Every headline about the new generation of AI campuses leads with megawatts. Stargate Abilene is 1.2 GW. Meta's Hyperion in Richland Parish, Louisiana was expanded in July to 5 GW. xAI's Colossus 2 in Memphis is permitted at roughly 1.2 GW with a $659 million expansion filed on top. Microsoft calls its Fairwater sites in Wisconsin and Atlanta one AI superfactory. Power gets the attention. But inside the fence, what determines whether a gigawatt of GPUs behaves like one computer is glass, and the numbers on glass are larger than most of the industry has absorbed.
What a gigawatt looks like in strands
Start with the rack. An Nvidia GB200 or GB300 NVL72 system draws in the range of 120 to 140 kW, and the Vera Rubin generation arriving in 2026 and 2027 pushes higher. Once you subtract cooling and distribution overhead, a 1 GW campus lands somewhere between 5,000 and 7,000 of these racks, or roughly 350,000 to 500,000 GPUs. Microsoft describes Fairwater Atlanta as hundreds of thousands of Blackwell GPUs under one roof, and Meta's networking lead Omar Baldonado told Hot Interconnects in August that Meta has moved from 24,000 to 32,000 GPU clusters, through a 129,000 GPU system, to multi-building gigawatt deployments. Four of Abilene's eight Phase 1 buildings were already running Oracle Cloud workloads for OpenAI when the last shell topped out this spring.
Now follow one GPU out of the rack. Each accelerator gets its own 800G scale-out port on the Spectrum-X or InfiniBand fabric. An 800G DR8 link is eight parallel single-mode fibers. A non-blocking three-tier Clos fabric means every GPU port is mirrored by a leaf-to-spine link and a spine-to-core link, so each GPU drags roughly 24 fiber terminations through the building before you count storage, front-end, management and the inter-building tier. Run that across 400,000 GPUs and the scale-out fabric alone is on the order of 10 million fiber endfaces. FiberPulse's own estimate, which lines up with vendor claims that AI halls need about four times the fiber of a conventional cloud hall, is that a fully built 1 GW campus consumes between 8 and 15 million connector terminations and somewhere north of 50 million fiber-meters of cable. A single Hyperion-class 5 GW site is a decade of old-style enterprise structured cabling demand delivered in three years.
Scale-across turns the campus into a cable plant
The architectural shift that makes this a campus problem rather than a hall problem is what Nvidia christened scale-across at Hot Chips in 2025. NVLink handles scale-up inside the rack. Spectrum-X or InfiniBand handles scale-out across the hall. Spectrum-XGS Ethernet handles scale-across between buildings and between sites, and Nvidia's own documentation now talks openly about dark fiber, DWDM and ZR coherent pluggables as part of the AI factory bill of materials. Gilad Shainer's keynote in August put Spectrum-X Ethernet Photonics co-packaged optics in production, with Corning, Lumentum, Coherent, SENKO, Foxconn and TSMC named as the ecosystem.
Microsoft's version of the same idea is the most concrete. Its Fairwater buildings are two stories specifically to shorten fiber runs and cut latency, and the company has deployed 120,000 miles of dedicated fiber for an AI WAN linking Atlanta, Wisconsin and sites still under construction, a 25 percent jump in its total fiber mileage in a single year. That is why the campus data center interconnect tier is where the strand counts go vertical. Each building in an eight-building campus like Abilene needs enough inter-building capacity that a training job can be striped across all of them without the fabric becoming the bottleneck. Expect 3,456-fiber and higher ribbon trunks in campus duct banks, multiple diverse routes per building pair, and coherent 800ZR and 1.6T optics on the longer campus hops. Ciena, Nokia and Cisco are all pitching this tier as their growth engine, and hollow-core fiber from Corning, Heraeus and Microsoft's own program is being pulled in for the latency-sensitive links. AWS has said publicly it wants more hollow-core than it can get.
The connector is where density breaks
Inside the hall, the practical fight is at the panel. Classic MPO-12 and MPO-24 connectors built the 400G era, but 800G and the 1.6T links on the Rubin roadmap default to MPO-16 and push patch field density past what a 4RU panel can hold. US Conec's MMC and the broader very small form factor connector class deliver roughly three times the fiber count per panel footprint and are moving from option to spec in new AI halls. Co-packaged optics changes the calculus again: when the switch radix leaves the ASIC as fiber rather than pluggables, the cabling contractor is terminating against the switch face itself, with tighter loss budgets and no second chance to clean a connector that is now part of a $100,000 system. Expect MMC-terminated pre-connectorized trunks, factory-tested and shipped by the hall, to become the default procurement unit rather than field-terminated cable.
The supply chain is already being rebuilt
The manufacturers have made their bets public. Corning signed a $6 billion supply agreement with Meta in January, then announced three new plants in North Carolina and Texas in May with a $500 million investment from Nvidia, promising a 10x increase in US optical connectivity manufacturing capacity and a better than 50 percent increase in US fiber output. Amazon followed in June with its own multibillion-dollar Corning deal and a thousand new North Carolina manufacturing jobs. Prysmian answered in July with a ten-year, up to 5.5 billion euro agreement with Molex for cable deployed inside data centers, backed by a 550 million euro upfront payment, then committed $1.25 billion in August to more than double US fiber and cable capacity across Claremont, Jackson and Lexington. Corning has acknowledged a supply crunch. Analysts expect the new plants to take one to three years to reach volume, which puts the supply relief squarely in 2027 and 2028, the same window in which Abilene Phase 2, Hyperion, Fairwater Wisconsin and the New Mexico, Wisconsin and Pennsylvania Stargate sites are all scheduled to be in fit-out.
What installers should plan for through 2028
For the structured cabling and installation trades, three things follow. First, the unit of work is changing from the rack to the hall. Abilene has run 7,000 trades on site at peak, and white-space fit-out takes four to six months per building, with liquid-cooled GPU pods commissioning more slowly than CPU halls. Fiber crews are on the critical path for every one of those months, and the firms that can staff, certify and test a 10 million endface campus on a compressed schedule will be the ones hyperscalers call back. Second, test and inspection becomes the gating skill. MPO-16 and MMC endfaces at 800G and 1.6T loss budgets do not tolerate the cleaning and inspection discipline of the 100G era, and the certification volume on a single campus will exceed what most regional contractors have done in their history. Third, campus outside plant is now data center work. The duct banks, high-count ribbon trunks, splice vaults and coherent DCI between buildings are the same skills as metro fiber, but they are being bought by Crusoe, Meta, Microsoft and xAI rather than by carriers, on construction timelines rather than telecom timelines.
The gigawatt is the number on the press release. The strand count is the number that decides whether the campus works. Through 2028 the fiber industry's growth is not broadband and it is not BEAD. It is a few dozen fenced sites in Texas, Louisiana, Tennessee, Wisconsin and Georgia, each of which needs more fiber terminations than a mid-sized city, delivered on a schedule set by the GPU shipment calendar. The manufacturers have placed their capital. The installers are next.