NVIDIA seems to be preparing its own digital highway for artificial intelligence across the entire United States. The company buys unused fiber-optic lines and can get a network with a checkpoint of up to 7.6 petabits per second.
According to experts, NVIDIA acquires long lines of the so-called dark fiber in different parts of the country. In some areas, the company receives up to 100 pairs of optical fibers. Dark is called already laid, but not yet connected equipment that the owner can use to expand the network in the future.
The pair is needed for two-way data transmission, so 100 pairs correspond to 200 individual fibers. Each of them can be divided into dozens of independent channels using a dense spectral seal. DWDM (Dense Wavelength Division Multiplexing) technology transmits data simultaneously at different wavelengths, effectively turning one fiber into a multi-lane. Modern DWDM equipment can easily accommodate from 80 to 96 different data channels in one fiber optic cable.
If you use 96 channels at a speed of 800 Gbps each and connect all 100 pairs, the theoretical bandwidth will reach 7.68 million Gbps, or about 7.6 Pbps. The calculation shows the maximum potential of the network, and not the confirmed speed of the already operating infrastructure.
The first information about the project appeared in June 2026. Then the analytical company Needham reported that NVIDIA is preparing a large telecommunications network costing from 5 to 10 billion dollars. It is planned to build and connect the infrastructure in about three years.
Own optical network will allow NVIDIA to connect data centers into a single computing complex. The company will also be able to hand over channels to operators of cloud platforms, which are still inferior to the largest suppliers in terms of its own network infrastructure.
A more important goal may be to reduce dependence on the largest cloud companies. Now NVIDIA mainly sells them accelerators, but Amazon, Google and Microsoft are developing their own specialized chips for artificial intelligence. Over time, such solutions can reduce the demand for NVIDIA graphics accelerators and strengthen the control of cloud companies over the market.
Own network, investments in independent cloud platforms and ready-made artificial intelligence centers will allow NVIDIA to offer computing power to corporate clients directly. The company will be able to sell not only accelerators, but also a fully prepared infrastructure as a service. The scale of the project was reported by analysts of Wolfe Research.
Such a strategy is also indicated by NVIDIA’s participation in the cloud company Nebius. The accelerator maker recently reported a share of 9.3%, which strengthened assumptions about the intention to create their own distributed platform for artificial intelligence.
According to experts, NVIDIA acquires long lines of the so-called dark fiber in different parts of the country. In some areas, the company receives up to 100 pairs of optical fibers. Dark is called already laid, but not yet connected equipment that the owner can use to expand the network in the future.
The pair is needed for two-way data transmission, so 100 pairs correspond to 200 individual fibers. Each of them can be divided into dozens of independent channels using a dense spectral seal. DWDM (Dense Wavelength Division Multiplexing) technology transmits data simultaneously at different wavelengths, effectively turning one fiber into a multi-lane. Modern DWDM equipment can easily accommodate from 80 to 96 different data channels in one fiber optic cable.
If you use 96 channels at a speed of 800 Gbps each and connect all 100 pairs, the theoretical bandwidth will reach 7.68 million Gbps, or about 7.6 Pbps. The calculation shows the maximum potential of the network, and not the confirmed speed of the already operating infrastructure.
The first information about the project appeared in June 2026. Then the analytical company Needham reported that NVIDIA is preparing a large telecommunications network costing from 5 to 10 billion dollars. It is planned to build and connect the infrastructure in about three years.
Own optical network will allow NVIDIA to connect data centers into a single computing complex. The company will also be able to hand over channels to operators of cloud platforms, which are still inferior to the largest suppliers in terms of its own network infrastructure.
A more important goal may be to reduce dependence on the largest cloud companies. Now NVIDIA mainly sells them accelerators, but Amazon, Google and Microsoft are developing their own specialized chips for artificial intelligence. Over time, such solutions can reduce the demand for NVIDIA graphics accelerators and strengthen the control of cloud companies over the market.
Own network, investments in independent cloud platforms and ready-made artificial intelligence centers will allow NVIDIA to offer computing power to corporate clients directly. The company will be able to sell not only accelerators, but also a fully prepared infrastructure as a service. The scale of the project was reported by analysts of Wolfe Research.
Such a strategy is also indicated by NVIDIA’s participation in the cloud company Nebius. The accelerator maker recently reported a share of 9.3%, which strengthened assumptions about the intention to create their own distributed platform for artificial intelligence.
