Inside the Race to Scale Indium Phosphide: How JX Advanced Metals and Its Rivals Are Responding to a 70% Supply Gap

In under a year, the company's spending target ballooned by a factor of 80, with the final plan representing a sixfold increase over the tripling target set just four months earlier .

A Perfect Storm: AI Demand and Geopolitical Shocks

Two forces are driving this extraordinary urgency. The first is demand-side: more than 80% of downstream InP demand now comes from high-speed optical modules for AI data centers . As hyperscalers standardize on 800G and 1.6T optics and push toward co-packaged optics, the industry's appetite for InP-based lasers and wafers is expected to maintain a compound annual growth rate of 85% through 2030 for AI-related applications alone .

The second is a geopolitical shock. In February 2025, Beijing introduced export licensing controls on the compound semiconductor . The results were immediate and severe. InP prices surged roughly 250%, and the world's largest photonics suppliers were left with order backlogs stretching into 2028 . The shock threatened to slow the trillion-dollar AI infrastructure build-out, compelling the CEO of NVIDIA-backed chipmaker Coherent to personally join a US trade delegation to China in May 2026 to negotiate access .

These forces have collided with deeply structural supply constraints. Qualifying a new InP production line takes an arduous 18–24 months, most production still uses 2- and 4-inch wafers while the transition to 6-inch has barely begun, and capacity expansions started in 2025 will not ship qualified product before 2027 .

The Numbers Behind the Crisis

The supply-demand math explains the panic. In 2025, global demand for InP devices was projected at 2 million pieces against an effective production capacity of only 600,000 units — a supply deficit of nearly 70% . By 2026, demand estimates have run as high as 2.6 to 3.0 million wafers, against a capacity base that has grown only marginally to between 600,000 and 750,000 units .

The Other Players in the Global Capacity Race

JX Advanced Metals is not moving alone. The entire InP supply chain is in a state of frantic expansion to meet the demand from major optical transceiver customers.

• Lumentum is in the middle of an aggressive expansion across four existing fabs and has acquired a fifth — a former Qorvo facility in Greensboro, North Carolina — to boost InP-based optical device production . The company has stated it is currently under-shipping market demand by roughly 25–30%. A planned 40% wafer fab capacity expansion was delivered ahead of schedule by mid-2026, yet the company remains sold out through 2028 .
• Coherent is also scrambling to expand. It has tripled its InP capacity year-on-year, moved the world's first 6-inch InP production line into operations in Texas, and received a $2 billion investment from NVIDIA as part of a $4 billion dual-investment deal with Lumentum to secure multiyear purchase commitments and future capacity access .
• AXT, the major Beijing-based substrate producer, has been a source of disruption. Its export permit delays under the new Chinese controls have directly impacted Taiwanese optical module makers VPEC and LandMark Optoelectronics, forcing them to seek non-Chinese sourcing alternatives .
• IntelliEPI, a Taiwan-based epitaxial wafer supplier, has publicly stated that InP substrate shortages are constraining output and forcing operational changes across the AI transmission supply chain .
• Sumitomo Electric and other traditional Japanese players are also investing in additional capacity, though at a smaller scale than JX.

Despite the billions of dollars now being committed, the industry consensus is that the supply-demand gap will take years to close. NVIDIA's $4 billion dual-investment, Lumentum's fifth fab, JX's 7–10x target — all are efforts to catch up to a demand curve that is itself accelerating. With qualification cycles measured in years and AI optics roadmaps demanding ever-greater density, the InP substrate bottleneck looks set to be a defining feature of the AI infrastructure landscape through the end of the decade.