NVIDIA is quietly building a massive telecommunications network across the United States, with total bandwidth potentially reaching an incredible 7.6 Petabits per second. This strategic maneuver allows the company to offer turnkey AI factories and GPU-as-a-service directly to enterprises, effectively bypassing the need to rely on external cloud providers.
Acquiring the physical infrastructure
Needham disclosed in June 2026 that NVIDIA was working on a “massive” telecom project that could cost between $5 billion and $10 billion over the next three years. Wolfe Research has since confirmed that the company is buying up long-haul dark fiber all over the United States, with fiber counts ranging up to 100 pairs. Industry experts believe they are aiming to resolve bottlenecks within its GPU clusters.
To understand the scale, consider that a single fiber strand can only send an optical signal in one direction, so you typically need a pair for effective two-way communication. If NVIDIA is acquiring 100 pairs of dark fiber—fiber optic cables meant for future capacity that currently lie dormant—it equates to 200 individual strands of glass.
Calculating the true capacity requires looking at DWDM (Dense Wavelength Division Multiplexing) equipment, which splits a single beam of light into different colors or wavelengths, where each color acts as a separate data channel. A state-of-the-art DWDM can easily pack between 80 and 96 distinct data channels within a single strand. Assuming a data transmission speed of 800 Gbps, a DWDM that splits a single fiber shard into 96 channels, and up to 100 fiber pairs, the total capacity of NVIDIA’s network scales to 7.68 million Gigabits per second or 7.6 Petabits per second.
While the specific plans for this infrastructure are unconfirmed, there are several theories regarding its use. It could connect multiple data centers to form a gigantic cluster, or NVIDIA could lease this network to neoclouds to even the playing field against hyperscalers. However, the most significant strategic implication appears to be a defense against the rise of ASICs.
Selling only GPUs renders NVIDIA vulnerable to the ascendancy of ASICs currently being championed by hyperscalers, who can eventually become the gatekeepers of AI solutions. By leveraging this dark fiber network and its existing investments in neoclouds, the company can offer its own GPU-as-a-service directly to enterprise customers, evading a potential chokehold by these larger competitors. The filing notes that NVIDIA has reported a 9.3% stake in the neocloud Nebius Group, which supports this integrated strategy.
Related: TSMC to generate its own power under new Taiwan law
Strategic independence
Building a network of this magnitude represents a significant shift in the tech industry. It suggests a move toward vertical integration where control over both the hardware and the connectivity becomes a priority. The sheer volume of fiber required implies a long-term commitment to infrastructure that goes far beyond standard data center operations.
NVIDIA’s strategy is focused on providing turnkey AI solutions to enterprises, and this network is a key component of that plan. By offering GPU-as-a-service, the company can help enterprises implement AI solutions without having to invest in their own hardware and infrastructure.
The network’s capacity is impressive, with the potential to reach 7.6 Petabits per second. This is a substantial resource that will allow the company to support a large number of enterprise customers.
It is a single step.
NVIDIA’s move into the telecommunications industry is a significant development, and it will be interesting to see how the company’s plans unfold. The filing notes that NVIDIA has reported a stake in the neocloud Nebius Group, which supports this integrated strategy. As the company continues to build out its network, it will be important to watch for any developments related to its GPU products.