Last Updated on 2026 年 3 月 23 日 by 総合編集組
The AI-Driven Optical Revolution: From 800G to 1.6T and CPO in Data Centers – Key Insights for 2026 and Beyond
The rapid rise of generative AI and large language models is transforming global data center infrastructure. Traditional copper-based electrical interconnects face severe limitations at speeds beyond 200 Gbps, including high power loss, latency, and the so-called “I/O wall.” Optical communication technologies are emerging as the critical solution, shifting from electrical to photonic pathways for high-bandwidth, low-latency, and energy-efficient connections.
Core Technologies in Semiconductor Optical Communication Optical transceivers are classified into three main categories based on integration level and drive method: traditional pluggable optics (DSP-based), linear-drive pluggable optics (LPO), and co-packaged optics (CPO).
Pluggable optics remain the most deployed solution in data centers due to hot-swappable flexibility and standardized supply chains. These modules integrate digital signal processors (DSP), laser drivers, transimpedance amplifiers (TIA), and electro-absorption modulated lasers (EML). In the 800G era, QSFP-DD and OSFP form factors dominate, with OSFP preferred for better thermal performance toward 1.6T and higher rates. DSP handles forward error correction (FEC) and equalization but consumes over 50% of module power (typically 14–20 W for 800G).
LPO addresses AI’s demand for ultra-low latency and power by eliminating the DSP inside the module, shifting compensation tasks to the switch ASIC’s SerDes. This reduces 800G power to 5–8 W and latency below 3 ns, ideal for synchronized AI training clusters. However, LPO limits reach to under 2 km and faces interoperability challenges across vendors.
CPO represents the long-term direction, integrating optical engines directly with ASICs or GPUs in the same package to eliminate long electrical traces on PCBs. For 1.6T links, CPO can cut power from ~30 W to ~9 W (a ~70% reduction), enabling massive GPU clusters with hundreds of thousands of units. NVIDIA and Broadcom have positioned CPO as essential for future “AI factories.”
Key Modulation and Material Breakthroughs High-speed scaling relies on PAM4 modulation and silicon photonics. PAM4 doubles bandwidth per clock cycle by encoding 2 bits across four voltage levels, replacing NRZ used in 10G/25G eras. While sensitive to noise, DSP enables precise equalization to maintain signal integrity.
Silicon photonics leverages CMOS processes to fabricate modulators, detectors, and waveguides on silicon wafers, enabling mass production and cost advantages over traditional InP materials. It already powers cost-effective 800G DR8 modules and is evolving with thin-film lithium niobate (TFLN) for 200G–400G per lane.
Major U.S.-Listed Players and Supply Chain Dynamics NVIDIA (NVDA) leads AI networking via Spectrum-X Ethernet and Quantum-X InfiniBand platforms, demonstrating CPO switches with up to 3.5x energy efficiency gains. It plans multibillion-dollar investments in laser suppliers for 1.6T supply security.
Broadcom (AVGO) dominates switch ASICs with Tomahawk 6 offering 102.4 Tbps capacity and 64×1.6T ports with embedded CPO. About 40% of its AI revenue stems from networking interconnects, bolstered by open standards like Ultra Ethernet Consortium.
Marvell (MRVL) holds ~70% share in 800G DSPs with its Spica series, securing full Google 1.6T DSP demand and leading in LPO. It offers attractive valuation relative to peers amid growing low-power needs.
Coherent (COHR) and Lumentum (LITE) supply critical EML and CW lasers. NVIDIA’s $2 billion investment each (total $4 billion) secures capacity amid shortages, with delivery lead times exceeding 24 weeks for high-end EMLs. Coherent benefits from scale and vertical integration; Lumentum excels in 200G/lane performance.
Applied Optoelectronics (AAOI) gains traction with hyperscale 800G orders and Texas facility expansion. Cisco (CSCO) advances via Silicon One chips and acquisitions, while Astera Labs (ALAB) supports internal server connectivity bottlenecks.
Market Demand and Future Projections AI model training doubles roughly every 18 months, driving faster optical upgrades. 800G shipments are forecast to surge ~60% in 2025, with 1.6T becoming standard for hyperscalers (AWS, Google, Meta) in 2026.
Silicon photonics and CPO markets show strong growth: silicon photonics reaches ~$3.6B by 2026, while CPO expands from early adoption to $9B+ by 2030 and potentially $20B+ by 2036 (various forecasts indicate 30–37% CAGR). EML shortages push adoption of CW lasers and silicon photonics for better yield and supply stability.
DSP Mechanisms and Performance Metrics DSP performs equalization (compensating chromatic and polarization mode dispersion via LMS/DFE algorithms) and FEC (oFEC/CFEC) to handle high BER in PAM4 systems without retransmission—vital for uninterrupted AI training.
Community Sentiment and Investment Considerations Forum discussions highlight Broadcom’s structural moat in networking chips versus valuation and customer concentration risks. Coherent and Lumentum enjoy bullish sentiment from NVIDIA backing, likened to “picks and shovels” in the AI gold rush, though some caution over high P/E reflecting optimistic expectations.
Conclusion: Optics as AI’s Second Battlefield Bandwidth upgrades are non-negotiable. 2025 marks peak 800G ramp-up, while 2026 heralds 1.6T and CPO commercialization. Chip vendors like Broadcom and Marvell capture high margins via DSP/ASIC dominance; upstream laser suppliers benefit from scarcity. CPO’s power savings position it as the path to million-GPU clusters. This photonic shift redefines semiconductor investment logic in the AI era.
