AI data center networking and optics: transceivers, lasers, suppliers

AI clusters link thousands of chips with switches and fiber, and the optical transceivers at each end depend on indium phosphide lasers that are in short supply. Filings from Lumentum and Coherent show optics revenue rising fast, Nvidia has put 2 billion dollars into each of them, and new laser factories will not reach volume until 2027 and 2028.

Key numbers at a glance

Figures are tagged with source and date. Some come from company filings. Others come from trade commentary, and those are marked.

• Lumentum quarterly revenue, quarter ended 27 June 2026: 1,006.3 million dollars, up 109.3% from a year earlier (Lumentum Form 8-K exhibit, 11 August 2026).
• Lumentum fiscal 2026 revenue: 3,014.0 million dollars, up 83.2% (same filing).
• Lumentum next quarter revenue guidance: 1.225 to 1.275 billion dollars (same filing, company guidance).
• Coherent Datacenter and Communications revenue, quarter ended 30 June 2026: 1,615.0 million dollars, against 1,018.3 million a year earlier (Coherent Form 8-K exhibit, 12 August 2026).
• Coherent spending on property, plant and equipment, fiscal 2026: 1,102.9 million dollars, against 440.8 million in fiscal 2025 (same filing).
• Nvidia investment in Lumentum: 2 billion dollars for 2,876,415 shares of Series A convertible preferred stock at 695.31 dollars a share (Lumentum Form 8-K, 2 March 2026).
• Nvidia investment in Coherent: 2 billion dollars, plus a multi-billion dollar purchase commitment (Coherent and Nvidia release, 2 March 2026).
• Lumentum's new indium phosphide laser plant in Greensboro, North Carolina: production expected to ramp in mid 2028 (Lumentum release, 26 March 2026).
• Ethernet optical transceiver market growth: almost 90% in 2025, forecast 120% in 2026 and 76% in 2027 (LightCounting, as reported by Lightwave, 8 October 2026; analyst estimate and projection).
• High-speed datacom optics demand above supply: roughly 30% (Comsoc technology blog by Alan Weissberger, 14 August 2026, no underlying source named).
• Share of global refined indium produced in China: roughly 70% (LEDinside, 6 August 2026).
• Arista Q2 2026 revenue: 3.036 billion dollars, up 37.7% (Arista Form 8-K exhibit, 4 August 2026).
• Nvidia supply commitments: 279 billion dollars at 26 July 2026, up from 119 billion a quarter earlier, mainly memory and manufacturing facilities (Nvidia Form 10-Q for the quarter ended 26 July 2026).

The network inside an AI data center, in plain words

A modern AI model is trained or served on thousands of accelerator chips that must exchange data constantly. If the network is slow, the expensive chips sit idle. So the network is treated as part of the computer.

There are three layers in the way most people describe it. Inside a rack, chips talk over very short links, often copper. Between racks, switches route traffic across a hall, and these links are mostly fiber. Between buildings and sites, longer links use fiber and more advanced optics.

People in the industry also talk about scale up and scale out. Scale up means joining a small group of chips so tightly that they behave almost like one large chip, as Nvidia's NVLink does inside a rack. Scale out means joining many such groups across a hall with Ethernet or InfiniBand switches. Today most optics sit in the scale out network, but as scale up domains grow past one rack, optics move into that layer too.

A switch is a box that sends data packets to the right destination. A big AI cluster uses many of them arranged in layers, often called leaf and spine.

An optical transceiver is a small pluggable module that sits in a switch or server port and turns electrical signals into light, and back. At each end of a fiber link there is one. Speeds are written as 400G, 800G and 1.6T, meaning 400, 800 or 1,600 gigabits per second per port. A big cluster can need hundreds of thousands of them.

Inside a transceiver is a laser that makes the light, a driver chip, a receiver and some digital signal processing. Higher speed modules use several lanes, each with its own laser and each running faster than before. The lasers are the hardest part to scale.

Why not just use copper everywhere? Nvidia's blog of 16 June 2026 gives the plain answer. Copper cannot span the distances needed for its 576 GPU NVL576 configuration, which spreads across eight racks. Light has a one time conversion cost at each end, after which extra distance adds little cost, and Nvidia says light is the most power efficient option at that scale. Every pluggable optical module in such a network carries an indium phosphide laser, according to the same post.

Why lasers set the pace

Many of the lasers in these modules are made from indium phosphide, written InP. It is a compound semiconductor, grown as wafers like silicon but made in far smaller volumes and in fewer factories. The LEDinside piece of 6 August 2026 describes InP as the only commercially proven substrate for high-performance EML chips in long haul coherent communications. An EML, or electro-absorption modulated laser, is a laser with a built-in light modulator that can run at very high speeds.

Wafer size is part of the problem. Nvidia's June 2026 post notes that silicon factories use 12 inch wafers while most InP production still uses 3 and 4 inch wafers. A 6 inch wafer has roughly four times the usable area of a 3 inch one, which is why both Lumentum and Coherent are moving to 6 inch InP lines.

The LEDinside article quotes Lumentum's chief executive as saying the supply and demand gap for InP has surpassed that of DRAM and NAND, which are the memory chips that have had their own shortages. It says imbalance is expected to persist beyond 2027, that customers are ordering further ahead and paying substantial deposits, and that China produces roughly 70 percent of global refined indium and has tightened export controls on InP and indium since 2025. The article does not give shipment volumes, lead times in weeks or details of the controls.

The Comsoc blog of 14 August 2026 says current high-speed datacom optics demand exceeds available supply by roughly 30 percent. It does not name a source for that number, so treat it as an estimate from one commentator. It also reports that LightCounting, a market research firm, found cloud builders widening their supplier pools from two or three traditional optical suppliers to five to seven specialized manufacturers, which is what you would expect when the usual suppliers cannot deliver enough.

An alternative to the InP laser approach is silicon photonics, which builds optical parts on silicon chips in standard chip factories. It still needs a laser, usually a continuous wave one that is supplied separately, so it moves the pressure rather than removing it. The Comsoc blog reports that Tower Semiconductor had triple digit year over year growth in silicon photonics revenue and that GlobalFoundries plans to double its silicon photonics business in 2026, both as company statements.

Two ideas aim to cut the number of optical parts or the power they draw. Linear pluggable optics, or LPO, drops some of the signal processing from the module. Co-packaged optics, or CPO, moves the optics next to the switch chip. Arista's 4 August 2026 release says its new 7060XE7 switches, with up to 100 Tbps of system bandwidth, use LPO that the company says cuts interconnect power by about 60 percent. The Comsoc blog says CPO can cut optical interface power by up to 75 percent, but gives no source. Lumentum's August 2026 release cites increasing demand for ultra-high-power CPO lasers as an early sign that optics are moving closer to the processors, and Nvidia's June post says Coherent supplies external laser modules for Nvidia's co-packaged optics switches.

Nvidia's 2 billion dollar bets on laser makers

On 2 March 2026 Nvidia announced matching deals with two United States laser makers, Lumentum and Coherent.

Lumentum's filing that day shows Nvidia buying 2,876,415 shares of new Series A convertible preferred stock at 695.31 dollars a share, for 2 billion dollars in cash, in a private placement. The joint release describes a multi-year, non-exclusive agreement with a multi-billion dollar purchase commitment and future capacity access rights for advanced laser components. Lumentum said the money helps fund a new fabrication facility.

Coherent's release the same day says Nvidia is investing 2 billion dollars in Coherent to support research and development, future capacity and operations. It also describes a non-exclusive agreement with a multi-billion dollar purchase commitment from Nvidia and future access and capacity rights for advanced laser and optical networking products. The release does not say what form the investment takes.

A purchase commitment is a promise to buy a set amount over time. Capacity access rights give the buyer a claim on future output. Together they are how a large customer makes sure that new factory lines it helps fund will build for it first. LEDinside had reported in August that Nvidia committed 2 billion dollars to each of two key suppliers without naming them; the two filings match that description.

New laser capacity and when it arrives

Lumentum, Greensboro. On 26 March 2026 Lumentum said it was buying a 240,000 square foot operating plant in Greensboro, North Carolina, from Qorvo. It will be retrofitted to make InP continuous wave and ultra-high-power lasers on 6 inch wafers, with production expected to ramp in mid 2028. Lumentum said it plans to invest hundreds of millions of dollars over several years, preserve and create more than 400 jobs, and it named Nvidia as a customer of the site.

Coherent, Sherman, Texas. Nvidia's 16 June 2026 post covers the groundbreaking for an expanded building at Coherent's Sherman campus, which Coherent describes as the world's first volume production 6 inch InP fab. Coherent announced a letter of intent for up to 50 million dollars in CHIPS Act funding, on top of about 17 million dollars in earlier Texas and local support, and said the site will support more than 550 direct jobs at full capacity. A letter of intent is not a final grant agreement.

The timing matters. Factories announced in 2026 add little output in 2026. LEDinside says expansions run into 2027 and 2028, and Lumentum's own date for Greensboro is mid 2028. In the meantime, extra supply has to come from running existing lines harder and from more module assemblers.

What the company filings show

Lumentum. Its 11 August 2026 filing shows revenue of 1,006.3 million dollars for the quarter ended 27 June 2026, up 109.3 percent from a year earlier and 24.5 percent from the previous quarter, and 3,014.0 million dollars for fiscal 2026, up 83.2 percent. Components were 649.4 million dollars of the quarter and systems 356.9 million dollars. It guided the next quarter to 1.225 to 1.275 billion dollars and listed cloud module 1.6T adoption and optical circuit switch solutions among growth drivers.

Its balance sheet hints at expansion. Net property, plant and equipment rose to 1,159.1 million dollars from 726.4 million, and inventories rose to 691.6 million from 470.1 million. The filing does not quantify laser output, and its risk factors flag the company's ability to increase manufacturing capacity and the ability of suppliers and contract manufacturers to deliver, along with tariffs, export controls and rare earth minerals.

Coherent. Its 12 August 2026 filing reports fourth quarter fiscal 2026 revenue of 2,046 million dollars, up 34 percent from a year earlier. The Datacenter and Communications segment brought in 1,615.0 million dollars, against 1,018.3 million a year earlier, and 5,274.6 million for the full year, against 3,755.2 million. The Industrial segment shrank over the same period, so the growth is concentrated in data center optics. Spending on property, plant and equipment was 1,102.9 million dollars in fiscal 2026, against 440.8 million the year before. Coherent guided the next quarter to 2.2 to 2.4 billion dollars. Its chief executive said AI data center architectures are increasingly moving from copper to optical connectivity.

Arista. Its 4 August 2026 filing reports second quarter revenue of 3.036 billion dollars, up 37.7 percent from a year earlier and 12.1 percent from the first quarter, its first quarter above 3 billion dollars. It guided the third quarter to about 3.3 billion dollars and introduced 1.6 Tbps AI fabric platforms, some liquid cooled.

Cisco and Nvidia, via secondary reports. The Comsoc blog relays Cisco's networking revenue up 28 percent with orders up 40 percent. MarketsDaily summarizes Nvidia's fiscal second quarter 2027 results, reported 26 August 2026, as a record networking quarter up 18 percent from the previous quarter, with Spectrum-X Ethernet revenue up 2.6 times from a year earlier. Nvidia's own 10-Q for that quarter reports total revenue of 96,221 million dollars but does not split networking out in the segment note, so the networking figures rest on those summaries. Check them against the companies' own releases.

Demand behind it: capital spending

The Comsoc blog relays Amazon's second quarter 2026 capital spending of 54.2 billion dollars, up 73 percent, with 2026 guidance raised from 200 to 220 billion dollars, and says capital spending as a share of revenue was 51 percent at Meta and 46 percent at Microsoft in the second quarter. It also says telecom operators' capital spending for 2026 is expected to be flat or down about 1 percent. Spending by cloud and AI builders is rising fast while spending by traditional carriers is not, so the optics industry is being pulled by one buyer group.

LightCounting's latest view, as reported by Lightwave on 8 October 2026, is that the Ethernet optical transceiver market roughly doubled in 2024, grew almost 90 percent in 2025 and will grow about 120 percent in 2026 before slowing to 76 percent in 2027 and 20 percent by 2031. The firm said the first nine months of 2026 derailed its earlier expectation of slower growth. It also warned that the supply chain is "clearly overreacting now", that 2004 to 2025 history shows two or three year growth runs broken by flat or negative years, and that its base case is a soft landing with a sharper correction as the alternative. These are analyst projections, not company data.

Nvidia's 10-Q adds a supplier side signal: its supply commitments rose from 119 billion dollars to 279 billion dollars in one quarter, described as primarily memory and manufacturing facilities. Not all of that is optics, but it shows how far ahead the largest buyer is booking.

There is a counterweight. LightCounting notes that TPU based clusters and multi-plane networks need fewer transceivers per accelerator, and that high-radix switches, meaning switches with many ports, can flatten networks and remove layers. So transceiver demand does not rise one for one with chip counts.

Who supplies what

The sources cited here name these companies. This is a list, not a ranking.

• Lasers and optical components: Lumentum and Coherent. LEDinside also names Yunnan Germanium as China's only mass producer of 6 inch InP substrates, now expanding, and Yuanjie Semiconductor as having commercialized 100G EML chips.
• Transceiver module assemblers: Innolight, which reported close to 200 percent year over year growth in optical transceiver sales for the first half of 2026, as cited by LightCounting.
• Silicon photonics foundries: Tower Semiconductor and GlobalFoundries, per the Comsoc blog.
• Switches: Arista, Cisco and Nvidia, with Nvidia's networking including Spectrum-X Ethernet.
• Market research referenced: LightCounting and Dell'Oro Group.

None of these sources gives market shares, and this article makes no claim about them.

What is firm and what is uncertain

Firm enough. Cloud capital spending is rising sharply. Lumentum's revenue more than doubled year over year and Coherent's data center segment grew strongly, both per their filings. Nvidia has invested 2 billion dollars in each of them and signed purchase commitments. New 6 inch InP lines are being built, with Lumentum's Greensboro ramp dated to mid 2028.

Uncertain. The size of the shortage, since the 30 percent figure has no stated source. Exact lead times, which none of the cited sources gives in weeks. How quickly new capacity helps. How much silicon photonics, LPO and CPO reduce laser demand. Nvidia's exact networking revenue, which comes from secondary summaries. Whether LightCounting's 2027 slowdown arrives as a soft landing or a correction.

What to watch next

• Lumentum's and Coherent's next quarterly filings, including capital spending, inventory and data center segment revenue.
• Progress at Greensboro and Sherman, and whether Coherent's CHIPS letter of intent becomes a final award.
• Chinese export control changes touching indium and InP.
• Whether co-packaged optics move from announcements to volume shipments.
• Whether the share of AI networking that uses Ethernet continues to rise, as Nvidia's Spectrum-X number suggests.

Related reading: AI chip market share in 2026, Anthropic's compute commitments, OpenAI's Stargate and 26 gigawatts and how power equipment is supplied.

Safeguard does not make network hardware, but it can scan the dependencies of software that manages it and build an SBOM, and there is a free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• Lumentum Form 8-K exhibit, fourth quarter and fiscal 2026 results, 11 August 2026
• Lumentum Form 8-K, sale of Series A convertible preferred stock to Nvidia, 2 March 2026
• Lumentum and Nvidia joint release (Form 8-K exhibit 99.1), 2 March 2026
• Lumentum, new U.S. manufacturing facility for advanced lasers, 26 March 2026
• Coherent Form 8-K exhibit, fourth quarter and fiscal 2026 results, 12 August 2026
• Coherent and Nvidia release (Form 8-K exhibit 99.1), 2 March 2026
• Nvidia blog, Coherent breaks ground on expanded Texas facility, 16 June 2026
• Arista Networks Form 8-K exhibit, second quarter 2026 results, 4 August 2026
• Nvidia Form 10-Q, quarter ended 26 July 2026
• Lightwave (Sean Buckley), Ethernet optical transceiver set to grow 120% in 2026, 8 October 2026
• Comsoc Technology Blog (Alan Weissberger), Impact of optical component shortages and bottlenecks explained, 14 August 2026
• LEDinside (TrendForce), InP shortage emerges as AI optical interconnect bottleneck, 6 August 2026
• MarketsDaily, Nvidia Q2 earnings call highlights, August 2026

Last reviewed: 19 October 2026.