أزمة فوسفيد الإنديوم: عنق الزجاجة الخفي الذي سيعطل طموحات الذكاء الاصطناعي لسنوات

The explosive growth of artificial intelligence is hitting a physical wall, and it's made of a compound semiconductor called indium phosphide (InP). Essential for the lasers and detectors that drive high-speed fiber-optic communication inside data centers, InP is now in critically short supply. The numbers are stark: current demand is roughly triple the effective global production capacity, and the structural reasons behind this gap mean it will not close any time soon—likely not by 2030.

The Current State of the InP Supply Gap

By multiple independent estimates, the supply-demand imbalance for InP substrates is severe and immediate. Global demand for InP devices is estimated to have reached approximately 2.0 to 2.1 million wafers (in 2-inch equivalents) in 2025 . Against this, effective global production capacity stands at only around 600,000 to 700,000 wafers per year, leaving a supply gap exceeding 70% . This figure is corroborated by Chinese brokerage China Galaxy Securities, which describes a supply-demand gap for actual production capacity exceeding 50% .

The situation is intensifying quickly. Shipments of 800G and faster optical transceivers are expected to jump from roughly 24 million units in 2025 to 63 million units in 2026, a 2.6× increase that will far outstrip any near-term growth in substrate output . The transition to 1.6T optics compounds the problem, as a 1.6T optical module can demand 2.7 to 2.8 times more InP substrate than an 800G module .

Why 2030 Doesn't Look Like a Return to Balance

The outlook to 2030 is one of sustained pressure, not equilibrium. While some forecasts suggest that the introduction of larger-diameter wafers and improved fabrication yields could begin to ease wafer-level availability toward the end of the decade , this supply-side relief runs headlong into a demand curve that is accelerating faster. The co-packaged optics (CPO) market for AI data centers alone is projected to grow from $9.0 billion in 2025 to approximately $122.5 billion by 2034, with 3.2T CPO interfaces expected to become the standard by 2030 .

McKinsey's analysis projects that production of 800-Gbps transceivers will fall 40 to 60 percent short of demand through 2027, with 30 to 40 percent shortfalls for 1.6-Tbps transceivers likely persisting through 2029 . The same structural constraints that created the current crisis are not expected to be resolved within the decade. The baseline outlook for the InP substrates market from 2026 to 2035 is described as "one of sustained, high-value growth constrained by supply-side capabilities rather than demand" .

The Structural Reasons Manufacturers Cannot Scale to Meet Nvidia's Ambitions

Nvidia is not waiting passively. The company has invested $4 billion in photonics leaders Lumentum and Coherent, securing multibillion-dollar, multiyear purchase commitments for advanced laser and optical components . It has also committed $500 million to a multi-year collaboration with Corning to expand U.S. optical connectivity production tenfold .

Yet even this unprecedented spending cannot solve the InP bottleneck quickly because the constraints are physical, not financial.

Qualification cycles stretch for years. Bringing new InP substrate capacity online is not a matter of flipping a switch. The qualification cycle for new production lines stretches 18 to 24 months, meaning capacity expansions planned today will not yield qualified output until at least 2027 .

The supply base is dangerously concentrated. High-quality InP substrate production is controlled by a very small number of global suppliers—roughly five to six firms—making the entire supply chain brittle and unresponsive to sudden demand spikes . Only 2 to 3 companies produce InP substrates at scale .

Physical throughput is the real ceiling. Indium phosphide manufacturing requires specialized metal-organic chemical vapor deposition (MOCVD) reactors and molecular beam epitaxy (MBE) systems. The order books for this equipment already extend through 2027, and skilled operators are scarce . CEO of Lumentum, the dominant player with 50–60% of the global EML laser market, stated in early 2026 that the company was "undershipping customer demand by somewhere around 30%," and that even after adding 20% additional capacity, the demand-supply imbalance had increased . All of Lumentum's EML capacity is locked under long-term agreements through calendar 2027 .

The scale of demand is unprecedented. AI data centers aren't scaling linearly. As they grow from thousands to hundreds of thousands of accelerators, optical interconnect density requirements have multiplied by 8–16× . A single AI server can require ten times more optical modules than a traditional server, and over 80% of current InP demand now comes from AI data centers .

How This Fits the Broader Pattern of AI Infrastructure Bottlenecks

The InP shortage is not an isolated crisis. It is the latest manifestation of a structural mismatch between the explosive demand for AI infrastructure and the slow-moving nature of specialized component supply chains.

High-bandwidth memory (HBM) has already been a well-documented chokepoint, with all three major DRAM/HBM manufacturers reported to be operating at full capacity . Optical transceivers themselves are a central constraint, with hyperscalers starved for sufficient 800G and 1.6T modules to connect their GPU clusters .

Now, the InP laser and substrate shortage sits inside that optical bottleneck as the most acute pressure point. One supply-chain analyst described the situation as a "catastrophic-level" shortage, with InP lasers and optical devices characterized as a "complete disaster" . Goldman Sachs has noted that the supply-demand gap for InP and several other electronic components has "continued to widen over the past four months" and is "unlikely to reverse before 2027" .

From HBM to optical transceivers to InP substrates, a consistent pattern emerges: each new wave of AI infrastructure demand crashes into a physical supply chain that was never designed to scale at AI velocity. These supply chains share long qualification cycles, highly concentrated supplier bases, and specialized manufacturing equipment with multi-year backlogs. The InP crisis is simply the latest point of failure, and with demand for AI optical components projected to grow at a compound annual rate of 85% through 2030 from data centers alone, it will not be the last .