Nvidia is building an escape route from hyperscaler control with dark fibre, AI factories and GPU rentals for enterprise punters.
Wolfe Research says Nvidia is “buying up long-haul dark fibre all over the United States, with fibre counts ranging up to 100 pairs”. That is not just a networking side quest.
Needham said in June 2026 that Nvidia was working on a “massive” telecom network project. The bill could run between $5 billion and $10 billion during the next three years.
The move looks like insurance against hyperscalers, ASICs and anyone else fancying a toll booth between Nvidia and AI customers. It gives Jensen’s lot another way to keep the GPU money machine fed.
Dark fibre is fibre-optic cable installed for future use but left dormant until someone lights it up. Buying loads of it gives Nvidia a private backbone without begging hyperscalers for room on theirs.
If Nvidia is chasing up to 100 pairs, that means 200 individual glass strands. Modern kit can squeeze 80 to 96 channels into one strand. If you assume 800Gbps per channel, 96 channels per strand and 100 fibre pairs. That gives a theoretical 7.68 million gigabits a second, or about 7.6 petabits a second.
That sort of capacity could stitch multiple data centres into one vast AI cluster. It could let Nvidia offer neoclouds something closer to hyperscaler-grade pipes.
It might equally be rented to neocloud operators that cannot match the fibre hoards of Microsoft, Amazon, Google and Meta. That would make Nvidia the landlord, plumber and chip dealer in one tidy bundle. The bigger point is ASICs. If hyperscalers keep pushing their own AI silicon, Nvidia needs a way to sell finished AI capacity rather than just expensive chips.
A dark-fibre backbone, tied to neocloud investments, lets Nvidia pitch GPU-as-a-service straight to enterprises. That dodges the future in which hyperscalers become the gatekeepers of AI infrastructure. The timing is interesting because Nvidia has just reported a 9.3 per cent stake in neocloud outfit Nebius.
