Lumentum expects to miss about 70% of demand for some AI optical parts in 2027

For those parts, a 70% shortfall means customers want about 3.3 units for every one Lumentum can ship in 2027 ^[18]. At the roughly 30% gap projected for 2028, the ratio falls to about 1.4 ^[19]. The date the gap closes has moved the other way. The sold-out horizon moved out by about a year in six months ^[3], so the backlog grew by roughly two months for each month on the calendar ^[20]. "We are completely sold out, so there's no end in sight," Hurlston told Bloomberg ^[5].

The buyers are paying toward the catch-up. Nvidia's $2 billion came with multibillion-dollar purchase commitments attached ^[6], and hyperscalers have agreed to take on some of the capital-expenditure risk ^[9]. "These hyperscalers are willing to put money behind it," Hurlston said ^[10]. Lumentum's own stated commitment of at least $350 million for its Tokyo-area sites ^[8] is about 35% of one quarter's revenue ^[16]. Nvidia's investment is about 5.7 times as large ^[17].

For new product lines, Lumentum plans to buy, and its own plant money goes to parts it already sells out ^[8][14]. "I think probably in the next year, we'd like to add additional capability to what we currently have," Hurlston said, according to Bloomberg ^[15]. The June quarter's $1.01 billion compares with about $483 million a year earlier ^[21], and the company credited most of that growth to components ^[11].

Whether optics now limits how fast AI data centres get built is a larger claim than these figures support. The 70% applies to some products ^[1], mostly indium phosphide devices such as EML chips and pump lasers ^[4], at one supplier. Bloomberg reported that the US may move ahead with a ban on new Chinese optical transceiver models ^[13]. Counterpoint Research, cited by Bloomberg, said Lumentum and Coherent could not fill the gap left by Chinese makers such as Innolight ^[13]. Nvidia has put money into both of those two this year, according to Quartz ^[7]. The sources do not show any data-centre project delayed for want of optical parts. They do show a long lag: building new capacity takes three to five years for companies like Lumentum, according to Bloomberg ^[12].

If the projected fall from 70% to 30% holds ^[2], the squeeze is a one-year problem in specific laser chips. A window that keeps sliding as it did between April and October ^[3] would mean demand is outgrowing a Tokyo plant that has already scaled twelvefold in two years ^[8]. The transceiver ban, if it goes through, would widen the gap beyond what Counterpoint says Lumentum and Coherent can cover ^[13]. I think the evidence points to the second. The one revision on record moved the horizon out by a year in six months ^[3], and customers are putting money toward capacity they cannot yet buy ^[9][6]. The case against is that every figure here comes from the chief executive in an interview ^[1], and a sold-out supplier has reason to tell its buyers so. A 2027 gap below 70% at the next update ^[2], or components revenue that stalls while Lumentum still calls itself sold out, would show the second reading is wrong.