AI Capital Cycle ⑤ — The Next Bottleneck: Optical & Networking

The 1849 California Gold Rush. Who got rich? *Not the miners*. Most miners went home poor. The rich were the ones who *sold picks and work clothes to the miners*. Levi Strauss began making denim work clothes in San Francisco in 1853; his 1873-patented jeans became the single standard of American fashion. *On all those miners' labor — those who sold picks and jeans became the real wealth of the cycle*.

This is the repeating pattern of capital cycles. In the 1860-80s US railroad cycle, the rich weren't the railroad companies — they were *the people who bought land alongside the rails, and Carnegie who made the steel the railroads needed*. The 1920s auto-cycle rich were not only Ford and GM but also *Goodyear who made the tires, and the cement companies that paved the roads*. The 1990s-2000s internet-cycle rich were not Pets.com or Webvan but *Corning who laid the fiber, Cisco who sold the routers* (and short-term, Lucent too). In oil cycles, as bottleneck moved from refining → pipelines → gas stations, different people got rich at each stage.

The bottleneck moves. At every stage of the cycle, *whoever holds the next bottleneck* enjoys the largest revenue growth.

In the 2026 AI capital cycle, GPU (NVIDIA) was the protagonist of the last stage. But *before* the GPU's absolute revenue curve slows, the cycle's capital has already begun moving to the next bottleneck. Optical.

Structural reason. Modern AI training clusters *outgrow the rack*. NVIDIA Blackwell NVL72 packs 72 GPUs into one rack with 1.8 TB/s NVLink (about 14x PCIe Gen5). But frontier model training has *thousands of GPUs* communicating *across many racks*. Inter-rack communication runs on — *optics*.

Jensen Huang at GTC 2025: "When the data centers are now the size of a stadium, we need something new — and that's where silicon photonics comes in."

Two distinct networks:
- Scale-up (inside rack): NVLink/NVSwitch — currently copper-dominated, but moving to optical/CPO (co-packaged optics) as GPU counts per domain grow.
- Scale-out (between racks): InfiniBand or Ethernet — *already optical*. Drives 800G → 1.6T transceiver demand.

Why supply is constrained. EML (electro-absorption modulated laser) chips 40-60% undersupplied through 2027 (per McKinsey analysis). InP wafer capacity is the single chokepoint. DSP supply: Broadcom holds >80% share of the high-end PAM4 DSPs used in 800G/1.6T modules; Marvell (Inphi acquisition) is the only credible #2.

One 1.6T transceiver requires ~8 EML chips + 1 DSP + advanced silicon photonics packaging. *Every Blackwell GPU pulls roughly 1.5-3 optical ports*. When NVIDIA ships millions of GPUs, optical transceiver demand is 1.5-3x that.

LightCounting forecast: AI cluster optics market grows $5B (2024) → $10B+ (2026), $100B market by 2030. 800G shipments +60% in 2025 (~20M units). 1.6T expected >5M units in 2026 — *the volume ramp year*.

The strongest data showing the cycle is moving is — *the revenue acceleration of the beneficiaries*.

Lumentum (LITE). Optical components and transceivers. Acquired Cloud Light for $750M in Nov 2023 — direct 800G/1.6T transceiver supply chain. Customers: NVIDIA, Google, Meta, MSFT, AMZN.
- FY25 (ended June 2025): revenue $1.65B
- Q1 FY26: $533.8M (+58% YoY)
- Q2 FY26 (Dec 2025): $665.5M (+65.5% YoY), NG-GM 42.5% (+1,020bps YoY), NG-OM 25.2%
- *AI+cloud revenue mix 60%+*
- Prior guidance "~$600M Q FY26" hit *two quarters early*

Coherent (COHR). Post-II-VI/Coherent merger, 70%+ revenue from data center and comms. Vertical from InP wafer to transceiver.
- Q1 FY26: $1.58B (+19% YoY pro-forma)
- Q2 FY26: $1.70B (+17.5% YoY)
- First 1.6T transceiver revenue: Q4 FY25 (June 2025 quarter)
- CEO Jim Anderson: "1.6T modules ramping *much faster* than expected a year ago. Orders extend through 2028."
- Capacity shift: 3-inch → 6-inch InP wafers — *4x chips per wafer, half the cost*

Astera Labs (ALAB). PCIe/CXL retimers, fabric switches. IPO March 2024. The "connectivity glue" of AI servers.
- 2024 revenue +242% YoY
- Q3 2025: $230.6M (+104% YoY), NG-OM 41.7%
- Aries 6 (PCIe Gen6 retimer) — >20% of Q3 revenue. The only Gen6 retimer in volume
- Scorpio (fabric switch) — >10% of Q2 2025 revenue, fastest ramp in company history
- May 2025 NVIDIA NVLink Fusion integration — hyperscalers can build custom XPUs that talk directly to NVIDIA GPUs via NVLink using Astera silicon
- Market cap $55-56B (May 2026), trading 50-60x forward sales — *premium proxy on the bottleneck rotation thesis*

Marvell (MRVL). Two pillars: custom AI silicon (AWS Trainium 2/3, MS Maia) + optical DSP (Inphi heritage, Nova 2 = industry's first 1.6T DSP).
- FY26 revenue ~$8.2B (+42% YoY); data center revenue >$6B (+46%)
- FY27 guidance: revenue >+30% to ~$11B; data center +40%
- FY28 guidance: ~$15B; "second tier-1 hyperscaler XPU ramping"
- Broadcom + Marvell = ~95% of custom ASIC market

Arista (ANET). The Ethernet-vs-InfiniBand bet. Leads Ultra Ethernet Consortium.
- FY25 revenue $9.0B (+28.6%), first-ever $1B quarterly net income in Q4
- FY26 AI networking target doubled to $3.25B (from prior $1.5B)
- Meta + Microsoft = 40-48% of revenue
- Meta chose Arista 7700R4 for its largest Ethernet AI cluster

If optical is the first next-bottleneck, CXL (Compute Express Link) and PCIe Gen6/Gen7 retimers are the second.

Structural reason. Inside AI servers, GPU-CPU-memory-storage communicate over PCIe. Going from PCIe Gen5 (32 Gb/s) to Gen6 (64 Gb/s) halves the reachable copper trace distance. *Without a retimer, the signal doesn't reach far enough*. Astera Labs' Aries 6 is the only Gen6 retimer in volume production.

CXL is memory pooling/sharing technology. Multiple servers share memory pools. CXL 1.0 (2019) → 2.0 (2020) → 3.0 (2022) → 3.1 (2024). CXL 4.0 announced Nov 18, 2025, running on PCIe Gen7. From 2025, 90%+ of new servers are CXL-capable. *2026 is the first real deployment year of multi-rack memory pools*.

Why this matters: when CXL works — *the entire data center's memory can be shared as a pool*. HBM is expensive and limited. If CXL lets GPUs access external DRAM, training memory constraints relax. *Astera makes the silicon that enables this*.

So Astera's revenue trajectory is the direct proxy for "how fast AI's second bottleneck is unclogging."

Honestly — *Korea has weak direct beta in optical*. Korea's semiconductor DNA is memory (DRAM/NAND/HBM), not III-V (InP/GaAs). EML/InP supply chains favor Japan, US, Europe (Mitsubishi, Sumitomo, Lumentum, Coherent). Korean fabs (SK siltron, DB HiTek) lack InP capacity.

Korean listed optical names (mostly small caps):
- OE Solutions (138080) — Korea's biggest datacom transceiver pure-play. 2024 revenue -30.4% (3-inch InP cycle bottom) → 2025 revenue +79.2% to ~₩57.3B (800G/1.6T pull-in). 1.6T+CPO roadmap announced May 2026.
- Opticore — Dec 2025: ₩4.0B 400G/800G AI DC transceiver supply contract (17.2% of prior year revenue).
- BNE Technology — 800G/1.6T qualification ending; revenue inflection expected H2 2026.
- Lightcom — optical amplifiers, fiber laser. Less direct AI DC exposure.
- RFHIC — GaN PA + photonics package facility investment announced (May 2026).

*But Korea's real win is not optical — it's HBM* (see Memo 4). Korean investors wanting optical exposure: honest answer is to buy LITE/COHR/ALAB/MRVL/ANET directly in the US market.

Two *indirect* paths by which Korea benefits from optical:
1. For SK Hynix HBM to sell more — optical must unclog (if optical bottlenecks, data center clusters can't expand → HBM demand softens).
2. Samsung Silicon Photonics 2028 mass production roadmap — 2027 launch / 2028 mass production. Samsung showed CMM-D (CXL 3.1) memory module at OCP Global Summit Oct 2025. If successful, Korea creates a direct optical beta 5 years out.

For now — *Korean optical beta is most accurately viewed as a function of SK Hynix HBM*.

The single proposition of this memo: at the end of a cycle, the rich aren't the protagonists — they're the ones holding the next bottleneck.

In 1849 the rich were Levi Strauss, not the miners. In the 1990s-2000s internet, the rich were Corning and Cisco, not Pets.com (short-term Lucent too — see Memo 1). In 2026 AI — the next rich are *already being decided*. The indicators show.

*First*, optical transceivers. Lumentum, Coherent — both ramping faster than expected a year ago. The 800G → 1.6T transition is the 2026-27 core. Coherent CEO's "orders extend through 2028" signals 4-5 years of cycle visibility.

*Second*, CXL/PCIe retimers. Astera Labs is the only Gen6 retimer in volume. Quarterly revenue growth +100%+. NVIDIA NVLink Fusion integration makes Astera the key silicon for "hyperscaler custom XPU + NVDA GPU hybrid."

*Third*, custom ASIC silicon. Marvell + Broadcom = 95% combined share. Marvell guides FY26 $6B → FY28 $15B ("second tier-1 XPU entering").

*Fourth*, Ethernet networking. Arista leads Ultra Ethernet Consortium, doubled AI networking guidance. The next stage of the architectural war between NVIDIA InfiniBand and Ethernet.

These four beneficiaries carry *less valuation burden than NVDA* — and can ride the cycle *longer* than NVDA. *The optical cycle lags the GPU cycle by 2-4 quarters*. Even after GPU peaks, optical can ramp for those quarters (Coherent's "orders through 2028" is the evidence).

From the Korean view: direct exposure is weak — but *SK Hynix HBM revenue is a function of the optical cycle*. If optical bottlenecks, clusters can't expand, and HBM doesn't sell. So Korean investors should also — *use Lumentum/Coherent/Astera quarterly revenue as a leading indicator for SK Hynix*.

*If NVIDIA was the NVIDIA of the GPU cycle, Lumentum may become the NVIDIA of the optical cycle. Just — before the GPU cycle stops.*