by Macdonald Amoah, Jahara Matisek, and Morgan Bazilian
The future of artificial intelligence may depend less on processors than on light itself. Inside hyperscale AI data centers, computation is no longer constrained primarily by transistor density or even electricity supply. The emerging bottleneck is the movement of information between processors at the speed required to hold giant AI systems together as a coherent machine. That shift has quietly transformed a little-known semiconductor material into one of the most strategically important substances in the global technology economy: indium phosphide (InP).
Few policymakers understood its value until Beijing placed InP substrates under export licensing controls in 2025. Yet, inside hyperscale data centers, InP is the critical photonic link, enabling the lasers that push data through optical interconnects, fusing tens of thousands of individual processors into a single computing fabric.
Since 1993, America has been 100 percent import reliant for the refined indium metal required to produce these substrates. China produces about 70 percent of global refined indium production, and as of late 2024, it directly supplied a quarter of U.S. imports. The remainder of the U.S. supply chain is similarly concentrated, relying on the Republic of Korea (29 percent), Japan (18 percent), and Canada (14 percent). For the companies building AI and fifth-generation (5G) communications systems, the vulnerability gets even worse at the wafer stage. Two Japanese firms and one U.S.-headquartered, China-based AXT collectively control 90 percent of InP wafers globally.
AI is Becoming a Materials Science Problem
Scaling AI requires transitioning from electrical signals to light; generating that light requires a continuous supply of InP. Securing this critical material is essential for America and its allies to maintain their economic strength and military power.
Artificial intelligence is often described as a software revolution. In practice, scaling frontier AI increasingly resembles an industrial mobilization problem. Every major expansion in AI capability now carries a parallel expansion in physical infrastructure requirements: electricity generation, cooling systems, high-voltage transformers, transmission capacity, fiber optics, semiconductor fabrication, and critical mineral supply chains.
As model sizes and GPU clusters continue growing, the constraints increasingly shift away from code and toward the physical systems capable of moving energy and information at extreme scale. The AI race is increasingly constrained not by algorithms, but by the periodic table.
Optical Fiber
Training a large language model requires continuous, high-volume data exchange across thousands of processors operating in parallel. As GPU clusters expand from the thousands into the hundreds of thousands, the interconnect fabric becomes the primary limiting factor. Industry projections suggest GPU clusters could reach one million units by 2026. At this massive scale, the physical properties of copper wire become actively hostile to performance. At data rates of 224 gigabits per second per lane (the industry standard in late 2024), the effective reach of a passive copper cable collapses to less than one meter.
Optical fiber bypasses these physical limits entirely. It transmits photons through glass at roughly 200,000 kilometers per second with losses measured in fractions of a decibel per kilometer. A single rack of advanced AI hardware can now draw up to 125 kilowatts, which is more than an average American household consumes in a month. Within that massive power budget, the transceivers converting electrical signals to optical pulses represent a disproportionate share of energy consumption. NVIDIA chief executive Jensen Huang quantified this concern at GTC 2025, noting that with six pluggable transceivers per GPU each consuming 30 watts, scaling to a million GPUs would devote enormous capacity solely to signal conversion. Transitioning to co-packaged optics (integrating photonic and electronic functions into a single package) reduces this energy burden by approximately 73 percent. This cuts power usage from around 17 watts per 800-gigabit link down to under 5 watts.
The commercial scale of this hardware transition is massive. The data center optical component market exceeded $16 billion in 2025, growing by more than 60 percent year over year. Shipments of 800-gigabit optical transceiver modules doubled over the same period. The silicon photonics market serves as the integrated platform underlying much of this infrastructure. This specific sector stood at $3.1 billion in 2025 and is forecast to reach $10.4 billion by 2030, advancing at a 27 percent compound annual growth rate. Today, over 80 percent of data center links in hyperscale environments now rely on optical interconnects.
At sufficient scale, interconnect performance stops being a secondary engineering issue and becomes a hard computational limit. A modern AI training cluster cannot simply “slow down” when optical transmission bottlenecks emerge. GPUs begin sitting idle while waiting for data synchronization across the cluster. Training efficiency collapses even as electricity consumption remains elevated. In practical terms, photonic constraints can strand billions of dollars of AI infrastructure capacity.
Why Indium Phosphide Matters
Silicon has dominated semiconductor manufacturing for six decades because it is abundant, highly scalable, and compatible with mature fabrication infrastructure. However, silicon has a fundamental physical limitation since it cannot produce laser light. It is an indirect-bandgap material, meaning that when an electron transitions between energy states, it cannot efficiently emit a photon. Indium phosphide (InP) solves this exact problem.
As a direct-bandgap compound semiconductor, electron-hole recombination within InP releases energy as photons with high probability. InP is the only major integrated-photonics platform that can both conduct and produce light. This makes it uniquely capable of generating the on-chip laser sources required by AI optical systems. Its emission characteristics are also perfectly matched to the telecommunications C-band centered at 1,550 nanometers. This is the precise wavelength window at which single-mode optical fiber exhibits minimum attenuation of 0.2 decibels per kilometer. No other commercially mature material combines these properties at this wavelength with comparable performance.
In earlier industrial eras, strategic power flowed through oil fields, pipelines, shipping lanes, and telecommunications cables. In the AI era, geopolitical leverage may increasingly depend on command of the interconnect: the materials and systems that allow information to move at light speed across giant computational fabrics.
Consequently, the silicon photonics industry relies completely on InP to generate light. Silicon is used to fabricate the waveguides, modulators, and passive components in a modern optical transceiver. Manufacturers must then bond InP laser dies directly onto silicon substrates or use InP-based external cavity lasers to feed silicon chips with continuous-wave light. Leading research institutions (including UC Santa Barbara, TU Eindhoven, and IMEC) continue to advance heterogeneous integration by printing InP gain material onto silicon waveguides. Regardless of the integration method, the InP material itself remains indispensable. As Professor Martijn Heck of TU Eindhoven recently noted, there are no substitutes for III-V semiconductors like InP in their specific applications.
Weaponizing the Optical Supply Chain
The strategic problem with indium actually occurs before the mine. Indium is almost never targeted for primary mining. It occurs in trace concentrations (one to 100 parts per million) within zinc-sulfide ore deposits. Consequently, indium is recovered as a byproduct of zinc smelting. This means a highly inelastic supply curve. Spikes in indium demand do not equal more indium mines. Production only surges if global zinc economics justify expanding zinc smelting capacity or improving recovery rates at existing plants.
Because Beijing controls half of global zinc smelting capacity, it commands the global indium supply, leaving Western markets exposed to a major electronics vulnerability. Beijing has steadily weaponized these dependencies through a calibrated escalation of export controls. The mineral control campaign began in mid-2023 with licensing requirements on gallium and germanium, followed closely by graphite and antimony. In December 2024, China escalated significantly by banning gallium, germanium, and antimony exports to the United States alongside a blanket prohibition on all dual-use item exports to U.S. military end users.
In early 2025, China imposed export controls over tungsten, tellurium, bismuth, molybdenum and indium. This caused indium prices to jump from $2,600 per kilogram to $3,000 in less than a month, and over time it has led to indium prices peaking around $4,700 per kilogram in March 2026.
Price shocks are damaging, but Chinese licensing regimes are even more dangerous. They operate as a mandatory intelligence-gathering mechanism. To obtain export approval, Western applicants must submit exhaustive end-use information to Chinese authorities, serving as a dual-use intelligence purpose.
Furthermore, the early 2025 directive did not just target raw indium metal. It explicitly restricted indium phosphide substrates and the highly specialized precursor chemicals required to manufacture them. The distinction is critical. Producing a finished substrate requires advanced crystal growth infrastructure and decades of manufacturing expertise. It is the finished substrate, rather than the raw ore, that feeds directly into the optical transceivers powering AI data centers.
Ultimately, this strategy injects severe uncertainty directly into the investment and procurement decisions underpinning the American tech sector. Beijing is doing more than driving up commodity costs. It is actively holding the hardware layer of Western AI development at risk.
Strategic leverage does not require a full embargo. It only requires enough uncertainty to distort procurement decisions, delay infrastructure investments, and force competitors into defensive supply-chain behavior. In that sense, export licensing regimes function not only as trade restrictions, but as instruments of industrial pressure.
Breaking the Photonic Chokepoint over AI
The convergence between AI infrastructure and energy infrastructure is becoming impossible to separate. Optical interconnect efficiency now directly affects data center electricity demand, cooling loads, rack density, and transmission planning. In this sense, indium phosphide is not simply a semiconductor issue. It is an emerging energy systems issue.
The AI competition has become a materials race, stretching from power transformers to obscure rare earths like yttrium. During the twentieth century, great powers competed for command of oil supply chains because fuel determined military mobility and industrial output. In the twenty-first century, command of the interconnect may play a comparable role for computational power. Nations that cannot secure the physical infrastructure required to move information at scale may find themselves structurally disadvantaged in AI development regardless of software sophistication.
The global market for indium phosphide wafers is relatively small, valued at just over $200 million in 2026. Yet it operates as a massive strategic chokepoint. A tiny physical input governs the performance of trillion-dollar cloud computing and AI systems, and a disruption to its supply can damage economies and degrade military operations.
Expanding capacity means overcoming decades of highly rigid zinc smelting economics and building up specialized crystal growth expertise that does not yet exist at scale outside of East Asia. Silicon and other alternative materials cannot physically replicate InP's light-emitting properties at the required telecommunications wavelengths.
The strategic importance of photonic materials is likely to increase rather than diminish. AI architectures are already moving toward co-packaged optics, photonic switching, and increasingly dense optical interconnect fabrics designed to overcome the physical limitations of electrical transmission. Future generations of AI infrastructure may become even more dependent on advanced photonic materials than current systems. Three lines of effort are essential.
First, the full value chain must be mapped with operational precision. The Defense Logistics Agency is already investigating how best to stockpile indium, and the newly established Economic Defense Unit should be tasked with completing the picture. Existing USGS mineral commodity surveys track raw material flows at a national level, but they do not trace the specific path from allied zinc concentrates through indium refining, precursor chemical synthesis, wafer crystal growth, and final optical transceiver assembly. An operational map must identify chokepoints at each stage: which smelters recover indium as a byproduct and which do not, where precursor trimethylindium is synthesized, which crystal growth facilities can produce substrate-quality boules, and how many qualified transceiver manufacturers source from non-Chinese wafer suppliers. Without this granularity, stockpiling raw indium addresses only one link in a multi-stage chain.
Second, allied wafer and refining networks must be forged in lockstep. America cannot rebuild this entire photon-for-AI ecosystem domestically in the near term. Washington needs to co-invest with Japan, South Korea, and Canada to support non-Chinese refining, recycling, and wafer production capacity. These are niche material markets that private capital routinely ignores because individual investments appear too small to justify diligence costs. Government-directed subsidies and co-financing arrangements, modeled on the joint ventures that expanded allied semiconductor packaging capacity under the CHIPS Act, can close this gap. Coordination must extend beyond funding to include joint qualification standards, so that allied producers can be certified as drop-in alternatives without years of redundant testing by each end-user.
Finally, there need to be bankable demand signals. Production capacity will not materialize if end-users refuse to qualify alternative suppliers, and suppliers will not invest in expensive crystal growth infrastructure against the risk that a single customer defects back to cheaper Chinese sources. The Pentagon and the broader federal government must use their procurement power to mandate secure InP supply chains for all future advanced communications and defense AI systems. Long-term purchase agreements that guarantee demand for allied producers would convert what is currently a geopolitical aspiration into a commercially investable proposition.
No amount of software innovation or venture capital can substitute for the physical ability to generate photons at the wavelengths required to hold large AI systems together. The next great technological competition may not be decided solely by who builds the smartest algorithms or the largest data centers. It may ultimately be decided by who controls the obscure materials, optical systems, energy infrastructure, and industrial supply chains that allow artificial intelligence to move at the speed of light.
