AI’s Next Supply Chain Crisis: How the InP Laser Shortage Threatens Data Centers
2026/08/04 11:13:00
The artificial intelligence boom has already exposed shortages in GPUs, high-bandwidth memory, advanced packaging, electricity and cooling. The next constraint may be far less visible but just as important: the lasers that move enormous volumes of data between processors inside modern AI data centers.
Indium phosphide, or InP, is a compound semiconductor used in high-speed optical components. Demand for InP-based lasers is rising as cloud companies move from 400G and 800G networks toward 1.6T connections and co-packaged optics. At the same time, export controls, limited production capacity and a concentrated supplier base are making expansion more difficult. For crypto investors, the issue could influence AI infrastructure costs, decentralized computing economics, AI-token valuations and the wider risk appetite connecting technology stocks with Bitcoin.
How the InP Shortage Became an AI Problem
The shortage was not caused by a single factory accident or a sudden wave of speculative buying. It developed as AI clusters became larger and more dependent on high-speed networking. Training and serving advanced models requires thousands of accelerators to exchange data with minimal delay. That has increased demand for optical transceivers, high-capacity switches and the lasers that generate the signals traveling through fiber-optic cables.
Lumentum says the expansion of AI infrastructure is driving unusually strong demand for optical components. The company has cited market estimates showing demand for InP electro-absorption modulated lasers and related laser chips running approximately 30% to 50% ahead of supply. The imbalance matters because an AI cluster cannot operate efficiently when processors are available but the network connecting them is incomplete.
Geopolitics has added another layer of uncertainty. China introduced tighter controls on InP-related exports in February 2025, while overseas buyers have reported closer scrutiny of indium shipments. China produces about 70% of global indium, the upstream metal used to manufacture InP. Reuters reported that licensing delays and the concentration of substrate production have complicated efforts by American and European companies to secure alternative supplies.
Corporate purchasing behavior provides further evidence that this is a structural problem rather than a temporary headline. AXT agreed to reserve InP substrate supply for Lumentum through December 31, 2031. Lumentum is also building a new U.S. manufacturing facility for advanced InP lasers, while Coherent is expanding its six-inch InP production capacity in Texas. These are multiyear investments designed to address a supply chain that was not built for the present pace of AI infrastructure growth.
Why AI Data Centers Need InP Lasers
An AI data center is not simply a warehouse filled with GPUs. It is a synchronized system in which processors, memory, networking, electricity and cooling must expand together. A large cluster may contain thousands or tens of thousands of accelerators. If those accelerators cannot exchange model parameters and intermediate calculations quickly enough, they spend more time waiting and less time computing.
Copper connections remain useful over short distances, but signal loss, heat and power consumption become harder to control as speeds and distances increase. Optical links can move data over fiber with higher bandwidth and better energy efficiency. InP is valuable because it can generate and modulate light at wavelengths widely used in fiber-optic communication. Lumentum’s data-center portfolio includes InP-based EML and continuous-wave lasers designed for high-bandwidth, low-latency networks.
EMLs and Continuous-Wave Lasers
Electro-absorption modulated lasers, or EMLs, combine a laser source with a high-speed modulator. They are commonly used in pluggable optical transceivers. Lumentum’s 200G EML, for example, is designed so that eight 200G lanes can support a 1.6T module. As data centers upgrade to faster connections, the performance and production requirements for these devices become more demanding.
Continuous-wave lasers provide a stable light source for silicon-photonics systems and co-packaged optics. Silicon photonics can guide, divide and modulate light efficiently, but silicon is not an efficient light-emitting material. NVIDIA’s co-packaged systems therefore continue to use serviceable external laser sources. The architecture changes where the laser is located and how it is connected; it does not remove the need for a qualified light source.
| Infrastructure component | Main function | Consequence of a shortage |
| GPU or AI accelerator | Performs training and inference calculations | Limits raw computing capacity |
| High-bandwidth memory | Supplies data to accelerators | Reduces processor utilization |
| Optical transceiver | Converts electrical and optical signals | Delays high-speed network deployment |
| InP laser | Generates light for optical connections | Constrains transceiver and photonics production |
| Network switch | Connects accelerators into a cluster | Limits cluster scale and efficiency |
| Power and cooling | Keeps the system operational | Restricts the amount of hardware that can run |
The simplest analogy is that GPUs are the engines of an AI data center, while the optical network is the highway system connecting them. InP lasers provide part of the light that keeps traffic moving. An operator may own enough engines, but the system cannot reach full performance if the highways are unfinished.
Why Supply Cannot Catch Up Quickly
The first difficulty is supplier concentration. InP substrates and high-performance laser devices come from a much smaller industrial base than conventional silicon chips. Reuters reported that three companies dominate approximately 90% of global InP substrate supply. When a market depends on only a few qualified producers, a disruption involving raw materials, wafers, epitaxy or device fabrication can affect the entire chain.
Customers also cannot replace suppliers immediately. Data-center lasers must meet strict requirements for output power, temperature stability, operating life and consistency. A new wafer or laser source must pass testing and customer qualification before it can be used in large commercial systems. Capacity that exists on paper may therefore take months or years to become usable production.
Larger wafers could improve the economics of the industry. Coherent says its six-inch InP platform can provide four times as much production capacity and reduce die costs by more than 60% compared with smaller-wafer approaches. However, those benefits depend on equipment availability, process maturity, yield improvement and customer adoption. Scaling compound semiconductor manufacturing remains more complicated than adding capacity to a mature silicon process.
Suppliers must also avoid expanding at the wrong point in the cycle. Building too slowly risks losing customers and prolonging shortages. Building too quickly could leave manufacturers with underused factories if hyperscaler spending later weakens. Long-term contracts, customer prepayments and strategic investments help divide this risk between companies that need secure supply and manufacturers that must finance expensive new capacity.
How the Shortage Could Slow Data Centers
An InP shortage is unlikely to shut down AI data centers that are already operating. The more realistic risk is that it delays new clusters and network upgrades. Operators could receive their GPUs but wait longer for 800G or 1.6T optical modules, switches or external laser systems. Expensive processors may then operate in smaller configurations, remain underused or generate revenue later than planned.
The impact will not be evenly distributed. Hyperscalers can offer deposits, sign multiyear contracts and negotiate priority access to production. Smaller cloud providers, specialized AI companies and decentralized-compute operators have less purchasing power. They may face higher prices, longer lead times or less favorable allocations, potentially strengthening the position of the largest infrastructure companies.
Co-packaged optics may improve energy efficiency and reduce signal loss, but it does not eliminate the supply problem. NVIDIA says its silicon-photonics switches can significantly reduce power consumption and improve signal integrity compared with traditional pluggable designs. Those systems still depend on external laser sources and a specialized photonics ecosystem. The bottleneck may shift from conventional transceivers to high-power laser arrays rather than disappearing.
The real challenge is system balance. More GPUs do not automatically create more usable AI capacity. Compute, memory, networking, power and cooling must arrive in the correct location at approximately the same time. A relatively inexpensive optical component can delay a facility whose total investment is measured in billions of dollars.
Nvidia’s Optics Bet Changes the Story
NVIDIA’s actions have turned optical networking into a mainstream AI investment theme. In March 2026, the company announced separate strategic partnerships with Lumentum and Coherent, including a $2 billion investment in each business. The agreements are intended to expand manufacturing capacity, deepen research and support the optical technology required by future AI infrastructure.
The logic is straightforward. NVIDIA can design increasingly powerful accelerators, but customers must connect them into reliable, efficient clusters. As model sizes and inference traffic grow, moving data becomes a larger part of total system performance and energy consumption. Securing optical capacity protects the value of NVIDIA’s broader computing and networking platform.
| Market signal | What it suggests | What it does not prove |
| NVIDIA invests in Lumentum | Optical capacity has become strategically important | Every optics company will earn higher margins |
| NVIDIA invests in Coherent | Multiple suppliers and technologies are needed | The shortage will last permanently |
| Lumentum builds a new U.S. facility | Domestic InP capacity is expanding | New production will be available immediately |
| Coherent expands six-inch manufacturing | Larger wafers may improve scale and cost | Manufacturing and yield risks have disappeared |
| AXT reserves supply through 2031 | Customers want long-term material security | Demand will rise in a straight line until 2031 |
NVIDIA has also entered a long-term partnership with Corning to expand the optical connectivity used in hyperscale data centers. Together, these agreements show that fiber, photonics and lasers are no longer secondary parts of the AI story. They are becoming strategic infrastructure alongside accelerators and memory.
What the Crisis Means for AI Tokens
The InP laser shortage has no direct mechanical relationship with the price of an AI token. Crypto projects generally do not manufacture optical components. The connection runs through expectations about AI industry growth, the cost of computing infrastructure and the ability of decentralized networks to deliver useful services at competitive prices.
Higher Costs Will Test Weak Business Models
If networking equipment becomes more expensive or difficult to obtain, the overall cost of building GPU clusters increases. Projects that purchase cloud capacity and resell it through tokenized marketplaces may experience narrower margins. Networks that use token emissions to subsidize demand could face additional pressure when real infrastructure expenses rise faster than customer revenue.
A token can coordinate incentives and payments, but it cannot manufacture scarce hardware or shorten factory qualification cycles. Investors should therefore examine whether an AI-crypto project has active computing providers, recurring customers, transparent pricing, verifiable workloads and enough revenue to support its incentives. A functioning marketplace can adjust to higher costs; a project relying mainly on the phrase “AI plus blockchain” has much less protection.
Scarcity Is Not Automatically Bullish
Some investors may assume that limited centralized capacity automatically increases the value of decentralized AI networks. That argument has some merit when a network improves the utilization of idle GPUs or connects customers with smaller regional providers. However, decentralized operators also depend on data centers, switches, fiber and optical components. They are not isolated from physical supply constraints.
The likely result is greater differentiation among AI tokens. GPU marketplaces, inference networks, data platforms and agent protocols should not be valued as one category. Their exposure to hardware, bandwidth, latency and capital spending varies considerably.
The shortage is therefore better understood as a fundamental test than as a universal crypto catalyst. It may reveal which projects understand the economics of delivering AI services and which are primarily trading on a popular narrative.
Can Decentralized AI Benefit?
Decentralized computing networks have a genuine opportunity when centralized capacity is expensive or difficult to access. They can aggregate idle GPUs, connect smaller customers with regional operators and provide alternatives to long-term contracts with major cloud platforms. Token incentives can help attract computing resources and coordinate cross-border payments.
These networks may become particularly useful for inference, model serving, rendering, data processing and other workloads that can be divided across providers. If optical shortages increase hyperscaler costs or slow new data-center construction, customers may become more willing to test alternative sources of computing capacity.
However, decentralized networks cannot fully replace hyperscale clusters. Training frontier models requires tightly synchronized accelerators, extremely high bandwidth and very low latency. A geographically distributed collection of consumer GPUs cannot easily reproduce the performance of a purpose-built cluster connected by advanced InfiniBand or Ethernet fabrics.
The strongest decentralized projects will be those that select workloads compatible with their structure instead of claiming to replace every centralized provider. Decentralization can improve market access and hardware utilization, but it cannot remove the physical constraints of bandwidth, energy, reliability and distance.
What the Shortage Means for Bitcoin
Bitcoin is even further removed from the InP supply chain. Its security depends mainly on ASIC mining hardware and electricity rather than the high-speed optical networks used for advanced AI training. The shortage is therefore unlikely to become a direct driver of Bitcoin adoption, mining economics or network activity.
The more relevant connection is market sentiment. AI capital expenditure has become important to technology-sector earnings and valuations. If networking bottlenecks delay data-center revenue or reduce expected returns on AI investment, technology stocks could weaken and broader demand for risk assets could fall. Bitcoin may be affected when investors trade it as a high-volatility asset within the same liquidity environment.
A more positive outcome is also possible. Supply constraints could trigger additional investment, manufacturing expansion and government support without materially slowing the AI buildout. In that scenario, the event would reinforce the long-term infrastructure cycle. Interest rates, global liquidity, ETF flows, regulation and crypto-specific demand would still be more important to Bitcoin than the price of InP components.
What Investors Should Watch Next
The shortage narrative should be tested against operating data rather than repeated as a slogan. Investors need evidence that demand remains above supply, that new factories are progressing and that customers are willing to pay enough for manufacturers to earn attractive returns.
In a bullish scenario, new capacity arrives while AI demand remains strong, allowing optical suppliers to increase output without destroying pricing. The base case involves persistent tightness, gradual expansion and priority allocation to the largest customers. The bearish scenario would combine delayed data-center projects with weaker AI spending, leaving customers with lower returns and suppliers with expensive expansion commitments.
The decisive question is not whether demand for optical connectivity will grow. It is whether qualified supply can expand quickly enough to prevent networking from becoming a material limit on AI deployment.
The AI Boom Is Now a Full-Stack Race
The InP shortage changes how investors should view artificial intelligence infrastructure. The first phase of the boom focused on processors. The next phase depends on the complete system: memory, packaging, networking, lasers, fiber, electricity, cooling and software. Weakness in any one layer can reduce the value of spending across all the others.
For crypto investors, this is neither a guaranteed AI-token rally nor a direct Bitcoin shock. It is a reminder that digital markets ultimately depend on physical systems. Projects with real users, efficient resource allocation and credible infrastructure strategies may gain relevance, while tokens supported mainly by broad AI narratives may struggle to justify their valuations. The next phase of the AI boom may be decided not only by who owns the best chips, but by who can connect, power and operate them at scale.
|
KuCoin is celebrating its 9th anniversary with a special platform campaign filled with exclusive rewards, trading activities, and limited-time offers. Don’t miss the chance to participate and enjoy the benefits as the exchange marks nine years of growth and innovation. Visit the official campaign page now:
|
FAQs
Is indium phosphide used inside NVIDIA GPUs?
InP is not the principal material used for the logic circuits inside NVIDIA GPUs. Those processors are mainly manufactured using advanced silicon processes. InP is relevant to optical components that move data between accelerators, switches and data-center locations.
Will the shortage directly affect ordinary Bitcoin mining?
The direct effect should be limited. Bitcoin mining farms need networking equipment, but they do not require the same tightly synchronized, ultra-high-bandwidth connections used in frontier AI training clusters. Businesses operating both large GPU facilities and crypto-based computing marketplaces have greater exposure.
Could cloud-based AI services become more expensive?
Higher optical and networking costs can raise the total expense of deploying AI infrastructure. Whether customers pay more depends on competition, utilization, purchasing contracts and the amount of additional cost absorbed by cloud providers. Price increases may therefore appear in constrained services without affecting every AI product equally.
Which crypto sectors have the greatest exposure?
Decentralized physical infrastructure, GPU marketplaces, AI inference networks and tokenized cloud-computing platforms have the clearest connection. Payment tokens, decentralized exchanges and most Layer 1 networks are less directly exposed unless their valuations depend heavily on the AI narrative.
Can companies redesign optical systems to use fewer lasers?
Design improvements can reduce component counts or allow light sources to be shared more efficiently. Co-packaged optics and external laser arrays are examples of this evolution. However, these architectures still require qualified lasers and introduce their own manufacturing, testing and reliability challenges.
How will investors know when the shortage is easing?
Evidence would include shorter delivery times, stable or falling substrate prices, higher commercial shipment volumes, improved manufacturing yields and less urgent capacity reservation. Factory announcements alone are insufficient; investors should look for qualified production and customer deliveries.
Could alternative suppliers solve the problem quickly?
Additional suppliers can improve resilience, but qualification takes time. Customers must confirm that new materials and devices meet strict performance and reliability standards. New manufacturing capacity may therefore exist for some time before it contributes meaningfully to large-scale data-center deployments.
Is the InP shortage likely to last for years?
The exact duration remains uncertain. Supply agreements extending into the next decade and current factory-expansion plans indicate that companies expect supply security to remain strategically important. The imbalance could ease earlier if larger wafers improve output rapidly, more suppliers qualify or AI infrastructure spending slows.
Disclaimer: This content is for informational purposes only and does not constitute investment advice. Cryptocurrency investments carry risk. Please do your own research (DYOR).