Michael Hurlston delivered the news from Tokyo on Friday with the cheerful tone of a chief executive who has a problem most of his peers would envy. Lumentum Holdings, the optical-component maker he runs, has sold out its entire production capacity through early 2029. “We’re completely sold out,” he said, “and right now there’s no end in sight to the shortage.”
The cause is the AI buildout. Technology companies are racing to stand up faster data centers, and the components that carry light inside and between those facilities, Lumentum’s advanced indium phosphide devices, have become a hard bottleneck. Hurlston said the company cannot meet roughly 70% of demand for some products through next year and will not cover about 30% of demand for others through 2028.
The crunch arrived faster than Lumentum predicted. As recently as April, Hurlston said the company was on track to sell out its full-year 2028 capacity. Six months later the shortage horizon has stretched into 2029, and customers are starting to lock in orders for 2030.
Nvidia saw the bottleneck forming early. Earlier this year the chipmaker invested $2 billion in Lumentum and a matching $2 billion in rival Coherent, a bet that the optical layer of AI data centers would become the next supply choke point after the processors themselves. The bet has aged well.
The two investments pointed at the same conclusion. Lumentum and Coherent make the transceivers and lasers that carry traffic inside the fastest data centers, the links that connect thousands of servers into a single machine. As chips got faster, the optical layer had to keep pace, and the companies that mastered it suddenly found their order books full.
Lumentum’s constraint is not demand but time. The company has expanded capacity at its main plant in the greater Tokyo area twelvefold over the past two years and plans to spend at least $350 million more there, with further growth at a plant in Caswell in the United Kingdom. A new facility, though, takes three to five years to bring online, far slower than the pace at which customers are asking for parts.
Hurlston was blunt about the gap. “We are really trying to catch up, add capacity as fast as we can, but we are very, very far behind what customers want,” he said in the interview. The company is also weighing acquisitions to broaden its product lineup over the next year, he said.
The tightness reaches beyond any single supplier. Every hyperscaler building an AI data center needs optical transceivers and lasers to move data between servers and racks, and that supply is now committed years in advance. When one layer of the AI stack sells out, the projects downstream slow with it.
Markets read the shortage as a signal of demand, not of excess. Lumentum shares rose Friday even as the broader chip complex wobbled on separate news about OpenAI’s revenue accounting. To investors, a company with nothing left to sell through 2029 is a company whose customers have nowhere else to go.
The episode shows where the AI supply chain’s limits now sit. Optical components were a comparatively quiet corner of the semiconductor trade. AI has turned them into a hard constraint on how fast data centers can be built, much as advanced packaging and high-bandwidth memory became constraints in earlier years.
Lumentum, based in San Jose, California, helped define the field it now leads. Its indium phosphide devices underpin high-speed cloud computing and data transmission, and its order book runs far past its production horizon.
Hurlston, who joined Lumentum in 2025 after running chip designer Synaptics, has spent his tenure trying to turn the company’s technological lead into manufacturing scale. He has described the situation as a footrace the company cannot win outright, only run harder. The expansions in Tokyo and Caswell are meant to narrow the gap, but he has conceded that new capacity arrives too slowly to erase it.
For the industry, the message is that AI’s next shortage may not be chips at all but the unglamorous parts that connect them. Companies that thought their supply problems ended once they secured graphics processors are discovering a new bottleneck further down the line.
The pattern will be familiar to anyone who has watched the chip business over the past two years. First it was high-bandwidth memory, which sold out as Nvidia’s processors multiplied demand. Then it was advanced packaging, the process that bonds memory to processors. Now the pinch has moved to the parts that carry light, and the lesson each time is the same: the bottleneck rarely sits where the market expects it.
For Hurlston, the near-term task is simple: build more. He acknowledged that new capacity cannot come online quickly enough to meet demand and that the company sees no point at which supply and demand will balance. The sold-out sign, by his own account, is likely to stay up.
That is good news for Lumentum’s shareholders and bad news for the data-center operators waiting on its parts. The AI boom has produced a company that cannot make its products fast enough, and the gap, for now, keeps widening rather than closing.
