AI Networking's Indium Phosphide Problem Is Below the Switch

The bottleneck starts before the laser

The chain begins with indium and phosphorus being formed into high-purity single-crystal InP, sliced and processed into semiconductor substrates. Public suppliers include AXT and Sumitomo Electric. Those wafers then enter epitaxial growth and wafer fabrication, where precisely engineered semiconductor layers are deposited before devices are patterned into components such as electro-absorption modulated lasers, continuous-wave lasers, and higher-power external light sources.

From there, the supply chain widens. Laser dies feed transceiver and optical-module manufacturers. Those modules connect servers and accelerators to switches based on platforms such as Nvidia Spectrum-X and Broadcom Tomahawk. The completed fabrics are ultimately deployed by hyperscalers and other operators building large AI clusters.

The important point is that adding optical-module assembly capacity does little if the upstream laser wafer cannot be produced. Epitaxy, wafer fabrication, device yield, laser packaging, and module assembly are sequential constraints. A shortage near the beginning propagates through every layer below it.

Sumitomo Electric's own expansion plan makes the mismatch unusually clear. The company expects its production capacity for optical devices used inside data centers to increase about 12 times from 2023 to 2028, while its InP substrate capacity, including non-data-center applications, rises only about 2.4 times over the same period. [2] Those figures describe Sumitomo's capacity rather than the entire industry, but they illustrate the problem: downstream optical-device output can be scaled far faster than the crystalline substrate base supporting it.

AXT is showing the same pressure from another direction. Its second-quarter 2026 substrate revenue rose to $37.6 million from $11.3 million a year earlier, with the company attributing the increase primarily to stronger InP wafer demand for data-center and passive optical-network applications after additional Chinese government export approvals. [1] This makes InP exposure partly a manufacturing-capacity issue and partly a trade-policy issue.

Silicon photonics does not necessarily remove InP

One reason this bottleneck is easy to miss is the assumption that silicon photonics eventually replaces compound-semiconductor lasers. In many architectures, it does not.

An EML can generate and modulate the optical signal directly on an InP device. Silicon-photonics modules move more of the modulation and routing function onto silicon, but silicon is a poor native light emitter. These designs still commonly require an external CW laser, often built on InP. Co-packaged optics can push the optical engine closer to the switch ASIC while moving the laser into an external laser source. The physical location changes, but the demand for high-quality light does not disappear.

That means the industry's migration from 800G pluggables to 1.6T modules, silicon photonics, and eventually more co-packaged designs can shift which InP device is required rather than eliminating InP from the chain.

Broadcom's March 2026 introduction of its 400G-per-lane Taurus optical DSP shows where the downstream roadmap is heading. The platform is intended for 1.6T transceivers and future 3.2T modules, paired with high-speed optical components and ultimately 102.4T and 204.8T switch generations. [3] Every increase in optical lane speed puts additional demands on laser performance, power, yield, and manufacturing precision upstream.

Not every AI link depends on InP. Short-reach multimode connections can use gallium-arsenide VCSELs, and optical architectures continue to diversify. The relevant investment conclusion is therefore not that all AI networking requires InP. It is that several of the fastest-growing single-mode optical architectures still terminate at an InP light source somewhere in the system.

The economic exposure sits several layers above the hyperscaler

This changes how the AI-networking supply chain should be read. Nvidia and Broadcom may define switch architectures, and hyperscalers may determine capital spending, but neither can create additional InP crystal growth or laser-wafer capacity simply by ordering more switches.

The more asymmetric exposure is upstream. AXT is small enough that individual capacity agreements with Lumentum and Coherent are economically meaningful. Lumentum and Coherent sit one layer downstream but have responded by securing substrate capacity and expanding their own InP manufacturing. Sumitomo Electric occupies an unusual position because it participates in both substrate production and optical devices, giving it exposure to more than one constrained layer.

The hyperscalers are much less financially dependent on any single substrate producer, but operationally the chain still converges there. A Meta, Microsoft, Amazon, or Google AI cluster ultimately needs optical links between accelerators, racks, and switches. If InP substrate growth trails laser demand, the effect can appear several steps later as fewer available lasers, constrained transceiver output, delayed switch population, or higher optical-component costs.

That is the core asymmetry: the largest buyers sit at the end of the chain, but the most concentrated capacity can sit near its beginning.

For Altsets research conventions and interpretation limits, see the research methodology.

Conclusion

AI networking's InP problem is not simply an optical-module shortage. The relevant chain runs from AXT and Sumitomo Electric substrates, through epitaxy and laser fabrication at companies such as Lumentum and Coherent, into transceivers and optical modules, then into Nvidia and Broadcom switching platforms and finally hyperscale AI infrastructure.

The most important signal is not a forecast for optical traffic. It is customers paying in advance to reserve semiconductor substrate capacity years before the finished network is deployed. As AI fabrics move from 800G toward 1.6T and 3.2T, investors should watch whether InP substrate and laser-wafer capacity can expand at the same rate as the optical systems built on top of them.

Sources

1. AXT, Inc. Quarterly Report for the quarter ended June 30, 2026, U.S. Securities and Exchange Commission, 2026. https://www.sec.gov/Archives/edgar/data/1051627/000143774926027677/axti20260630_10q.htm

2. Growth Strategy for Data Center Related Business, Sumitomo Electric Industries, November 2025. https://sumitomoelectric.com/sites/default/files/2025-11/download_documents/Growth%20strategy%20for%20data%20center-related%20business_2025.pdf

3. Broadcom Delivers Industry's First 400G/lane Optical DSP for Next-Generation AI Networks, Broadcom, March 11, 2026. https://investors.broadcom.com/news-releases/news-release-details/broadcom-delivers-industrys-first-400glane-optical-dsp-next

How to Cite This

For research inquiries or data access: press@altsets.com

Go Deeper