SiPh vs. EML: 1.6T Transceiver Device Divergence

Introduction: The 1.6T Inflection Point in AI Networking

As the artificial intelligence (AI) arms race accelerates into 2026, the demand for bandwidth in hyperscale data centers has become insatiable. We have transitioned from 400G to 800G at breakneck speed—a cycle that used to take three to five years has compressed into eighteen months. Now, the industry stands at the precipice of the 1.6 Terabit Ethernet (1.6T) era. This transition is not merely a linear speed upgrade; it represents a fundamental divergence in optical device roadmaps that will define the next decade of networking hardware. The debate between Electro-absorption Modulated Lasers (EML) and Silicon Photonics (SiPh) has intensified, driven by the unforgiving physics of 200G per lane signaling and the brutal economics of AI cluster scaling.

For network architects, procurement leaders, and investors, the choice between EML and SiPh at 1.6T is no longer just about component cost—it is a strategic decision impacting power efficiency, supply chain resilience, and the long-term path to Co-Packaged Optics (CPO). As we witness the deployment of 51.2T switch ASICs (like Broadcom Tomahawk 5 and NVIDIA Spectrum-4) and the looming arrival of 102.4T chips, the optical interconnect has officially become the primary bottleneck in system performance. This article dissects the technical realities, supply chain dynamics, and performance trade-offs of these competing technologies as we move from 800G deployments to 1.6T mass adoption.

The Physics of 200G/Lane: Why the Roadmap Diverges

The leap to 1.6T is primarily enabled by increasing the single-lane rate from 100G (used in 800G modules like 2x400G or 8x100G) to 200G PAM4. This doubling of the baud rate places immense stress on the optical modulator, forcing a divergence in technology paths. At 100G per lane, the industry had multiple viable options (VCSEL, DML, EML, SiPh). At 200G, the herd thins dramatically.

EML: The Proven Incumbent for Performance

Electro-absorption Modulated Lasers (EML) have long been the workhorse of high-speed optics, particularly for reaches beyond 100 meters (DR/FR/LR).

• Mechanism: EMLs integrate a Distributed Feedback (DFB) laser and an Electro-Absorption Modulator (EAM) on a single Indium Phosphide (InP) chip. This monolithic integration allows for high-efficiency coupling and compact design. The DFB provides a stable continuous wave light, while the EAM acts as a fast shutter, modulating the light intensity.
• The 200G Advantage: At 200G per lane, the bandwidth requirements for the modulator are extreme. InP-based EAMs have inherently high bandwidth and high extinction ratios, making them the “safest” and most mature path to achieving 200G signaling with acceptable Bit Error Rates (BER) and Transmit Dispersion Eye Closure Quaternary (TDECQ) margins.
• Thermal Challenges: The catch is thermal management. EMLs are notoriously sensitive to temperature. As lane speeds increase, they require robust Thermoelectric Coolers (TEC) to maintain wavelength stability, which adds to the power budget—a critical factor when you pack eight of them into a standard OSFP module.
• Supply Chain & Cost: Manufacturing EMLs requires complex III-V semiconductor processing, which is less scalable and more expensive than silicon. The supply chain is highly concentrated among a few players like Lumentum, Coherent, and Broadcom (Avago). During the 400G and 800G cycles, EML shortages were a frequent bottleneck, leading to “allocation wars” between hyperscalers.

Silicon Photonics: The Integration Challenger

Silicon Photonics (SiPh) leverages the mature CMOS ecosystem to etch optical components (waveguides, modulators, detectors) onto silicon wafers, using external light sources.

• Mechanism: A high-power Continuous Wave (CW) laser (usually a DFB) is coupled into a silicon chip via a fiber or direct bonding. The light is then modulated using Mach-Zehnder Modulators (MZM) or Ring Modulators fabricated in silicon.
• The Integration Edge: The superpower of SiPh is integration. You can print thousands of optical components on a large wafer alongside control electronics. This makes it the natural platform for multi-channel scaling (e.g., 1.6T, 3.2T) where the component count for discrete EMLs becomes unmanageable.
• The 200G Hurdle: Modulating light at 200G/lane in pure silicon is challenging due to the material’s lower electro-optic coefficient compared to InP. Achieving the necessary bandwidth often requires complex segmented modulator designs or higher drive voltages, which can negate power savings.
• The Hybrid Solution: Innovations like Thin-Film Lithium Niobate (TFLN) on silicon are emerging to solve this. TFLN offers the best of both worlds: silicon’s manufacturing scale with Lithium Niobate’s superior modulation speed and low voltage, potentially outperforming both pure SiPh and EML at 200G.

The VCSEL Bottleneck: Why Short Reach is Dying

Historically, Vertical-Cavity Surface-Emitting Lasers (VCSELs) dominated the short-reach (SR) market (<100m) due to their incredibly low cost. However, VCSELs are hitting a wall at 100G/lane. Reliability issues (wear-out mechanisms) and bandwidth limitations make 200G VCSELs extremely difficult to produce with high yields. As a result, 1.6T may mark the death of the traditional “SR” reach, with single-mode solutions (SiPh/EML) cannibalizing the <100m market segment previously held by multimode fiber.

Architecture Battles: Pluggable vs. LPO vs. CPO

The device physics (EML vs. SiPh) directly influences the module architecture and the form factor wars.

Standard Pluggables (DSP-based)

Most initial 1.6T deployments (OSFP-XD, OSFP) will use standard DSP-based modules.

• EML Dominance: For early 1.6T-DR8 (500m) and 1.6T-2xFR4 (2km) variants, EMLs are the default choice due to their high launch power and signal quality.
• Power Penalty: The Digital Signal Processor (DSP) inside the module is necessary to clean up the 200G signals but consumes ~50% of the module’s total power. A fully loaded 1.6T module can draw 25-30W. In a 64-port switch, this translates to nearly 2kW of power just for the optics—a thermal nightmare for air-cooled racks.

The Rise of LPO (Linear Pluggable Optics)

Linear Pluggable Optics (LPO) is the industry’s attempt to “have its cake and eat it too.” It removes the power-hungry DSP from the optical module, relying instead on the specialized SerDes (Serializer/Deserializer) in the switch ASIC to drive the optics directly.

• The SiPh Connection: LPO strongly favors Silicon Photonics. SiPh modulators (specifically MZMs) have highly linear transfer functions, which makes them easier to drive directly with analog signals compared to the non-linear response of EMLs.
• Benefits: By removing the DSP, LPO reduces module power consumption by ~40-50% and cuts latency by ~100ns. For AI training clusters, where “all-to-all” collective operations are latency-bound, this is a massive performance unlock.
• Adoption: Major hyperscalers (like Meta and Google) are aggressively evaluating LPO to reduce the “optical tax” on their power budgets.

Co-Packaged Optics (CPO): The End Game?

CPO moves the optical engine off the front panel and onto the same substrate as the switch ASIC, connecting via ultra-short electrical links.

• SiPh is Mandatory: You simply cannot co-package hundreds of discrete EML lasers and lenses. SiPh is the only viable platform for CPO due to its reliability, yield, and thermal density.
• Status: While CPO promises the ultimate power efficiency (<5 pJ/bit), it remains a “next-generation” technology. The success of LPO and 1.6T pluggables has pushed the CPO timeline out to the 3.2T or 6.4T generations. However, for specific workloads like TPU pods, CPO-like implementations (like OCS) are already reality.

Supply Chain Map & Market Landscape

The technology divergence creates two distinct supply chains, each with its own winners and risks.

The EML Camp (Performance-First)

• Key Chip Suppliers:
  • Lumentum: The market leader in high-speed EMLs.
  • Coherent (formerly II-VI): Vertical integration from wafer to module.
  • Mitsubishi Electric: A key supplier for Japanese and global markets.
  • Broadcom: Produces high-performance EMLs for its own modules and select partners.
• Module Vendors: Coherent, InnoLight (high-end EML SKUs), Eoptolink.
• Supply Chain Risk: EML production is capital intensive and yield-sensitive. Capacity expansion takes 12-18 months. If NVIDIA or Google suddenly double their 1.6T orders, EML supply will tighten instantly, driving prices up.

The Silicon Photonics Camp (Integration-First)

• Key Chip Suppliers:
  • Intel: A pioneer in SiPh, though currently navigating foundry restructuring.
  • Broadcom: Leveraging its “Tomahawk” dominance to push SiPh-based interconnects.
  • Marvell (Inphi): Strong DSP + SiPh portfolio (via COLORZ).
  • Cisco (Acacia): Vertical integration for coherent and DC interconnects.
  • Foundries: TSMC (COUPE), GlobalFoundries (Fotonix), and Tower Semiconductor are enabling fabless startups to enter the market.
• Module Vendors: Intel, InnoLight (investing heavily in SiPh lines), and hyper-scale internal designs.
• Opportunity: The entry of TSMC into silicon photonics packaging (COUPE) is a game-changer. It signals that optical engines are becoming standard semiconductor components, potentially commoditizing the market and lowering the barrier to entry for new players.

Comparative Analysis: 800G to 1.6T Evolution

To understand the shift, we must look at the metrics.

Feature	800G Era (Current)	1.6T Era (Next-Gen)	Trend / Implication
:—	:—	:—	:—
Lane Signaling	100G PAM4	200G PAM4	Requires faster modulators; EML has edge, SiPh catching up.
Dominant Laser	EML (70% share)	EML (Early), SiPh (Ramping)	SiPh share will grow as LPO/CPO gain traction.
Form Factor	OSFP / QSFP-DD	OSFP-XD / OSFP	Modules getting larger and hotter.
Power Efficiency	~12-14 pJ/bit	Target <10 pJ/bit (via LPO)	Power is the primary constraint for AI clusters.
Reach Focus	SR8 / DR8 / FR4	DR8 / 2xFR4	<500m reaches dominate; SR (VCSEL) struggles at 200G.
DSP Role	Essential	Optional (LPO/CPO)	The “DSP-less” revolution begins at 1.6T.

Strategic Recommendations for Industry Players

1. For Hyperscalers: Diversify your optical portfolio. Do not rely solely on EML-based optics for 1.6T. Qualify SiPh-based LPO solutions for your “East-West” traffic inside AI clusters to save power and cost.
2. For Module Makers: Vertical integration is key. If you don’t own the chip (EML or SiPh), you are just a packaging house with thin margins. Acquiring or partnering with chip startups (especially in TFLN or novel modulators) is critical.
3. For Investors: Watch the CPO timeline. While pluggables are extending their life, the physics of copper reach and signal integrity will eventually force a switch to CPO. Companies like Broadcom, Marvell, and TSMC are best positioned for this long-term shift.

Conclusion

The 1.6T era forces a choice: pay the premium for EML performance today to guarantee 200G link budgets, or invest in the Silicon Photonics ecosystem for long-term scaling and cost reduction. For AI clusters where power is the new currency, the transition to SiPh-based LPO and eventually CPO seems inevitable. However, for the immediate 2026 deployment cycle, EML remains the king of reliability and reach. The roadmap has diverged, and the winner will be determined not just by who has the fastest laser, but by who can deliver the most bits per joule at the scale of a million-GPU cluster.