NVIDIA in the Saudi AI Compute Landscape
No vendor shapes Saudi Arabia’s AI compute buildout more profoundly than NVIDIA. The $77 billion sovereign AI program anchored by the Public Investment Fund treats NVIDIA’s GB300 Grace Blackwell architecture as the de facto standard for frontier model training and large-scale inference, a position NVIDIA has cemented through a combination of hardware dominance, software lock-in, and high-level diplomatic engagement that few semiconductor companies can replicate.
The scale of NVIDIA’s Saudi commitment is extraordinary even by the standards of the current AI infrastructure boom. Humain, the PIF-backed sovereign AI company created specifically to anchor Saudi Arabia’s compute ambitions, signed a landmark agreement covering 18,000 GB300 NVL72 systems in Phase 1 alone, with a trajectory toward 500,000 to 600,000 GB300 systems over a three-year horizon. To put that in context, a single GB300 NVL72 rack-scale system contains 72 Blackwell GPUs interconnected via fifth-generation NVLink, delivering roughly 1.4 exaflops of AI compute at FP4 precision per rack. Scaling that across hundreds of thousands of systems represents a compute concentration with no precedent outside the United States.
NVIDIA CEO Jensen Huang attended the Riyadh technology summit on May 13, 2025, signaling the depth of executive commitment on both sides. His presence alongside Crown Prince Mohammed bin Salman and a roster of US tech executives underscored that the Humain-NVIDIA partnership is not simply a hardware transaction but a strategic alignment between Saudi sovereign capital and Silicon Valley’s most powerful AI infrastructure company. The Saudi government views NVIDIA technology as foundational to Vision 2030’s AI pillar, and NVIDIA views Saudi Arabia as its most strategically important sovereign customer outside the United States.
Beyond Humain, NVIDIA’s Saudi footprint extends to the Saudi Data and AI Authority, which secured 5,000 Blackwell units for national AI programs including Arabic language model development, and to stc and Center3’s joint venture infrastructure allocation. The aggregate compute target across these deployments points toward a 500-megawatt AI campus footprint, making Saudi Arabia one of the world’s largest concentrations of NVIDIA-powered AI infrastructure.
Hardware Architecture and Specifications
The GB300 Grace Blackwell Superchip represents NVIDIA’s most aggressive architectural leap since the original Volta transformer era, and its specifications explain why Saudi Arabia’s planners selected it as the backbone of a multi-decade AI investment. Understanding the GB300 architecture is essential to understanding the scale and ambition of the Saudi compute program.
At the chip level, the GB300 integrates two Blackwell GPU dies with a Grace CPU via NVLink-C2C, NVIDIA’s chip-to-chip interconnect that delivers 900 gigabytes per second of bidirectional bandwidth between processor and memory. This tight integration eliminates the PCIe bottleneck that limited earlier GPU-CPU communication and allows the CPU and GPU to share a unified memory address space across 288 gigabytes of LPDDR5X system memory on the Grace side and 192 gigabytes of HBM3e on the Blackwell GPU die. Total memory accessible to a single GB300 Superchip reaches 480 gigabytes in the NVL72 configuration, a figure that matters enormously for fitting the largest frontier models—those with hundreds of billions of parameters—without model parallelism overheads that degrade effective throughput.
The Blackwell GPU die itself delivers 20 petaflops of FP4 tensor core performance, a precision format that NVIDIA introduced with Blackwell specifically to maximize inference throughput for quantized large language models. At BF16, the more common training precision, each Blackwell die delivers approximately 5 petaflops, and in an NVL72 rack the aggregate BF16 throughput reaches 360 petaflops—enough to perform trillions of floating-point operations every second across a dense transformer attention layer.
The NVLink 5.0 interconnect fabric binding the 72 GPUs within an NVL72 system delivers 1.8 terabytes per second of all-to-all bandwidth. This is the architectural choice that differentiates NVIDIA’s rack-scale approach from disaggregated GPU clusters: by treating the entire 72-GPU rack as a single logical GPU with coherent high-bandwidth interconnect, NVIDIA enables model parallelism strategies that keep activations and gradients in-fabric rather than traversing the slower InfiniBand host network. For the very largest models—GPT-4 scale and above—this distinction between in-rack NVLink bandwidth and inter-rack InfiniBand bandwidth determines whether training is bottlenecked by compute or communication.
Between racks, NVIDIA’s NDR InfiniBand at 400 gigabits per second per port provides the scale-out fabric. Humain’s campus designs call for HDR/NDR InfiniBand at full-fat-tree topology, meaning every rack communicates with every other rack at line rate without oversubscription. At the 500-megawatt campus scale, this implies an InfiniBand switching infrastructure comprising thousands of Quantum-2 or Quantum-3 switches, all supplied by NVIDIA’s Mellanox networking division.
The GB300’s power envelope runs approximately 1,000 watts per Superchip in sustained training workloads, and the NVL72 rack consumes roughly 120 kilowatts under full load. Scaling to hundreds of thousands of systems means the Saudi compute campuses require purpose-built power and cooling infrastructure, driving parallel investment in liquid cooling systems and grid-scale power procurement from Saudi Aramco’s power subsidiaries and NEOM’s renewable energy projects.
Saudi Deployment and Partnerships
The primary deployment channel for NVIDIA hardware in Saudi Arabia is Humain, the sovereign AI company established by the Public Investment Fund in 2024 with an initial capital commitment of $40 billion and a mandate to build Saudi Arabia into one of the world’s top five AI nations by 2030. Humain operates as a commercial entity with sovereign backing, meaning it can execute procurement at the scale of a nation-state while retaining the operational flexibility of a technology company.
Phase 1 of the Humain-NVIDIA partnership covers the deployment of 18,000 GB300 NVL72 systems, representing roughly 1.3 million Blackwell GPUs and approximately 26 exaflops of AI compute at BF16 precision. The systems are planned for deployment across multiple data center campuses in the Riyadh metropolitan area and along the Red Sea coast, with initial capacity expected to come online in late 2025 and Phase 1 completion targeted for 2026. The three-year roadmap to 500,000 to 600,000 systems represents capital expenditure on the order of $50 billion at current NVIDIA pricing, making it the single largest AI hardware procurement contract in history.
SDAIA’s 5,000-unit Blackwell deployment is operationally distinct from the Humain deal and serves a different set of use cases. SDAIA, as the national regulator and operator of Saudi Arabia’s government AI programs, uses its Blackwell allocation primarily for Arabic language AI development—specifically for training iterations and fine-tuning runs on the Allam Arabic large language model family—and for the National Data Management Office’s AI inference services. SDAIA’s deployment is physically hosted within Saudi Arabia’s government cloud infrastructure and operates under the PDPL, Saudi Arabia’s personal data protection law, which requires certain categories of government data to remain within Saudi borders.
The stc and Center3 JV allocation represents the commercial cloud layer of Saudi Arabia’s NVIDIA deployment. stc Group, Saudi Arabia’s dominant telecommunications operator, and Center3, its data center subsidiary, are building hyperscale AI cloud infrastructure that will offer NVIDIA GPU compute as a service to Saudi enterprises, government agencies, and regional customers across the Gulf Cooperation Council. This commercial layer is essential to the economic sustainability of Saudi Arabia’s compute buildout: sovereign investment in hardware only generates returns if that hardware runs at high utilization, and cloud services to paying enterprise customers provide the demand signal that justifies the infrastructure investment.
Competitive Position vs Other Silicon Vendors
NVIDIA’s position in Saudi Arabia is not simply dominant—it is structurally entrenched in ways that make displacement by AMD, Intel, or specialized vendors extremely difficult over any near-term horizon.
The most durable source of NVIDIA’s competitive advantage in Saudi is not hardware performance but software: the CUDA ecosystem. CUDA, NVIDIA’s parallel computing platform and programming model, has been the foundation of AI research and production deployment for fifteen years. Every major AI framework—PyTorch, JAX, TensorFlow—runs natively and most efficiently on CUDA. Every major LLM training codebase, from Megatron-LM to GPT-NeoX, is written against CUDA primitives. When Saudi Arabia’s AI teams hire engineers trained at Google, Meta, OpenAI, or top universities, those engineers arrive knowing CUDA. Switching to AMD’s ROCm or Intel’s oneAPI requires retraining engineering teams, porting production code, and accepting a period of reduced productivity. At the scale of a national AI program, those transition costs are prohibitive.
AMD’s MI300X is the most credible alternative for inference workloads, offering 192 gigabytes of HBM3 per chip with higher memory bandwidth than the H100 generation. For inference-only deployments where memory capacity determines how large a model can be served efficiently, AMD presents a genuine cost-competitive option. But for training—which is where the largest NVIDIA allocations in Saudi are directed—CUDA’s ecosystem advantage and NVLink’s intra-rack bandwidth keep NVIDIA in a position AMD cannot currently challenge at scale.
Against Groq’s LPU architecture, NVIDIA competes differently: NVIDIA cannot match Groq’s deterministic latency for single-stream inference, but Groq cannot match NVIDIA’s flexibility to run training workloads, multimodal models, or custom architectures. Saudi deployments increasingly separate training clusters (NVIDIA-dominated) from inference serving clusters (where Groq, AMD, and specialized vendors compete), a segmentation that actually benefits NVIDIA by cementing its role in the highest-capex part of the stack.
Intel Gaudi 3 and SambaNova’s RDU are in evaluation or niche deployment stages in Saudi Arabia and do not represent near-term competitive threats to NVIDIA’s volume. The 500-megawatt campus scale Humain is building is effectively an NVIDIA monoculture, deliberately designed around NVLink, NVSwitch, and InfiniBand as an integrated system.
Export Controls and Geopolitical Considerations
Saudi Arabia’s access to NVIDIA’s most advanced hardware, specifically the GB300 Grace Blackwell, was not guaranteed until a diplomatic breakthrough at the highest levels of US-Saudi relations. Under the Biden administration’s AI Diffusion rule, Saudi Arabia was classified as a Tier-2 country—meaning it could receive advanced AI chips only under a government-to-government framework or under strict volume and end-use controls that made sovereign-scale procurement practically impossible.
The Trump administration’s revision of the AI Diffusion framework, culminated at the May 13, 2025 Riyadh summit, created a pathway for Saudi Arabia to receive GB300 hardware at sovereign scale through a bilateral AI agreement that includes US oversight provisions, data center inspection rights, and commitments from Saudi Arabia not to re-export the hardware to Tier-3 destinations. The diplomatic channel that enabled this agreement—the Trump-MBS relationship developed through the Abraham Accords era and renewed in 2025—is a material factor in NVIDIA’s Saudi business, because the hardware access itself is contingent on that political relationship remaining stable.
The export control framework creates both a moat and a risk for NVIDIA. The moat: because access to GB300 hardware in Saudi Arabia required US government approval, it is not easily replicable by Chinese competitors like Huawei (whose Ascend 910C chip is export-restricted from Chinese companies operating in US-aligned markets). The risk: any deterioration in US-Saudi relations could trigger export restriction reviews that interrupt hardware deliveries or require renegotiation of the bilateral AI agreement.
Outlook: Saudi Silicon Roadmap
NVIDIA’s three-to-five-year roadmap in Saudi Arabia is more clearly defined than any other vendor’s: the 600,000-system trajectory represents a capital deployment schedule, not a wishlist. Phase 1 hardware deliveries through 2026 will consume the majority of NVIDIA’s GB300 production allocation for the Gulf region. Phase 2 deployments in 2027 and 2028 will likely incorporate NVIDIA’s next-generation Rubin architecture, currently scheduled for sampling in 2026 with production ramp in 2027.
The 500-megawatt campus target implies that Saudi Arabia will build AI compute infrastructure at a scale comparable to the largest hyperscaler campuses in the United States, but concentrated in a single sovereign entity rather than distributed across commercial cloud providers. This creates a unique dynamic: Humain operates at hyperscaler scale but with sovereign priorities, meaning it will allocate compute to national AI programs, Arabic language model development, and Vision 2030 economic diversification goals that pure commercial operators would not prioritize.
For NVIDIA, Saudi Arabia represents both a revenue milestone—a single sovereign customer accounting for tens of billions in hardware revenue over three years—and a strategic template for the sovereign AI market. The Humain deal is being watched by UAE’s G42, Qatar’s QIA, and other Gulf sovereign funds as a model for national AI infrastructure. NVIDIA’s ability to deliver at scale in Saudi Arabia will directly influence whether it captures the broader Gulf sovereign AI market or whether alternative vendors find an opening in the second tier of procurement.