Coherent's PhotonLink and the Optical Bottleneck You Can't See From the Top

Coherent will unveil a product called PhotonLink on September 21 at a European optics conference in Málaga. It is an integrated optics platform covering the full signal chain — light generation, beam shaping, transmission, detection, and conversion back to an electrical signal. It works with co-packaged optics, near-packaged optics, and other architectures.

The name is new. The idea behind it has been in the headlines for six months, and the money behind it arrived in March when NVIDIA committed $2 billion to Coherent.

What matters for an investor is not the platform name. What matters is whether Coherent sits at a genuine chokepoint in the AI optical interconnect chain, and whether a stock trading around $297 — up roughly 210% over the past year — has already priced the answer yes.

To see the chain, start with what the market already knows. AI training clusters are scaling from thousands to tens of thousands of GPUs. The copper wires that connect them are hitting physical limits at 800 gigabits per second and beyond. At that speed, electrical links consume too much power and degrade over even short distances. The only practical alternative is light.

That means optical transceivers — modules that convert electrical signals to light and back. And they need to get faster: from 800G to 1.6T to 3.2T and eventually 12.8T per module. Each lane in those modules needs its own laser to generate light and a detector to receive it. A 1.6T module uses eight lanes, nearly doubling the number of lasers per module versus 800G.

The global optical transceiver market is on track to more than double between 2026 and 2036, with AI network demand as the fastest-growing segment. Coherent reported that its datacenter and communications segment brought in $1.6 billion in its last quarter, up from $1 billion a year earlier. Full-year datacenter revenue hit $5.3 billion, up from $3.8 billion.

But here is where the visible bottleneck ends and the real one begins.

Every high-speed optical transceiver, including the ones built on silicon photonics, depends on one material: indium phosphide, or InP.

Silicon can route, modulate, and split light. It cannot generate it. So even in a silicon photonics architecture — which is exactly where the industry is moving — an InP-based laser die is bonded onto the silicon chip. The more silicon photonics is adopted, the more InP lasers are needed, not fewer.

The problem is that InP epitaxial growth — the process of growing the layered crystal structures that become lasers — has inherently lower throughput and yield than silicon wafer production. It is done by a small number of specialty manufacturers, not a broad foundry ecosystem. Equipment for InP fabs has lead times of 18 to 24 months. New fabs take years to build. Skilled photonics labor is scarce.

Lumentum's CEO Michael Hurlston told industry reporters in August that the supply-demand gap for InP has now surpassed that of DRAM and NAND. Current transceiver demand exceeds supply by a factor of two.

China controls roughly 70% of global refined indium production and has tightened export controls on InP-related materials since 2025. The substrate and epitaxy layers of the supply chain touch China and Taiwan, and most fabs hold only weeks of inventory. Months of friction could halt downstream production of transceivers entirely.

This is the real bottleneck. The transceiver companies — Coherent, Lumentum, Broadcom — are the visible layer. The constraint lives beneath them, in a narrow stack of crystal growth and epitaxial fabrication that cannot be expanded quickly and cannot be bypassed.

Coherent's structural advantage is that it owns more of this stack than most competitors. The company has vertical integration across InP substrates, epitaxy, and electro-absorption modulated lasers (EMLs). When other optical module makers must buy their lasers from a constrained third party, Coherent controls its own supply from substrate to finished device.

This is not a monopoly — Lumentum also operates multiple InP wafer fabrication facilities and dominates high-margin InP EMLs for 1.6T transceivers. But Coherent's broader vertical integration, combined with its CPO integration and optical packaging capabilities, gives it depth that is hard to replicate quickly.

NVIDIA appears to recognize this. On March 2, 2026, NVIDIA announced parallel $2 billion investments in both Coherent and Lumentum — $4 billion total. The Coherent deal is a multiyear, nonexclusive agreement running through 2030, with a multibillion-dollar purchase commitment. Revenue from the new co-packaged optics and advanced laser products is expected to start flowing in calendar 2027.

The deal is not circular. NVIDIA provides equity capital, Coherent builds capacity, NVIDIA buys product under a long-term contract. It is the same playbook NVIDIA used with CoWoS packaging capacity and HBM memory: pre-fund the constraint before it stops production.

So the chain checks out. InP is real and constrained. Coherent has vertical integration that matters. NVIDIA has locked in capacity through 2030.

But structure and valuation are two different questions.

Coherent's stock trades around $297 with a market capitalization of roughly $68 billion. That implies a trailing P/E ratio near 100x. Even on a forward basis, using the company's own FY2027 guidance of $2.2 to $2.4 billion in quarterly revenue and non-GAAP margins of 39.5% to 41.5%, the earnings multiple remains well above 60x.

The stock is up approximately 210% over the past twelve months. Year-to-date, it has returned 52%. One month before the PhotonLink announcement, it had already surged more than 21%.

Then there is the margin question. Coherent beat earnings expectations in its last reported quarter — $2.05 billion in revenue, non-GAAP EPS of $1.74 versus a consensus of $1.62. And yet the stock fell roughly 5% in premarket trading. The reason was not revenue. It was margins.

Analysts flagged that the datacenter transceiver business is lower-margin than Coherent's traditional industrial photonics, and the mix shift toward high-volume, lower-margin transceiver products is compressing gross profitability. The "marginal beat" in gross margins left investors wanting more, according to Barclays' Tom O'Malley. Even with the NVIDIA deal's multibillion-dollar purchase commitment, Coherent risks becoming a higher-volume, lower-margin manufacturer if hyperscalers push pricing pressure through multi-vendor designs.

Here is the investment picture as it stands.

Coherent occupies a genuinely difficult position in the AI supply chain. The InP bottleneck is not theoretical — it is being experienced now, with transceiver demand exceeding supply two-to-one and industry leaders describing it as more severe than the DRAM shortage. Coherent's vertical integration through that bottleneck is a real moat, and the NVIDIA commitment through 2030 is a meaningful floor on future revenue.

But three things separate a strong structural position from an attractive stock.

First, timing. The NVIDIA revenue contribution does not start until calendar 2027. The PhotonLink platform is being unveiled in September 2026. Both are forward-looking. The stock's 210% one-year run has already absorbed an optimistic view of what comes next.

Second, execution risk. Capacity expansion in InP is hard. New fabs take years. Yield improvement is slow. The company guided to its first quarter exceeding $3 billion in revenue, which is aggressive. A miss on margins — already the concern that weighed on the stock after a strong earnings beat — at a 100x multiple is expensive.

Third, the competition is not absent. Lumentum received the same $2 billion NVIDIA investment. Broadcom has deep silicon photonics content. Asian suppliers are gaining capability. The dual-investment structure is deliberate: NVIDIA is not betting on a single supplier because single-supplier bets have burned them before.

PhotonLink is a credible platform for the architecture that is coming. But the question for an investor is not whether Coherent is important to the optical interconnect chain. It is whether you are buying a structural advantage at a price that assumes perfect execution, flawless margin expansion, and zero competitive erosion — all starting tomorrow, not in 2027.

The material is there. The stock may not be.