AMD, Cisco and HUMAIN bring MI355X GPU cluster live in Saudi Arabia
AMD, Cisco and HUMAIN have confirmed that GPU-accelerated AI infrastructure in Saudi Arabia is now serving production workloads, marking the first live milestone in a joint venture that has publicly targeted 1 gigawatt of deployed AI capacity by 2030. The announcement, timed to coincide with the companies' broader strategic briefing on 31 August, signals a tangible shift from deal-signing to operational delivery in the Gulf's most ambitious sovereign AI programme.
The live deployment runs on AMD Instinct MI355X GPUs paired with AMD EPYC CPUs, interconnected over Cisco's N9000 Series switching platform using Cisco Silicon One ASICs and 800G optics. HUMAIN, a company owned by Saudi Arabia's Public Investment Fund, is offering the capacity as GPU-as-a-service, targeting use cases from large-scale model training to real-time inferencing for customers in the Kingdom and the wider region.
Next phase and long-range targets
The companies have outlined a second construction phase set to begin in 2027, with capacity expected to start coming online in the second half of that year. The next tranche will add up to 250 MW of infrastructure and will transition the GPU layer to AMD's forthcoming Instinct MI400 Series, alongside AMD's ROCm open-source software stack and Cisco networking. Combined with the live systems already operating, the 2027 phase will materially expand Saudi Arabia's available AI compute before the decade closes.
The joint venture remains on track, the partners say, to reach the headline 1 GW figure by 2030. AMD chief executive Lisa Su said the platform is "serving customers today and will scale significantly in the coming years," while HUMAIN chief executive Tareq Amin noted that "demand for sovereign, high-performance AI infrastructure continues to grow." The release does not name specific customers, disclose revenue commitments, or provide utilisation figures for the live cluster.
Market and regulatory context
Saudi Arabia's AI infrastructure push sits within a broader regional competition for sovereign compute capacity. The UAE, Qatar and Egypt have each announced large-scale data-centre investments, with hyperscalers including Microsoft, Google and Amazon committing capital alongside local sovereign vehicles. The AMD-Cisco-HUMAIN structure is notable for its explicit rejection of proprietary closed ecosystems: ROCm provides an open alternative to NVIDIA's CUDA software stack, and Cisco's networking layer is positioned as interoperable rather than locked to a single silicon vendor.
The sovereign framing is commercially significant. Gulf states are increasingly requiring that AI workloads involving government data or national-security applications reside on locally operated infrastructure, a requirement that mirrors European digital-sovereignty debates under the EU Cloud Rulebook and France's SecNumCloud scheme. HUMAIN's full-stack model, spanning hardware, networking and Arabic-language large language models, is designed to meet those requirements without routing sensitive data through US hyperscaler regions.
Export-control risk remains the most material external variable for the venture. US Bureau of Industry and Security rules have repeatedly tightened controls on advanced GPU exports to the Gulf, and the MI355X sits in a performance tier that warrants monitoring. AMD's own cautionary language in the release explicitly cites "export regulations" and "national-security-based regulations" among the factors that could cause results to differ from guidance. The 2027 phase, which depends on MI400 Series GPUs not yet shipping at volume, carries additional supply-chain and licensing uncertainty. Investors and enterprise buyers watching the HUMAIN platform will treat the 2030 targets as ambitious but directionally credible, contingent on the regulatory environment remaining stable.