On August 14, 2026, NVIDIA announced the full-scale mass production of its Spectrum-X CPO Ethernet optical switches, with the first batch of customers locked in as CoreWeave, Lambda Labs, and Oracle. This switch, utilizing CPO technology, features silicon photonics chips manufactured by TSMC (Taiwan Semiconductor Manufacturing Company), wafer-level packaging and testing by SPIL, laser chips supplied by Lumentum, laser module assembly by TFC Optical Communication, and complete system integration by Hon Hai (Foxconn).
The parameters provided by NVIDIA are highly impactful: compared to traditional pluggable solutions, the number of lasers is reduced by 75% (only a quarter of the original), power consumption is lowered by 80%, and the Mean Time Between Failures (MTBF) is improved by 10 times. The flagship model SN6800 can be configured with 512 800G ports, offering a total switching bandwidth of 409.6 Tbps, based on the next-generation Spectrum-6 chip with a single-chip switching capacity of 102.4 Tb/s.
With laser usage dropping by 75%, upstream demand for Indium Phosphide (InP) would logically shrink significantly. However, the situation on the front lines of the industry is exactly the opposite: the lead time for InP substrates has been extended to 24–40 weeks, and some customized specifications require a 30%–50% advance deposit before production can be scheduled.
Why is the supply of Indium Phosphide becoming tighter despite the decline in laser demand? And how long will this shortage pattern last?
Why CPO Actually Exacerbates the InP Shortage
Let us first look at the Bill of Materials (BOM) structure of CPO.
CPO packages the "optical engine" directly next to the switching ASIC, compressing the electrical signal transmission distance from tens of centimeters to a few millimeters. A typical optical engine consists of four parts: a silicon photonics PIC (Photonic Integrated Circuit, responsible for light guiding, splitting, and modulation), an InP laser source, a Fiber Array Unit (FAU), and micro-lens and coupling components, supplemented by TEC (Thermo-Electric Cooler) temperature control and packaging substrates.
In the CPO era, optics have become the most expensive part of the switch. Goldman Sachs once disassembled NVIDIA's Quantum-X CPO switch and found that in a switch with a BOM cost of $75,800, the optical engine alone accounts for 43% (72 units at $450 each), while the switching chip accounts for only 16%.
BOM Cost Breakdown of the Quantum-X CPO Switch (Goldman Sachs)
Taking the Quantum-X800 InfiniBand CPO switch as an example, the entire machine is equipped with 4 Quantum-X800 switching chips, a total of 72 silicon photonics engines, and 18 independent InP external Continuous Wave (CW) laser sources. The hardware architecture features 4 silicon photonics engines sharing 1 InP laser source. Compared to the traditional pluggable optical module solution with built-in lasers of the same bandwidth, the total number of lasers is reduced to 1/4 of the original, a decrease of 75%.
In the pluggable era, each EML (Electro-absorption Modulated Laser) laser directly modulates data signals. However, under the CPO architecture, NVIDIA physically separates the lasers from the heat-generating ASICs, designing them as front-end, replaceable external laser sources. A small number of high-power CW (Continuous Wave) lasers provide centralized light, which is then distributed to multiple optical engines via splitting and microring modulators. Although the number of lasers has decreased, it inversely amplifies the consumption of InP, for the following reasons:
First, the value content per unit surges. The high-power CW light sources required for CPO have a single InP substrate area 4 to 5 times that of traditional EML chips. According to industry chain calculations, at the same data rate, the total consumption of InP substrates in the CPO architecture has not decreased but has instead increased by about 50%.
Second, reliability costs drive up yield loss. Once packaged, CPO optical engines are almost impossible to rework, and NVIDIA has set the MTBF target at 10 times that of traditional solutions. This means the lasers must undergo more stringent burn-in screening, resulting in a lower good die yield. Behind every qualified chip is an InP investment several times higher than that of traditional solutions.
Third, reducing the quantity does not equal reducing the total volume. At least for now, CPO is not a pure substitute for pluggables but an overlay. In 2025-2026, 800G/1.6T pluggable optical modules themselves are in the fastest volume ramp-up cycle in history, with 50G/100G high-speed EMLs already in extremely short supply; the high-power CW brought by CPO represents an entirely new incremental demand.
Global Oligopoly of Three in InP Substrates, with Capacity Concentrated Release in 2028
In addition to the aforementioned explosion on the demand side, the supply side is equally highly rigid.
To lock in upstream capacity, on March 2, 2026, NVIDIA officially announced investments of $2 billion each in Lumentum and Coherent, signing long-term supply agreements (LTAs) worth billions of dollars along with capacity priority rights.
Let us first look at the most upstream InP substrates.
Over 90% of the global InP substrate capacity is held by three companies: Sumitomo Electric (about 42%), AXT (about 36%), and JX Metals (about 13%). InP single-crystal growth must be conducted at temperatures above 1000°C and a pressure of about 27 atmospheres, with only over 10 kilograms of raw materials per batch; it takes 18 to 36 months from the start of a production line to stable mass production, and customer certification takes another 1 to 2 years. The core MOCVD (Metal Organic Chemical Vapor Deposition) equipment is supplied globally by only Aixtron and Veeco, resulting in long delivery cycles.
Global InP Substrate Capacity Share
Yole data shows that in 2026, global demand for InP substrates will reach as high as 2.6 million to 3 million pieces (equivalent to a 2-inch diameter), while effective capacity is only about 750,000 pieces, resulting in a gap of over 70%. Further upstream, Indium is a by-product metal with China's reserves accounting for 75% of the global total and production accounting for 70%. After being included in export controls in February 2025, its price once surged to a ten-year high. The rigid constraints of resources make the act of capacity expansion itself full of uncertainties.
Faced with the gap, global players have voted with their hard cash.
At the substrate end, the three giants are expanding capacity simultaneously but with staggered paces.
Sumitomo Electric announced in July 2026 that it plans to invest about 18 billion yen to increase the total capacity of InP substrates to 3.1 times that of fiscal year 2024 by the end of fiscal year 2028;
JX Metals announced in June 2026 an expansion plan of up to 120 billion yen before fiscal year 2030, which will increase the capacity of InP substrates, the core material for optical communications, by 7 to 10 times compared to fiscal year 2025. This is the largest in scale but will take the longest to yield results;
AXT (Tongmei Beijing) completed an additional issuance of about $600 million in April 2026, expecting quarterly output value to reach $35 million by the end of 2026 and $65 million to $70 million from the end of 2027 to early 2028. Its backlog of orders has exceeded $100 million, and it has signed a three-year long-term contract with Coherent.
Domestically, Yunnan Germanium (Xinyao Semiconductor) is the absolute local leader and the only manufacturer in China capable of delivering 6-inch substrate samples in small batches. Yunnan Germanium's current 2-4 inch capacity is 150,000 pieces/year. It plans to produce 180,000 pieces in 2026, and the newly added capacity, equivalent to a 4-inch diameter, will reach 300,000 pieces/year, with full capacity expected to be achieved by the end of 2027.
As optical modules evolve from 800G to 1.6T and even 3.2T, the demand for InP chips per module grows exponentially. Moreover, the laser arrayization required by 1.6T and CPO technologies forces the substrate size to leap from 2 inches to 6 inches.
However, 6-inch InP wafers are currently recognized by the industry as a technical watershed. When the crystal size increases, the original micro-size crystal growth technology cannot be simply replicated, which imposes stricter limitations on crystal growth equipment. Enterprises must independently develop or upgrade large autoclaves, high-power heaters, and highly precise temperature and mechanical control systems. Therefore, there are very few companies globally capable of 6-inch mass production, making 6-inch capacity a scarce core asset.
At the optical device end, Coherent is moving the fastest. Its Sherman, Texas factory, as its first global scaled 6-inch InP production line, has already entered mass production, completing a capacity doubling by mid-2026 and planning to double it again by the end of 2027. Its CEO explicitly judged that the InP shortage will span both 2026 and 2027. In addition, Lumentum's new factory in North Carolina plans for its 6-inch InP production line to reach mass production only by 2028, which is too late to solve the immediate shortage.
Among domestic laser chip manufacturers, Yunjie Technology has mass-produced 100G EMLs, with 200G EMLs expected to be mass-produced by the end of 2026, and 70/100mW CW light sources positioned for 800G/1.6T silicon photonics and CPO scenarios. Source Photonics, under Dongshan Precision, achieved stable mass production of 200G EMLs in the first quarter of 2026, with 100mW CW passing AWS and Coherent certifications. The 300-400mW high-power version has been sampled to NVIDIA, with a long-term plan for an annual production of 1 billion units. Accelink Technologies is one of the few domestic InP IDM manufacturers, with a self-supply rate of 100G EMLs exceeding 70%, and has launched the world's first 3.2T silicon photonics NPO module.
At the CPO system end, NVIDIA is the frontrunner. The Quantum-X (InfiniBand) CPO switch was launched in the second half of 2025, the Spectrum-X CPO Ethernet switch will be mass-produced in 2026, Rubin Ultra (2027) will provide a hybrid copper cable + CPO solution, and Feynman (2028) will fully switch to CPO.
As a more advanced optical interconnect solution, as long as CPO technology matures and achieves stable mass production, the original optical module demand will inevitably continue to migrate to CPO.
Stretching out the timeline, the release rhythm of the shortage is roughly as follows:
In 2025-2026, substrates and EMLs will be the tightest, with lead times of 24-40 weeks, and shortages will persist throughout the year;
In 2026-2027, the volume ramp-up of CPO and the iteration of pluggables will highly overlap, making high-power CW and 6-inch substrates the most scarce varieties;
In 2028, the new capacities of Sumitomo Electric, JX, AXT, Coherent, Lumentum, etc., are expected to be released in a concentrated manner, leading supply and demand toward rebalancing.
The true turning point will most likely occur around 2028.