AI's Next Bottleneck Is Glass: A 180 Million Fiber-Kilometer Shortfall
TrendForce expects the world to need 577 million fiber-kilometers of optical fiber in 2026 and come up 180 million short, about 31% of demand. Factories are nearly full, inventories have shrunk to a month, and new capacity takes up to two years. When volume cannot grow, price does the work.
Implied supply is demand minus TrendForce's estimated shortfall (577 − 180 = 397), a NodeZero calculation. TrendForce does not publish a supply figure. Source: TrendForce, optical fiber shortage (19 October 2026).
AI data centers have spent two years competing for chips, electricity and memory. Their next scarce input is much older: glass.
Market research firm TrendForce estimates that the world will need 577 million fiber-kilometers of optical fiber in 2026 and come up 180 million short, roughly 31% of demand (TrendForce, 19 October 2026). Fiber factories are running at 92% of capacity, inventories have fallen from four months of supply to one, and building new capacity takes 18 to 24 months.
That combination produces a familiar pattern. When the amount that can be made is nearly fixed, extra demand cannot show up as extra volume. It shows up as price.
What a fiber-kilometer is
A fiber-kilometer counts each glass strand separately. A cable one kilometer long that contains 100 fibers adds up to 100 fiber-kilometers. Modern cables can hold hundreds or even thousands of fibers, so 577 million fiber-kilometers is a measure of how much glass is drawn, not of how many kilometers of cable are laid.
The same applies to miles. When Corning announced its agreement with Verizon for “80+ million miles” of fiber (Corning, 8 September 2026), it meant fiber-miles. TrendForce converts it to 129 million fiber-kilometers.
Why AI needs so much glass
An AI accelerator is only useful if data can reach it quickly. Large models are trained and run across thousands of chips, and those chips must constantly exchange data with each other, with storage, with other buildings and with other data centers. Over short distances, inside a rack, much of that traffic still runs over copper. Beyond a few meters, it runs over light in glass fiber.
TrendForce estimates that data centers account for about 18–20% of global fiber demand in 2026 and expects that share to reach 30–35% in 2027 Estimated. Hyperscale cloud providers lead, but chipmakers and telecom operators are also building links between data centers and higher-capacity backbone networks.
Source: TrendForce, optical fiber shortage (19 October 2026).
The rest of demand still comes from traditional uses such as home broadband and mobile networks. AI does not need to dominate fiber demand to cause a shortage. It only needs to add demand faster than factories can add capacity.
Why supply cannot respond quickly
Optical fiber is drawn from a large rod of ultra-pure glass called a preform. Making more preforms means installing new deposition equipment and qualifying the process, which TrendForce says typically takes 18 to 24 months. Expansions now being planned may not add much fiber before 2028.
Meanwhile, the existing system has run out of slack:
| Measure | Earlier | 2026 |
|---|---|---|
| Capacity utilization | – | 92% |
| Inventory | 4 months of supply (2024) | 1 month of supply |
| Order books | – | Some leading suppliers booked into Q1 2027 |
Source: TrendForce, 19 October 2026 Observed.
Falling from four months of inventory to one means three-quarters of the buffer is gone Calculated by NodeZero. With little stock and little spare capacity, an unexpected order cannot be filled from the warehouse or by running factories harder.
Three different price increases
TrendForce’s figures are for spot prices, the prices paid by buyers without long-term contracts, and they are expected averages for 2026 compared with 2025. Buyers with long-term agreements typically pay less. Even so, the increases are steep, and they are steepest for the fiber used inside data centers:
TrendForce describes bend-insensitive fiber as 'more than five times' its 2025 level; the bar is drawn at 5.0×. Source: TrendForce, optical fiber shortage (19 October 2026).
Bend-insensitive fiber can be routed around tight corners with little signal loss, which makes it suited to the dense cabling inside data centers. It is where AI demand is most concentrated, and where prices have risen most.
Volume capped, revenue up: the supplier view
If volume is close to its limit, a supplier’s revenue growth has to come mostly from price and mix. Corning, one of the world’s largest fiber makers, reported third-quarter 2026 sales in its Optical Communications segment up 44% to $2.38 billion. Enterprise Networks sales rose 84%, and sales of generative AI products more than doubled (Corning results, 27 October 2026) Reported.
Corning’s figures include cable and connectivity products as well as fiber, and the company does not split its growth into price and volume, so the 44% cannot be decomposed exactly. It also said it has been adding capacity, backed by long-term customer agreements that share the cost and risk of expansion.
The buyers locking up supply
When supply is scarce, large buyers stop buying on the spot market and sign long-term agreements (LTAs) instead. In the third quarter, Corning said it secured four long-term, multibillion-dollar agreements. Two are public:
- Verizon: “80+ million miles” of fiber and connectivity products for delivery from 2027 to 2032, for both home broadband and a long-haul backbone connecting AI data centers. In TrendForce’s conversion, that is 129 million fiber-km, equal to about 22% of global demand in 2026, or about 3.7% a year if spread evenly over six years Calculated by NodeZero.
- AT&T: a multi-year agreement valued at more than $3 billion for fiber and cable, mainly for expanding home and business internet (Corning, 29 September 2026).
These agreements give buyers certainty, and give suppliers the confidence to invest in new capacity. They also mean that a growing share of future output is already spoken for before it is made.
The wider shift: from copper to light
Fiber is the passive part of an optical network. The active parts, such as lasers, optical transceivers and switches, are seeing an even larger boom. TrendForce projects that AI networking and interconnect revenue will grow from $37.9 billion in 2024 to $279.8 billion in 2027, about 7.4 times in three years (TrendForce, 2 November 2026) Estimated.
2027 is a projection. The 2027 segments are TrendForce's; 'other' is the remainder, a NodeZero calculation. Source: TrendForce, AI networking and interconnect market (2 November 2026).
The same price-and-mix pattern appears here. TrendForce expects optical transceiver revenue to rise 76% in 2027, faster than shipments, because buyers are moving from 800G to 1.6T products that cost more than twice as much. Copper is not disappearing yet: TrendForce expects it to remain the main connection inside GPU racks in 2027, although Google already links its TPU clusters with optical connections.
On 2 November, chipmaker Marvell agreed to acquire Lumentum, a maker of lasers and optical transceivers, in an all-stock deal valuing Lumentum at about $115 billion (joint announcement) Reported. The companies cited “the physical limitations of copper-based connectivity” as a reason AI infrastructure is moving toward optical technology. The deal is expected to close in the first half of 2028, subject to approvals. It concerns optical components, not glass fiber itself, but it points in the same direction: AI’s data increasingly travels as light.
What could end the shortage sooner
- A slowdown in AI spending. Much of the extra demand depends on data center construction. If cloud providers cut their capital spending, demand growth would slow quickly.
- Double ordering. In a shortage, buyers often order more than they need, or from several suppliers at once, to make sure they get enough. When supply catches up, those extra orders can disappear, and apparent demand can fall faster than real demand.
- New capacity. Expansions already planned will add supply from around 2028. If several arrive at once, the market could move from shortage to surplus. The telecom boom of the early 2000s ended with a large glut of fiber.
- Technology. Denser cables and more efficient network designs could change how much fiber each data center needs, in either direction.
What the numbers do not tell us
- How the shortfall is defined. TrendForce does not publish a supply figure or explain exactly how the 180 million gap is measured. The 397 million supply figure is our arithmetic.
- Contract prices. The 2.8× to 5× figures are spot prices. Most large buyers pay contract prices, which are not public and have probably risen less.
- Forecast risk. Demand, shortfall and 2026 average prices are TrendForce estimates made in October 2026. The 2027 data center share and networking revenue are projections.
- Regional differences. Prices and availability vary by country and by fiber type; the figures here are global averages for broad categories.
- This is not investment advice. The article describes market data; it does not judge any company’s prospects or value.
What to watch next
- Fiber spot prices in early 2027. If prices keep rising while factories stay near full, the shortage is still deepening.
- Corning’s fourth-quarter results. Further growth in Optical Communications, and any new long-term agreements, would show whether demand is holding.
- Capacity announcements. New preform plants, and when they are due to start, will indicate when relief can arrive.
- AI capital spending plans. Cloud providers’ guidance for 2027 is the biggest single driver of data center fiber demand.
Key terms
- Fiber-kilometer (fiber-km): one kilometer of a single optical fiber. A 1 km cable with 100 fibers equals 100 fiber-km. A fiber-mile is the same idea in miles (1 fiber-mile ≈ 1.61 fiber-km).
- Preform: the large rod of ultra-pure glass from which optical fiber is drawn.
- Bend-insensitive fiber (BIF): fiber designed to lose little signal when bent tightly, widely used for cabling inside buildings and data centers.
- Spot price: the price for immediate purchase, as opposed to a price agreed in a long-term contract.
- Long-term agreement (LTA): a multi-year contract that guarantees a buyer supply, usually at agreed volumes and prices.
- Optical transceiver: a device that converts electrical signals into light and back, at each end of a fiber link.
- Scale-up / scale-out: connections between chips within one system or rack (scale-up), and between servers across a cluster (scale-out).
Notes on the data
- Fiber demand, shortfall, utilization, inventory, data center share, price multiples and capacity lead times come from TrendForce’s press release of 19 October 2026. TrendForce’s underlying report is not public.
- “Implied supply” (397 million fiber-km), the 31% and 69% shares, the 75% fall in inventory and the Verizon comparisons are NodeZero calculations from those figures. The Verizon comparison divides a six-year contract volume by one year of global demand, to give a sense of scale.
- The Verizon volume is stated by Corning as “80+ million miles”; the 129 million fiber-km conversion is TrendForce’s.
- AI networking revenue figures are TrendForce projections from its press release of 2 November 2026. “Other” (about $33.3 billion in 2027) is the total minus the three named segments.
- Corning figures are from its third-quarter 2026 results, furnished to the SEC on 27 October 2026.
Sources
Research: TrendForce – optical fiber shortage (19 Oct 2026) · TrendForce – AI networking and interconnect market (2 Nov 2026)
Companies: Corning Q3 2026 results (SEC) · Verizon and Corning agreement (8 Sep 2026) · AT&T and Corning agreement (29 Sep 2026) · Marvell and Lumentum announcement (2 Nov 2026)
Every figure in this article links to its source and was checked against that source before publishing. How we check figures