AMD, Cisco, and HUMAIN have brought a new artificial intelligence infrastructure online in Saudi Arabia, built on AMD Instinct MI355X GPUs, AMD EPYC processors, and Cisco Silicon One networking. It’s the first operational step of a much larger project: the companies plan to start a next phase of up to 250 MW in 2027 and are still targeting up to 1 GW of AI capacity by 2030.
The AMD, Cisco, and HUMAIN rollout in 20 seconds
- The first AMD Instinct MI355X infrastructure is already serving HUMAIN customers.
- It uses AMD EPYC CPUs and Cisco Silicon One networking with 800G optics.
- HUMAIN will offer GPU capacity as a service for training and inference.
- The next phase targets up to 250 MW starting in 2027.
- The project’s goal remains up to 1 GW by 2030.
The announcement marks a significant shift compared with many of the large AI projects announced in the Middle East over the past few years. In this case, operational infrastructure is already handling customer workloads, though AMD, Cisco, and HUMAIN have not disclosed the currently installed power, the number of MI355X accelerators deployed, or the investment behind this first phase.
The truly large scale remains a forecast.
The three companies expect to start deploying up to 250 MW of AI infrastructure in 2027, with the first capacity from this expansion expected in the second half of the year. After that, the joint venture previously announced by the partners keeps its goal of reaching up to 1 GW by 2030.
So the 250 MW and the gigawatt figures don’t represent currently installed capacity. They’re targets for the project’s next phases.
MI355X, EPYC, and an 800G Network to Deliver GPU as a Service
The infrastructure that just went into production combines three main components.
Accelerated compute comes from AMD Instinct MI355X accelerators, paired with EPYC processors. Interconnection runs on the Cisco N9000 platform built on Silicon One, with 800G optics.
HUMAIN will use this infrastructure to offer GPU as a Service, providing compute capacity to customers who want to train models or run inference without deploying their own systems.
Networking takes on particular importance in facilities like this.
A single server can hold several GPUs, but the largest models need to spread work across many accelerators. As the number of nodes grows, data movement between them can become one of the factors that determines performance.
Cisco says the architecture is designed to deliver low latency, room to grow, and operational resilience. The announcement, however, doesn’t publish independent performance metrics for the installation or training and inference results that would allow comparison with other architectures.
There’s also a technology gap between the current infrastructure and the one planned for the next phase.
The systems already in operation use MI355X, while the expansion of up to 250 MW planned from 2027 will be based on the future AMD Instinct MI400 family, EPYC processors, and the ROCm software stack, along with Cisco’s networking infrastructure.
The jump to MI400 signals that the project is designed to keep incorporating new generations of accelerators as capacity grows.
From the First Systems to a 1 GW Target by 2030
The energy scale announced puts the project in a very different category from a conventional enterprise data center.
The companies are talking about up to 250 MW in the next phase and up to 1 GW by 2030.
To make sense of these figures, it’s worth remembering that megawatts describe a facility’s electrical capacity, not its computing performance. There’s no direct conversion that lets you say how many GPUs a facility will contain based on power alone.
The final number will depend on the accelerators used, server configuration, cooling systems, networking, storage, and the rest of the facility’s power draw.
AMD, Cisco, and HUMAIN also haven’t announced how many MI400 units would correspond to the 250 MW.
They have set an initial timeline. The rollout would start in 2027, with part of that capacity expected to come online during the second half of the year.
The target of up to 1 GW by 2030 belongs to the joint venture previously announced by AMD, Cisco, and HUMAIN and remains subject to execution, demand, infrastructure availability, and other factors. Both AMD and Cisco include specific disclaimers about the forward-looking nature of these projections.
The distinction between operational infrastructure and projected capacity matters especially at a time when the industry is piling up announcements of data centers running several hundred megawatts.
Building that capacity takes more than just having servers.
It requires power supply, grid connections, cooling, buildings, distribution equipment, fiber, and a logistics chain capable of delivering thousands of specialized components.
The announcement says customer demand is supporting the project’s continuation, but it doesn’t provide contracts, utilization rates, or information about the customers currently consuming the capacity.
Saudi Arabia Wants to Become an AI Capacity Provider
HUMAIN occupies a central position in this strategy.
The company belongs to Saudi Arabia’s Public Investment Fund (PIF) and is developing activities spanning data centers, AI infrastructure, cloud platforms, models, and applications.
With AMD and Cisco, it aims to build a platform that can serve both the Saudi market and customers in other countries.
This adds another dimension to the country’s technology strategy.
Saudi Arabia isn’t just looking to consume AI services developed in the United States, Europe, or Asia. It’s trying to build enough infrastructure to offer compute capacity from its own territory.
Data sovereignty is part of the argument.
AMD, Cisco, and HUMAIN describe an architecture where governments, companies, research centers, and developers can decide where their data stays, how they customize their models, and under what mechanisms they’re deployed.
Locally operated infrastructure can appeal to organizations that need to keep certain data sets within a jurisdiction or that want direct control over the platform running their models.
That doesn’t automatically make infrastructure “sovereign.” The concept also depends on software, operations, the supply chain, ownership, jurisdiction, and actual control over the different components.
In this project there are U.S. technologies from AMD and Cisco, while HUMAIN provides local operation and presence.
The companies also lean on the concept of an open platform.
ROCm is an important part of that approach for AMD. The company needs its accelerators to work with a growing number of frameworks and models without depending on a closed software environment.
The strategy lines up with another initiative announced by AMD, Saudi Arabia’s Ministry of Communications and Information Technology, and the Digital Cooperation Organization to train developers in AI tools and ROCm.
The two announcements are different, but they show two sides of the same ambition: building physical capacity while developing the knowledge needed to use it.
AMD Seeks an Alternative to NVIDIA at Data Center Scale Too
For AMD, the Saudi project also provides a large-scale showcase for its Instinct family.
Competition in AI accelerators no longer depends solely on the power of a single GPU.
Modern facilities need servers, high-speed networks, software, storage, libraries, and management systems capable of working together.
The deployment with Cisco shows exactly that approach.
AMD provides CPUs and GPUs, ROCm forms the main acceleration software layer, and Cisco supplies a key part of the network connecting the systems.
If the project eventually reaches the hundreds of megawatts announced, the required capacity will call for coordinating a very large number of components.
It could also give AMD an important reference point in an industry where NVIDIA maintains a strong presence, both through its GPUs and through its interconnect and software technologies.
But there’s still a considerable gap between the deployment that’s already running and the final goal.
The announcement confirms that the MI355X systems are operational, while the 250 MW built on MI400 belongs to a future phase, and the gigawatt is a target for 2030.
That distinction helps put the real scope of the announcement in perspective.
AMD, Cisco, and HUMAIN have already moved from plans to a first production infrastructure. What remains to be seen is how fast they can scale it up over the next four years.
If they stick to the announced timeline, Saudi Arabia would end up with one of the largest AI platforms built on AMD technology, at an electrical scale that would place the project within the international race to concentrate compute capacity for training and inference.
Frequently Asked Questions
What GPUs power HUMAIN’s new infrastructure in Saudi Arabia?
The infrastructure already in production uses AMD Instinct MI355X accelerators together with AMD EPYC processors and Cisco networking based on Silicon One.
Is the announced 250 MW already installed?
No. AMD, Cisco, and HUMAIN plan to start that rollout in 2027 and expect the first capacity from this phase to come online during the second half of the year.
What hardware will the next phase use?
The companies plan to use AMD Instinct MI400-series GPUs, EPYC processors, and ROCm, along with Cisco networking and infrastructure.
Will AMD, Cisco, and HUMAIN have 1 GW of AI capacity by 2030?
That’s the goal announced by the joint venture, not currently available capacity or a guaranteed figure. The companies maintain their plan to deploy up to 1 GW by 2030, subject to the project’s future execution.
