Industry Deep Dive: The Third AI Supercycle: Why Optical Interconnects Are the Next Generation's HBM

Executive Summary: The global artificial intelligence infrastructure build-out is aggressively shifting its bottleneck from compute capabilities to network bandwidth and power constraints. While the 2023-2024 cycle handsomely rewarded GPU manufacturers and High Bandwidth Memory (HBM) pure-plays, the structural reality of 100,000+ GPU clusters exposes the terminal limits of copper interconnects. Based on our synthesis of recent industry data from the OFC 2026 conference, optical interconnects—specifically Co-Packaged Optics (CPO) and Hollow-Core Fiber (HCF)—represent the critical third wave of the AI investment cycle. We project the optical component Total Addressable Market (TAM) will expand by 3x to 5x heading into the 1.6T upgrade cycle in 2027-2028, fundamentally rerating the optical hardware supply chain.

Strategist's Core View

• Macro Driver: The exponential scaling of AI distributed training networks demands network bandwidth that copper cannot physically provide without breaking data center power density limits. Optics offer a 65% power reduction and are transitioning from long-haul telecom to chip-to-chip (Scale-Up) environments.
• Top Sector Pick: Small-cap precision optical components and CPO assembly/testing equipment. We favor domestic players exposed to Silicon Photonics (SiPh) alignment and external laser source packaging.
• Key Risk: A delayed transition to the 1.6T standard or thermal management failures in early CPO deployments could compress multiples for early-stage component suppliers.

The Macro Landscape: The Physical "Wall" of Copper and Power Density

To understand the urgency of the optical transition, investors must recognize the severe power and bandwidth limitations currently strangling hyper-scale data centers. The infrastructure that connects AI servers relies largely on electrical transceivers and copper Direct Attach Cables (DAC), utilizing standards designed over a decade ago for cloud computing, not localized heavy-compute AI workloads.

We are hitting a physical wall. Copper interconnects consume upwards of 10 picojoules per bit (pJ/bit), pushing interconnect power draw to an unsustainable 30% to 40% of total data center power. Furthermore, at 200Gbps per lane (the requirement for next-generation networks), copper signal attenuation restricts cable lengths to a mere 2 to 3 meters. In a 72-to-256 node GPU cluster environment, a 3-meter reach is physically unworkable. The signal degradation forces the use of power-hungry retimers (drawing 7-15W each) just to maintain signal integrity.

Optical interconnects resolve this exact bottleneck. By converting data to photons, power consumption drops to 1-3 pJ/bit. More importantly, it completely bypasses the electromagnetic interference (EMI) and density constraints of copper cables. The strategic takeaway is absolute: the transition from copper to optics is not an elective upgrade; it is a physics-mandated necessity for AI scaling.

Interconnect Generation	Data Rate (Total)	Lane Speed & Modulation	Copper Limitation	Deployment Phase
Current	400G	100G (PAM4)	Max 5-7 meters reach	2023-2024
Inflection (Now)	800G	100G x 8 / 200G x 4	Impossible at 200G beyond 3m	2025-2026
Next-Gen (CPO Era)	1.6T	200G x 8 (PAM4)	Copper physically unviable	2027-2028
Future State	3.2T	400G (PAM8/10)	Requires Silicon Photonics	2029+

Strategic Focus: Winning Sectors & Stock Picks

The institutional capital allocation playbook must bifurcate the optical theme into three distinct vectors: Scale-Out, Scale-Across, and Scale-Up. Each vector operates on a different commercialization timeline and offers distinct alpha-generation opportunities.

1. Scale-Out (Server-to-Server) & Scale-Across (DC-to-DC)

In the immediate term (2025-2026), the transition is already happening in the Scale-Out and Scale-Across domains. Hyper-scalers are facing an acute power density crisis. Because building 1GW+ data centers in a single location is increasingly unfeasible, companies must distribute AI training across multiple physical sites located tens of kilometers apart. This demands extreme-low-latency Scale-Across solutions.

Hollow-Core Fiber (HCF) is proving to be a game-changer here. By transmitting light through air rather than glass, HCF reduces signal attenuation by 50% (to 0.1 dB/km) and cuts amplification power requirements by 35-40%. Recently, a 600km distributed training run for a Large Language Model (LLM) was successfully demonstrated, effectively proving that data centers spread across the distance of Seoul to Busan can operate as a single logical AI cluster.

Investment Implication: Domestic fiber optic manufacturers and 800G optical transceiver makers are the frontline beneficiaries. Local strategy estimates identify companies like Opticore (800G optical transceivers) and Taihan Fiberoptics (cables) as primary recipients of this immediate CAPEX wave. Furthermore, network equipment providers like Solid are well-positioned for the scale-across DCI (Data Center Interconnect) infrastructure build-out.

2. Scale-Up (Chip-to-Chip) and Co-Packaged Optics (CPO)

The true structural alpha lies in the 2027-2028 timeframe with the commercialization of Co-Packaged Optics (CPO). Currently, optical modules are "pluggable"—inserted into the front panel of a switch. The electrical signal must travel 10-30cm across a PCB trace from the ASIC to the module, burning massive amounts of power. CPO integrates the optical engine directly onto the switch chip's substrate, reducing the distance to mere millimeters. This architectural shift eliminates power-heavy components like retimers and DSPs, yielding a verified 65% reduction in power.

Recent demonstrations of Broadcom's Tomahawk 6 CPO switch (102.4 Tbps capacity) and Meta's exhaustive 36-million device-hour reliability testing confirm that CPO is moving from prototype to production.

Investment Implication: As the industry adopts standard OIF CPO frameworks, demand for precision component manufacturing and automated alignment tools will skyrocket. The alignment tolerance for these micro-optical connections is incredibly tight (±25µm). Domestic unlisted and small-cap players such as ADST (CPO alignment equipment) and MPNICS (micro lenses) stand to see explosive earnings revisions as their TAM expands. Additionally, RF Materials, which is localizing laser diode components, provides an attractive angle into the External Laser Source (ELS) supply chain.

Valuation Reality Check & Fair Price Assessment

When assessing the domestic optics sector, we must strip away the AI hype and look at the underlying earnings mechanics. Currently, the market is pricing Korean optical transceiver and component stocks largely based on legacy telecom infrastructure multiples (P/E 10x - 15x). However, the shift into the core AI computing stack justifies a severe multiple expansion. As these companies transition from being categorized as "telecom equipment" to "AI semiconductor infrastructure," they should command multiples closer to the semiconductor backend and HBM supply chain (P/E 20x - 30x).

Analyst J's Verdict

While the market consensus expects a gradual, linear earnings growth for the domestic optical supply chain throughout 2025, we believe this is highly conservative. The consensus fails to price in the non-linear inflection point occurring in 2027 with the 1.6T standard and broad CPO adoption. We view current dips as a structural accumulation zone. A fair valuation framework for high-purity CPO equipment and micro-lens component players should apply a 25x forward P/E on 2026 estimates, reflecting the premium commanded by bottleneck-breakers in the AI hardware stack.

Sub-Sector	Primary Driver	Historical P/E Band	Analyst J Target P/E	Valuation Justification
Optical Transceivers (800G)	Scale-Out deployment in hyper-scale DC	12x - 15x	18x - 20x	Volume ramp imminent, but ultimate terminal value capped by CPO transition.
CPO Equipment & Test Sockets	1.6T CPO Standardization & Yield optimization	N/A (Nascent)	25x - 30x	High barriers to entry; operates identically to semiconductor advanced packaging capex logic.
Micro-lenses & Specialty Components	Silicon Photonics vertical coupling	15x - 18x	22x - 25x	Critical bottleneck for mass production yields; strong pricing power.

Key Risks & Downside Scenarios

No secular theme is without severe execution risks. For the optical interconnect supercycle, investors must monitor three critical friction points:

1. Thermal Management Failures: Co-Packaged Optics places highly temperature-sensitive laser components mere millimeters from ASICs generating massive heat (350-400 W/cm²). Semiconductor lasers degrade exponentially with heat; a 10°C rise cuts the lifespan in half. If flip-chip VCSEL designs or external laser sources (ELS) fail to maintain junction temperatures below 65°C, hyperscalers will stall deployment due to reliability concerns.
2. Silicon Photonics (SiPh) Foundry Bottlenecks: The economics of optical components rely entirely on the ability of foundries (like imec, GlobalFoundries, and TSMC) to mass-produce 300mm SiPh wafers at high yields. Any delay in CMOS-compatible SiPh scaling will keep unit costs too high to displace copper entirely in the Scale-Up domains.
3. The "THz Bridge" Threat: We note the emergence of Terahertz (THz) interconnects offering <3 pJ/bit efficiencies over 50-100m using standard ASIC manufacturing processes. If THz development accelerates rapidly, it could carve out a significant portion of the intermediate data center market, slightly delaying the full optical overhaul for mid-range connectivity.

Strategic Outlook & Actionable Advice

The market has clearly defined the AI hardware baton pass: GPUs in 2023, HBM in 2024, and Interconnects in 2025 and beyond. With AI accelerator architectures demanding up to 7.2 TB/s of bandwidth (e.g., the Rubin generation), the physics of copper are broken.

Institutional portfolios should strategically overweight the optical hardware supply chain immediately. Focus capital allocation on companies providing the fundamental building blocks of the CPO era—specifically those involved in Silicon Photonics integration, alignment testing (such as TFE's CPO test sockets), and high-speed modulation. The 800G cycle provides the cash flow floor today, but the 1.6T CPO integration in 2027 is the multi-bagger catalyst that will define the winners of the next decade's infrastructure race.

---

Disclaimer: The information provided in this article is for informational and educational purposes only and does not constitute financial, investment, or trading advice. Investing in the stock market involves risk, including the loss of principal. All investment decisions are solely the responsibility of the individual investor. Please consult with a certified financial advisor and conduct your own due diligence before making any investment decisions.