Nvidia Spends Billions on Dark Fiber, Expanding AI Battle from Chips to Pipelines

According to reports from Wolfe Research and Needham, NVIDIA is quietly acquiring and leasing dark fiber resources in North America, with investments expected to reach $5-10 billion over the next three years. As of press time, NVIDIA has not officially responded, and the market picture is largely based on analysts' cross-referencing of supply chain signals.

Dark fiber refers to unused optical cable cores reserved by telecom companies. After purchasing, NVIDIA must install DWDM equipment and high-speed optical modules, and overlay its own Spectrum-X networking platform to make it usable. This year, NVIDIA has invested $2 billion each in Coherent and Lumentum, signing multi-year procurement agreements targeting indium phosphide wafers, CPO, and other compute-side optical interconnect components.

The defensive motive is to counter cloud giants. Microsoft, Google, and Amazon are both NVIDIA's largest customers and potential threats—they are developing their own AI chips and building closed loops. By controlling dark fiber, NVIDIA can deliver standardized AI factories that bundle 'GPU + private network' to enterprises, and bypass channel blockades by investing in emerging cloud providers like Nebius (a $2 billion investment in March, with a 9.3% stake by July).

Cloud giants have deep pockets, already control most new transoceanic submarine cables, and are locking in power through nuclear agreements (e.g., Microsoft's deal with Constellation to restart the Three Mile Island nuclear plant). Their ultimate goal is to reduce dependence on NVIDIA.

Among hardware vendors, AMD, Intel, and Broadcom are core participants in the Ultra Ethernet Consortium (UEC), attempting to challenge NVIDIA's InfiniBand dominance with open Ethernet standards. However, Meta and Microsoft are also founding UEC members, so cloud giants are simultaneously building their own networks to compete with NVIDIA while aligning with hardware vendors on open standards—camps are intertwined.

The real winners may be neutral network operators. Lumen has disclosed that its 'Private Connectivity Fabric' business has accumulated nearly $13 billion in contracts. Coherent, Lumentum, Ciena, and other optical module and transmission equipment makers see rising orders. Traditional telecom operators that react slowly face the risk of being 'piped.'

A 100,000-GPU cluster requires a stable self-supplied power source. The gas turbine market is dominated by three players: GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries. GE Vernova has a backlog of $176 billion, with 116 GW of gas equipment capacity; Siemens Energy has an order backlog of approximately €154 billion, with 60 GW of gas turbines signed and 27 GW reserved. Multiple executives say new orders for heavy-duty gas turbines will not be delivered until 2028-2030, with customers typically waiting 3-4 years.

High-voltage transformer lead times have stretched from 24-30 months before 2020 to 3-5 years. Some institutions estimate that over half of U.S. data center capacity under construction or planned may be delayed due to inability to obtain transformers or switchgear. Hitachi Energy's CEO admits orders are 'overwhelming' and has added $1 billion in capacity expansion.

For emergency generators, Virginia had approved over 10,000 generator sets by the end of 2025, with a combined nameplate capacity of about 27 GW—more than half of the total U.S. nuclear fleet. For cooling, when per-rack power exceeds 100 kW, liquid cooling becomes essential. The cooling equipment market is expected to grow from $3 billion in 2025 to $7 billion in 2029, and LiquidStack has already secured a single 300 MW order, with supply unable to meet demand.

AI data centers are no longer a game where 'money in place means construction starts.' They require simultaneous scheduling across four or five highly monopolized heavy industrial supply chains. More and more new clusters are forced to locate in remote areas near power resources, and the more dispersed the sites, the more they rely on long-haul dark fiber connections—this is the macro logic behind the dark fiber scramble.

From competing for advanced process wafers, to laying underground dormant fiber, to scrambling for gas turbines, transformers, generators, and liquid cooling equipment, NVIDIA's cross-industry moves—regardless of how details are confirmed or corrected—point to a clear trend: AI-era compute competition increasingly depends on control over physical resources and heavy industrial capacity, not just leadership in chip design.