Last week we covered photonics testing. The week before, power transformers.. This week the bottleneck is moving to optical fibers.
You probably have optical fiber running into your home. It's what "fiber internet" means, and it's why the connection is fast. A hair-thin strand of glass carrying your data as pulses of light.
That same glass is now a binding constraint inside the most expensive buildout in tech.
In an AI data center, fiber becomes the high-speed wiring between servers, racks, and switches across the cluster. A training cluster runs tens of thousands of GPUs that have to behave like one machine, constantly shuttling enormous volumes of data to each other.
Past short distances, copper can't carry that much bandwidth without becoming too hot or losing signal. Light through glass can. So inside the cluster, especially between GPU racks, the connections are via glass fiber.
And AI is bending the demand curve. A GPU rack needs roughly 36x more fiber than a traditional CPU rack and the clusters keep getting bigger.
The result is a shortage.
Optical fiber demand is moving to the data center
For decades, optical fiber was a telecom product. The buyers were phone companies and carriers running it into homes, across cities, and along the long-haul backbone. Demand grew steadily with the rise of broadband and 5G.
That center of gravity is now shifting. The marginal growth buyer is increasingly the data center, and AI clusters consume fiber at a very different intensity than legacy telecom networks. Data centers are expected to swing from under 5% of global fiber demand in 2024 to about 30% by 2027.
The clearest signal is in the earnings of the largest US optical fiber maker, Corning (GLW):
Optical revenue is compounding: Optical Communications grew 36% year over year in Q1 2026 to $1.85B, with segment net income up 93%. Carrier and enterprise both grew 36%.
Buyers are pre-funding supply: Meta signed a multiyear deal worth up to $6B in January and anchored a plant expansion. On the Q1 call, Corning disclosed two more long-term agreements with unnamed hyperscalers, each comparable in size to the Meta deal. Nvidia followed with a partnership worth up to $3.2B tied to a 10x US optical-capacity expansion. Amazon added a multi-billion-dollar supply deal this month.
Management is racing to keep up: Pressed by analysts on the supply-demand imbalance and on international rivals already raising fiber prices, CEO Wendell Weeks described demand robust enough to require expansion across all of Corning's major optical operations, fiber included, with customers sharing the cost.
GLW Q12026 Earnings Takeaway on Tessara
Nobody locks up years of supply and helps pay for the factory unless the thing is scarce.
So how is fiber actually made, and where does it jam?
How fiber is made, and the bottleneck inside it
Optical fiber is not manufactured strand by strand. It's drawn i.e. pulled, hot and thin, from a much larger piece of glass called a preform.
A preform is a precision glass rod, often a meter long and a few inches across, with its light-guiding core built up chemically, layer by vaporized layer. You hang it at the top of a draw tower, heat the tip past 2,000°C, and pull a single continuous fiber out of the bottom, hundreds of kilometers of hair-thin glass from one rod.
Fiber Drawing Process (Source: hkcablemachine.com)
That rod is the bottleneck. It's the single largest cost in a finished fiber and the part that takes years of process know-how to make well. Draw towers are quick to add but preform capacity is not. When a fiber maker talks about expanding, preform is the piece that actually gates it.
And the makers are few. Only about 20 preform producers exist worldwide, the top five hold close to 60% of capacity, and the map tilts east:
China leads: YOFC trades the global #1 spot with Corning (USA) year to year, alongside Hengtong, FiberHome and ZTT.
The US narrows to one name: Corning, which is vertically integrated, making the rod, drawing the fiber, and jacketing the cable in-house.
Why you can't just make more
Adding preform is slow and hard for reasons that don't yield to capital:
The build time is ~18–24 months per line and that's before the output is qualified to a customer's spec. Capacity announced in 2025 doesn't ship fiber until 2026–27, and the large Western additions mostly land after 2028.
The know-how doesn't transfer: Preform chemistry is a decades-deep craft held by a couple dozen firms. Which is why a demand spike goes straight into price. Mainstream fiber hit a seven-year high in early 2026, up about 75% in a single month, with data-center-grade specialty fiber running harder still.
Existing capacity gets cannibalized: With preform maxed out, makers divert their limited rod toward high-margin specialty and data-center fiber, starving the standard telecom fiber underneath and spreading the shortage sideways.
The key dopant is a Chinese chokepoint: The light-guiding core is doped with germanium, delivered as germanium tetrachloride. China controls roughly 60% of germanium production and put it under export licensing in August 2023; the price has about tripled since early 2024. Helium and silicon tetrachloride, the other draw inputs, are up ~135% and ~50%. The rod that gates the fiber is itself gated by inputs that are tightening and partly sitting behind a license Beijing can pull.
Those forces are global. The US buildout mainly favours one name. Corning, the lone integrated domestic maker, is expanding straight into the shortage.
Its May partnership with Nvidia funds three new plants in North Carolina and Texas and lifts US fiber capacity more than 50%, the preform behind it included.
And it builds into a market tilting domestic. Chinese-origin fiber lands behind Section 301 tariffs, which went higher in the 2025 escalation. This raises the cost of imported supply for any buyer, data centers included. There's no domestic mandate on a private data center, but the hyperscalers are reshoring by choice anyway, locking in US-made fiber through the very deals driving the shortage.
None of this makes Corning a hidden bet because the move is largely priced in. What isn't likely priced is how huge the demand could get in the next few years as data centers explode. Demand is still climbing against preform that doesn't fully arrive until 2027 and beyond. While that gap stays open, pricing power sits with whoever holds the rod.
Corning (GLW) price is up over 100% since the beginning of the year
Who pays, who captures
Pays rent: the AI hyperscalers and the telcos and colos building for them — Microsoft, Google, Amazon, Meta, OpenAI and the operators between them, all of whom need fiber that has to be drawn from a rod that doesn't yet exist.
Captures rent: the preform holders at the glass layer. YOFC and the Chinese majors, Shin-Etsu and Sumitomo in Japan, Corning in the US
What to watch
Preform capacity timing: New rod lines confirmed for 2026–27, and whether the big Western additions still read "after 2028." That date is the constraint.
The price spread between standard and specialty/data-center fiber. A widening gap is the cannibalization signal.
Corning Q2 (Aug 4): optical margin and how fast the new US capacity actually ramps.
The Week Ahead
A few important earnings this week. Read our post and pre-call briefs here and stay prepared.
Earnings Tab on Tessara
Wednesday, June 24
MU (Micron) — Memory. Watch: HBM4 qualification and customer expansion
Want the live data behind The Chokepoint? Tessara is the research terminal for the AI buildout. We track what's binding in the supply chain and what it means for what you own. 300+ companies across compute, memory, foundry, networking, and power.
See you next week,
Teng & Arvind
This article is for informational and research purposes only. It is not financial advice, investment advice, or a recommendation to buy or sell any security. Tessara Research does not publish price targets. The views expressed here reflect our analysis at the time of publication and may change as new evidence arrives. Readers should do their own research and consult a qualified financial adviser before making investment decisions.