AI Affairs, home

Thursday 22 October 2026

Corporate

Humain expands Nvidia partnership with plan for up to 84MW of AI capacity

The proposed rollout spans Blackwell and Vera Rubin systems. A planned 50MW Vera Rubin allocation is for an unnamed US customer, subject to final agreements.

Two NVIDIA RTX 2080 graphics cards stand side by side against black.
Photo: Nana Dua via Pexels

Saudi Arabia’s Humain expanded its Nvidia partnership on 22 October with a plan for up to 84MW of AI computing capacity, building on a May 2025 plan for up to 500MW over five years, The National reported. Announced at Nvidia’s GTC conference in Berlin, the expansion covers approximately 27,000 graphics processing units, or GPUs, for Humain Cloud and dedicated customer infrastructure.

Key points

  • The first phase is 14MW of Blackwell systems, equivalent to about 6,000 GPUs, with deployment starting in the fourth quarter of 2026.
  • A further 70MW of Vera Rubin infrastructure, representing about 21,000 GPUs, is planned from mid-2027.
  • An unnamed US client is earmarked for 50MW of the Vera Rubin capacity. The companies say those projects remain subject to final agreements and deployment plans.
  • A separate preliminary framework at NEOM envisages Nebius taking up to 50MW of data centre capacity.

Humain starts with 14MW of Blackwell systems

Humain is owned by Saudi Arabia’s Public Investment Fund. Launched in May 2025, it is chaired by Crown Prince Mohammed bin Salman. Its original Nvidia partnership envisaged an initial deployment of 18,000 Blackwell processors as part of the plan for up to 500MW of AI computing capacity over five years.

The first phase of the expansion comprises 14MW of Nvidia Blackwell HGX B300 systems, equivalent to about 6,000 GPUs. Deployment is due to start in the fourth quarter of 2026. That capacity is intended for Humain Cloud customers, including AI companies, government entities and businesses.

Konsulteer, an AI industry publication, described the first-stage technology as Blackwell Ultra. The National described it as Blackwell HGX B300.

The later phase is planned around Nvidia’s Vera Rubin infrastructure, the successor to Blackwell. It accounts for 70MW and about 21,000 GPUs, with deployment planned from mid-2027. The two phases put different chip generations into Humain’s planned capacity rather than making the full 84MW available in the first rollout.

US customer earmarked for 50MW of Vera Rubin

Humain plans to begin deploying 50MW of the Vera Rubin capacity for an unidentified US client in June 2027. The capacity is intended for advanced AI workloads. It is part of the 70MW Vera Rubin phase, rather than an addition to the announced 84MW programme.

Humain expects to begin deploying the remaining 20MW for Humain Cloud in July or August 2027. The National reported that both companies have made the Vera Rubin work conditional on final contracts and agreed rollout plans. Those conditions apply to the phase containing both the US client’s planned allocation and the Humain Cloud capacity.

Humain chief executive Tareq Amin said the company was moving to introduce successive generations of Nvidia computing in Saudi Arabia. The initial Blackwell phase is directed towards Humain Cloud customers, while the later Vera Rubin phase includes capacity dedicated to a single US client.

Nebius plans separate 50MW commitment at Oxagon

Humain also has a separate customer arrangement under development at Oxagon, in NEOM. Humain, DataVolt and Nebius have signed a preliminary framework agreement under which Nebius, an AI cloud company, plans to take up to 50MW at the AI giga campus, Data Centre Magazine reported. The proposed offtake covers half of the initial 100MW under construction.

DataVolt is the developer, owner and operator of the site’s initial phase. The Oxagon development targets 360MW for its complete first phase. Its proposed Nebius commitment concerns data centre capacity at that campus, while the unnamed US client’s planned 50MW concerns Vera Rubin computing capacity under the Nvidia expansion.

Humain’s data centre construction plans call for about 3GW by 2030, then more than 6GW by 2034. The Oxagon framework and the Nvidia rollout give it separate routes to place planned infrastructure with customers: one for capacity at a campus under construction, the other for computing systems scheduled across two chip generations.

Nebius expects to launch operations at Oxagon in phases starting in the fourth quarter of 2027.

Sources

Topics: Chips, Data centres, Enterprise adoption, Sovereign AI