Key Takeaways:
AI data center operators are hitting a bandwidth wall, and the fastest path over it runs through glass, not copper.
Coherent's datacenter segment surged 40.6% to $1.36 billion in the latest quarter, now 75% of total revenue, as hyperscalers race to replace copper interconnects with optical links that can handle 800G and 1.6T data rates. The shift is no longer speculative — it is showing up in earnings, capital expenditure, and supply agreements across the optical chip supply chain.
"The pipe between GPUs is where the next leg of AI capex lands," Coherent management said on the earnings call, pointing to a $2 billion Nvidia investment tied to US manufacturing capacity. The company guided Q4 FY26 revenue to $1.91 billion to $2.05 billion, with non-GAAP operating margin expanding to 20.3% from 18.6% a year earlier.
The data is visible across the stack. Fabrinet's capital expenditure nearly doubled to $63.8 million year-over-year, a signal that contract manufacturing capacity is filling ahead of schedule. POET Technologies expects to ship more than 30,000 optical engines in 2026, backed by $430 million in cash after raising about $375 million gross. nLIGHT posted a 160.7% EPS beat on revenue up 55.2%, with defense product revenue nearly doubling to a record $33.1 million. IPG Photonics, the fiber laser incumbent many investors wrote off, saw emerging growth products reach 53% of revenue as it redeploys its laser stack into datacenter photonics roadmaps.
The optical interconnect market is emerging as one of the most strategically important segments in AI infrastructure. Nvidia has committed roughly $4 billion across component makers Coherent and Lumentum to lock up supply, while Elon Musk received FTC clearance to acquire Mesh Optical Technologies, a startup founded by former SpaceX engineers that develops 800G and 1.6T optical transceivers drawing about a third of the power of competing modules. Japan's NTT established the $500 million IOWN Fund to accelerate the global transition from copper to light-based AI clusters. The spending wave has not crested.
The technical bottleneck is straightforward. Per-lane signaling in copper interconnects is climbing toward 200 gigabits per second, but attenuation, crosstalk, and the skin effect worsen at higher frequencies, making passive copper impractical beyond a meter or two. Optical links solve this by converting electrical signals into light for transmission over fiber, carrying far more data over longer distances while consuming less power.
This physics problem is driving a structural shift in how AI clusters are built. Training and inference tasks on frontier models are split across tens of thousands of GPUs using parallel-computing techniques, requiring processors to exchange enormous volumes of data every fraction of a second. The industry calls this mismatch the "I/O wall" — per-chip compute has raced ahead, but the bandwidth linking those chips has not kept pace.
Chipmakers and networking vendors are racing to deliver faster 800G and 1.6T optical transceivers while shortening electrical paths with co-packaged optics, which place the optical engine alongside the switch ASIC. Tower Semiconductor said it has shipped more than five million coherent photonic integrated circuits to Marvell Technology for use in optical transceivers that link data centers. The technology must precisely control both the phase and polarization of light, making these chips harder to build than simpler semiconductor types.
The AI datacenter bandwidth story is not a single-name trade. It spans five layers: interposer engines (POET Technologies), merchant transceivers (Coherent), contract manufacturing (Fabrinet), fiber laser reinvention (IPG Photonics), and the defense-plus-datacenter wild card (nLIGHT). Each occupies a different position in the optical supply chain, and each offers a different risk-reward profile.
Coherent, now an S&P 500 component, trades at roughly 38 times forward earnings — the premium for leadership in a market where Nvidia is a customer, not a competitor. Fabrinet has pulled back nearly 19% in the past month despite its capex signal, with the average analyst price target at $749.11. POET, the smallest name on the list, is up 16% year to date after a 24% monthly pullback, with $430 million in cash funding its production ramp. nLIGHT is up 87.8% year to date, and its datacenter angle remains an option the market has not yet priced in.
For investors, the question is which layer of the stack offers the best entry. Coherent's segment growth, Fabrinet's capex doubling, and nLIGHT's triple-digit EPS surprise all point at the same underlying reality: the pipe between GPUs is where the next leg of AI capital expenditure lands, and the window to position ahead of the next round of hyperscaler announcements is narrower than the recent pullbacks suggest.
This article is for informational purposes only and does not constitute investment advice.
