufacturers; Google’s insourcing covers only ~12,000 of the 2026E total, with the outsourcing share likely expanding further in 2027E as volumes scale.

证据链

Demand and shipment gap. 2026E global 800G demand is estimated at 45–50 million units; shipments are expected around 45 million. 1.6T demand exceeds 25 million units but shipments limited to 15 million. The shortfall is driven by shortages in EML chips, isolators, and filters. 2027E shows a similar pattern: 800G demand 60–70 million vs. shipments 55–60 million; 1.6T demand 40 million-plus vs. shipments ~30 million.

Customer concentration. 1.6T demand in 2026E is overwhelmingly concentrated: NVIDIA ~15 million units, Google ~10 million. 800G demand is more dispersed—Meta ~14 million, AWS ~7–8 million, Oracle ~7 million, NVIDIA and Google each ~5 million, Microsoft ~3–4 million. This concentration means the 1.6T supply bottleneck directly impacts the two most aggressive AI infrastructure builders, reinforcing their need to secure chip supply.

Price divergence. 800G DR EML transceivers are priced at ~$380 (down ~10% y-y); 1.6T DR SiPh transceivers at $1,000–1,100, with another 10%-plus decline expected in 2027E. In contrast, 100G EML chips have risen from $5 to ~$7, with further increases likely; 200G EML chips are $20–25. Transceiver price erosion coupled with rising component costs means module gross margins face a double squeeze—except for those with captive or preferential chip access.

OCS trajectory. Google’s OCS demand is estimated at ~18,000 units in 2026E, of which ~12,000 are in-house and the remainder outsourced to Coherent and Lumentum. By 2027E, total OCS demand is projected at ~40,000 units, with Google accounting for ~30,000. Oracle adds ~2,000 units in 2026E. The incremental 10,000-plus units from Google alone in 2027E will likely flow to external suppliers. NVIDIA’s next-gen dragonfly architecture may adopt a combination of OCS (for rack-to-rack layers) and CPO, opening a second demand vector for OCS beyond Google.

XPO relevance fades. The 12.8T XPO solution shown at OFC 2026 uses eight 1.6T transceivers, requires liquid cooling due to high power consumption, and is suited primarily for intra-rack use. CSPs, according to the expert, are favoring CPO and NPO over XPO for next-generation architectures.

关键分歧与风险

The primary tension is timing versus structural shift. A slowdown in cloud capex would reduce transceiver unit demand and relieve chip shortages, undermining the component pricing thesis. CPO adoption could displace pluggable transceivers faster than the market expects, compressing the addressable market for independent module makers earlier. Optical chip capacity—particularly 100G/200G EML—could expand faster than modeled, eroding the current pricing umbrella. Finally, if Google re-internalizes more OCS production as volumes scale, the outsourced revenue trajectory for Coherent and Lumentum would be cut. XPO-related names face an additional technical-route risk: the expert’s view that CSPs prefer CPO/NPO suggests limited large-scale deployment.

估值或交易含义

Investors should differentiate between optical module assemblers exposed to ASP deflation and component suppliers benefiting from shortage-driven pricing. 100G/200G EML chip vendors and filter/isolator manufacturers sit at the structural bottleneck and capture disproportionate profit growth. Within the OCS ecosystem, the ~4x expansion in Google’s total demand from 2026E to 2027E—with the outsourcing ratio likely stable or rising—directly benefits Coherent (COHR US) and Lumentum (LITE US). Avoid overpaying for XPO-narrative names; CSP preference for CPO/NPO undermines the addressable volume thesis. The transceiver demand cliff risk is real, but the nearer-term trading signal is to overweight upstream optical component scarcity and OCS contract manufacturing exposure.