AI has pushed the bottleneck from GPUs to networking, and that shift is making indium phosphide lasers and optical transceivers the next constraint. NVDA and PLTR are used as proof of what happens when a company is built for the fastest-growing market early, then the video argues COHR and LITE sit in the same kind of setup for AI networking. The case hinges on exploding demand for 400G, 800G, and 1.6T optical links, a supply chain still dominated by a few indium phosphide makers, and two vendors with heavy AI exposure. The risk is the same as the upside: if hyperscaler spending slows, both COHR and LITE get hit immediately.
A $10,000 NVDA investment at the AI era’s start is now over $125,000, while $10,000 in PLTR from 4 years ago would be close to $250,000.
ChatGPT’s release almost 4 years ago exposed compute as the first AI bottleneck: GPU count and GPU speed limited training and prompting latency.
Hopper, Blackwell, and Rubin pushed NVDA’s chips so far that the bottleneck shifted from chips to the network.
Optical networking replaces copper with light in glass, with 400G, 800G, and 1.6T ports; 1.6T equals 1,600G and a 400G link is 400x the speaker’s 1-gigabit home internet example.
A 1.6 terabit transceiver uses eight laser chips at 200G each; with two transceivers per fiber cable and three cables per GPU, that implies 48 indium phosphide laser chips per GPU.
Inside-rack GPU links are still on copper and carry about 9x the bandwidth of the inter-rack network; moving those links to fiber would require many more lasers and about 20 additional kilowatts per rack.
Claude Fable 5 was described as so powerful that the U.S. government forced Anthropic to keep it offline until just a few weeks ago.
The optical transceiver market is forecast to rise from $23 billion last year to $112 billion in 2031, a 30% CAGR that is said to be 2x the S&P 500’s pace over the last 10 years.
Most of the world’s indium phosphide supply comes from just three companies: Sumitomo and JX in Japan, and AXT in Beijing; indium phosphide is on China’s export controls list.
COHR gets more than 70% of revenue from data centers, while LITE’s revenue rose 83% year over year even after it stopped reporting its data-center revenue line separately.
COHR reported $2 billion in revenue last quarter, up 34% year over year, with 38.5% gross margin, and full-year adjusted EPS of $5.61 versus $3.53 the year before.
Over the last year, COHR generated $5.3 billion of its $7.1 billion revenue from the data center and communication segment, while the older industrial laser business shrank.
COHR moved indium phosphide production from 2-in to 4-in to 6-in wafers, says 6-in lines make 4x more chips at about half the cost, and expects to double indium phosphide output by the end of this year and double it again by end-2027.
NVDA put $2 billion into COHR and owns just under 4%; it also signed a multi-year agreement with a multi-billion dollar purchase commitment and access to five more COHR co-packaged optics product lines.
LITE reported $1 billion in revenue last quarter, up 109% year over year, with adjusted EPS of $3.23 versus $0.88 and adjusted gross margin of 50.4%.
LITE’s main product is the EML laser; it also makes ultra-high-power lasers for silicon photonics and pump lasers for telecom amplifiers, and it is effectively sold out of pump lasers.
Co-packaged optics can cut network power by 3.5x and use 4x fewer lasers; Nvidia says that saves about 13 kilowatts on a Grace Black rack, roughly 10% of the rack’s power budget.
NVDA invested $2 billion in LITE on the same day and on almost the same terms as the COHR investment.
If only one of COHR or LITE is chosen, COHR is the pick because it built the world’s first production line for 6-in indium phosphide wafers and is on track to quadruple capacity by end-2027.
Both COHR and LITE are framed as worth spots on every long-term investor’s watch list, but both are highly exposed to hyperscaler spending and could be hit hard if that spending slows.
Outskill is the sponsor and pitched a Claude AI Mastery Workshop this coming weekend, 10:00 a.m. to 7:00 p.m. Eastern, with 2-day training, the first 1,000 signups getting a free seat, and bonuses including 50 secret Claude codes, a full AI prompt library, and a personalized AI toolkit builder.