By Martin Sharratt, Managing Director, AP Technologies

 

Meeting the optical power demands of next-generation AI data centre interconnects requires a fundamentally different class of high-performance light source and optoelectronic solution. This article examines the ultra-high-power InP device platforms available through AP Technologies and the practical integration considerations for UK OEM teams working at the forefront of co-packaged optical engine and silicon PIC (photonic integrated circuit) design.

 

Key Takeaways

• AI data centre scaling is simultaneously increasing bandwidth density requirements and minimum acceptable optical output power.
• Copper interconnects are reaching practical limits at data rates of 800G and above.
• High-power O-band DFB lasers deliver up to 250mW continuous wave today, with CWDM channel expansion on track for Q3 this year and a roadmap targeting 400mW and 1W-class platforms.
• Ultra-high-power SOAs provide greater than 1W output with approximately 25% power conversion efficiency at 25°C and 20% at 50°C.
• A broad gain bandwidth of approximately 80nm enables scalable WDM and multi-channel architectures.
• InP RSOA gain chips support external cavity laser architectures and direct integration with silicon photonic PICs, with device customisation available to specific mode-field and geometry requirements.
• Early-stage sampling and structured engineering collaboration significantly reduce integration timescales and qualification risk for OEMs.

 

 

AI infrastructure is scaling faster than the physical layer was designed to support. GPU clusters have grown from occupying a few racks to filling entire data centre halls and every step up in power density demands a proportional increase in the bandwidth, reach and reliability of the optical links connecting them. Copper interconnects cannot sustain signal integrity at 800G and above across the distances modern AI clusters demand. As optical engines migrate from front-panel pluggables to co-packaged configurations sitting directly adjacent to switch silicon, power budgets tighten, thermal envelopes shrink and the small form factor requirements of these designs become considerably more demanding. Standard transceiver-grade sources were not engineered for this environment.

 

The solution lies in a new generation of indium phosphide (InP) semiconductor devices built specifically for high-power, high-efficiency operation in these conditions. AP Technologies provides direct access to SemiNex's ultra-high-power InP range in bare chip, chip-on-carrier and fibre-coupled packages, with pre-release prototype support and structured engineering collaboration to accelerate the path from first sample to qualified production.

 

The Optical Power Imperative

Higher signalling rates shrink insertion loss headroom across every connector, trace and interposer in the link. Higher optical launch power is necessary to maintain signal-to-noise ratios while preserving pJ/bit efficiency targets. However, increasing optical power without improving power conversion efficiency simply moves the problem from bandwidth to heat. In rack environments operating at 100kW and above, that is a real constraint.

 

Three complementary InP device architectures address this directly:

• High-power O-band DFB lasers deliver stable continuous-wave sources up to 250mW today, with a clear roadmap to 1W-class operation.
• Ultra-high-power semiconductor optical amplifiers (SOAs) extend optical margin beyond 1W while maintaining practical efficiency at elevated temperatures.
• Reflective semiconductor optical amplifiers (RSOAs) gain chips enable external cavity and silicon photonic PIC integration for coherent and tunable architectures.

 

High-Power O-Band DFB Lasers

The O-band dominates intra-data-centre interconnects due to its low chromatic dispersion in standard single-mode fibre and its mature ecosystem of passive WDM components. Current high-power DFB platforms deliver up to 250mW continuous wave (roughly three to five times the output of the previous generation of data centre sources) while maintaining single-mode emission, side-mode suppression exceeding 40dB and narrow spectral linewidth across this elevated power range.

 

SemiNex's current O-band DFB portfolio spans two production platforms:

• Model 1, rated at 200mW at 25°C and over 150mW at 50°C.
• Model 2, rated at 250mW at 25°C and over 200mW at 50°C.

 

The temperature-qualified figures are the ones that matter in practice as they reflect conditions inside real switch and server chassis rather than laboratory baselines. The 250mW DFB is on track for Telcordia qualification by Q4 of the 2026 fiscal year. Both models are available now in bare chip and chip-on-carrier formats suited to small form factor OEM designs, with a 14-pin butterfly package in development.

 

The current portfolio covers the 1310 and 1311nm O-band channels, with 1271, 1291 and 1331nm CWDM channels expected in Q3 2026. The roadmap extends further to a 400mW device at 50°C in advanced development and a 1W-class target platform beyond that. For OEM integrators designing products with three-to-five year lifespans, the visibility of a compatible higher-power successor within the product life cycle is a material engineering consideration. AP Technologies facilitates direct roadmap engagement so that architecture decisions are based on current manufacturer intelligence rather than published datasheet information alone.

 

Ultra-High-Power Semiconductor Optical Amplifiers (SOAs)

 

In a high-radix co-packaged switch with hundreds of optical ports, the amplification architecture is one of the primary levers on total system power consumption, not a peripheral design detail. Driving each port from its own laser source scales power draw directly with port count. Distributing a single high-power source across a passive splitter network and restoring margin at each branch with an SOA breaks that relationship entirely, and in facilities where power costs are measured in megawatts, the difference is substantial.

 

The SOA platforms available through AP Technologies deliver output in excess of 1W, well above the 100–300mW typical of the broader SOA market, but raw output power is only part of the story. Power conversion efficiency determines how much of that electrical input becomes useful light and how much becomes a problem. At 25% efficiency at 25°C and 20% at 50°C, these devices remain productive amplifiers under the thermal conditions that actually exist inside production chassis. Inefficiency does not disappear but becomes heat, deposited inside a chassis already running close to its thermal ceiling. The 80nm gain bandwidth across the O-band accommodates full WDM channel spacing without wavelength-selective amplification, removing component count and eliminating sensitivity to upstream laser wavelength variance in a single design decision.

 

Where the SOA sits mid-link rather than as a final-stage booster, noise figure and polarisation dependence move from secondary to critical parameters, since noise accumulated at an intermediate stage compounds through everything that follows. Full characterisation across both operating temperature points gives link designers real numbers to work with rather than forcing conservative derating to cover gaps in the data. SOA chips and chip-on-carriers are available for sampling today, with Telcordia qualification in progress.

 

RSOA Gain Chips for Silicon Photonic Integration

Silicon photonics can do many things well, but generating light efficiently is not one of them. The indirect bandgap of silicon makes it a poor gain medium, and that fundamental materials constraint is not going to be engineered away. The RSOA is how that gap gets bridged.

 

In a hybrid external cavity configuration, the indium phosphide-based RSOA provides the gain while a silicon photonic chip supplies wavelength-selective feedback through gratings, ring resonators or MEMS elements manufacturable at wafer scale. Each material contributes what the other cannot. InP delivers high gain and high output power. Silicon delivers the cost structure, integration density and manufacturing scalability that AI optical engine programmes require. The combination is more capable than either platform alone and increasingly it is the architecture that production-ready co-packaged designs are converging on.

 

For coherent links, the case is more specific still. As the growth of AI clusters pushes interconnect distances from metres to hundreds of metres, coherent transmission at those distances requires narrow-linewidth sources with the phase noise characteristics that coherent DSP engines can actually work with. Using an RSOA enables external cavity laser configurations to meet that specification without requiring a populated array of fixed-wavelength gain chips, since the 80nm gain bandwidth supports multi-channel and tunable operation from a single device.

 

Integrating InP gain chips with silicon PICs demands sub-micron alignment and mechanically stable, low-thermal-resistance bonding across the full operating temperature range. AP Technologies leverages support from SemiNex for integrators working through this to bring flexibility to mode-field design, facet geometry and layer thickness to match the specific requirements of a given PIC platform.

 

Wrap Up

The optical power requirements of AI data centre interconnects now exceed what transceiver-grade devices were designed to address. High-power O-band DFBs, ultra-high-power SOAs and RSOA gain chips form a technically mature InP family that meets those demands today, with a roadmap that tracks the trajectory of AI infrastructure scaling. UK OEM integrators are invited to contact AP Technologies to discuss sampling, application discussions and roadmap planning.

 

For further information on SOAs and RSOAs, including specification details, please visit the product page. Further details on the complete SemiNex product range available from AP Technologies is available on the dedicated supplier page.

 

Working with AP Technologies

AP Technologies provides UK OEMs with early access to engineering samples across all three platforms, typically ahead of general availability, and draws on direct engagement with SemiNex and accumulated experience of hybrid integration programmes across the UK photonics industry. Where device customisation is required, AP Technologies engages with SemiNex directly on the customer's behalf. Integration work beginning at sample stage can reach qualification maturity by the time production volumes arrive, a meaningful advantage in a market where AI deployment timelines leave little room for sequential iteration.

 

Frequently Asked Questions

Q: Why is the O-band preferred over the C-band for AI data centre interconnects?

A: The O-band exhibits near-zero chromatic dispersion in standard single-mode fibre at the reach distances typical within a data centre campus, eliminating the need for dispersion compensation at high signalling rates. The C-band remains dominant in long-haul applications where its low attenuation and mature EDFA ecosystem are decisive advantages, but within the data centre the O-band's dispersion characteristics are the more important factor.

Q: What is the difference between a DFB laser and an SOA in this context?

A: A DFB laser is a self-contained source that generates light at a specific stable wavelength; an SOA takes an existing optical signal as input and increases its power. In a practical AI interconnect architecture, a DFB typically seeds the link while an SOA compensates for splitting loss or insertion loss budget erosion downstream.

Q: What does power conversion efficiency mean for an SOA and why does it matter?

A: Power conversion efficiency is the ratio of optical output power to total electrical input power consumed. At 25% efficiency, a device producing 1W of optical output draws 4W of electrical power with the remaining 3W dissipated as heat. In a thermally constrained compute enclosure this has a direct and measurable cost in cooling infrastructure.

Q: What is an external cavity laser and why use an RSOA rather than a standard DFB?

A: An external cavity laser places the wavelength-selective feedback element outside the gain chip, enabling wavelength tuning using silicon photonic components manufactured at wafer scale. A standard DFB has its grating fixed at fabrication; the RSOA leaves wavelength control to the external cavity, enabling tunability and PIC integration that a fixed-wavelength DFB cannot provide.

Q: How long does a typical integration programme take from first sample to qualified production?

A: A co-packaged module using a characterised DFB in an established package format might reach initial qualification within nine to twelve months of first sample receipt; a hybrid silicon photonic integration involving an RSOA may require eighteen to twenty-four months. Early sample access through AP Technologies compresses whichever timeline applies by starting integration work ahead of general commercial release.

Q: Can these devices be used outside AI data centre applications?

A: High-power O-band DFBs and SOAs are applicable to high-capacity metro access networks, lidar and optical sensing. RSOA gain chips are used across coherent telecommunications, optical fibre gyroscopes and spectroscopy instrumentation. The AI data centre is currently the most demanding and fastest-growing market, but all three platforms are general-purpose InP components with broad applicability.