Silicon Suppliers

The Hardware at the Center of Everything

Every petaflop of AI compute in Saudi Arabia starts with a chip. Not software, not cloud platforms, not sovereign mandates — silicon. The physical semiconductors that perform the matrix multiplication operations underlying transformer models are the irreducible input to the entire Saudi AI buildout. Without a reliable supply of advanced AI accelerators, Humain’s $77 billion program is a roadmap without a vehicle.

The silicon suppliers sector tracks 12 entities with a combined capex of $4 billion and an average Saudi Compute Score of 6.22. That average is the second lowest of any sector, and the reason is important: silicon suppliers face the most constrained access path of any category in the ecosystem. They operate under export control frameworks administered by the US Department of Commerce, compete in an allocation-constrained market where demand significantly exceeds supply, and must navigate a geopolitical environment in which the US-Saudi bilateral relationship directly determines what can be sold and to whom.

Understanding this sector requires understanding not just who makes the chips, but why Saudi Arabia can buy from some suppliers and not others, how different silicon architectures map to different AI workloads, and what the current distribution of silicon supply reveals about the geopolitical structure of Saudi AI.

NVIDIA: The Dominant Supplier

GB300 and the Technical Case for Dominance

NVIDIA holds a position in AI compute silicon that has no precise historical analog. Its H100 and now Blackwell architecture (GB200, GB300) GPUs are the de facto standard for large language model training and high-performance inference. The technical reasons for this dominance are genuine: NVIDIA’s CUDA software ecosystem, developed over fifteen years of investment, means that virtually every AI training framework, every LLM architecture, and every enterprise AI workflow is optimized for NVIDIA hardware. Switching to alternative silicon is not merely a hardware swap — it requires rewriting and reoptimizing the software stack, which in practice takes months to years.

The GB300 NVL72 is NVIDIA’s flagship AI compute unit for the 2025-2026 deployment cycle. A single NVL72 rack contains 72 Blackwell GPUs connected via NVLink with 1.4 terabytes per second of inter-GPU bandwidth, capable of treating the entire rack as a single compute fabric. This architectural coherence is critical for large model training, where the ability to distribute computation across hundreds of GPUs with minimal communication overhead determines training efficiency. A 100,000-GPU deployment using NVL72 racks can train a trillion-parameter model in weeks rather than months.

Humain’s framework agreement with NVIDIA for 18,000 GB300 NVL72 systems in Phase 1 — representing approximately 1.3 million GPU equivalents at rack level — is the largest single AI hardware commitment by any entity outside the major US hyperscalers. The framework extends to 600,000 advanced GPUs over three years, which at current GB300 pricing would represent a capital commitment of $40-60 billion. This is not a purchase order for servers — it is a strategic partnership that positions NVIDIA as the primary silicon supplier for Saudi Arabia’s sovereign AI program.

Export Licensing Under the AI Diffusion Framework

NVIDIA’s GB300 sales to Saudi Arabia do not happen on the open market. They are governed by the US Bureau of Industry and Security (BIS) AI Diffusion framework, which classifies countries into tiers based on their geopolitical relationship with the United States and the assessed risk of advanced AI technology misuse.

Saudi Arabia occupies Tier 2 under the AI Diffusion framework — a classification that allows exports of advanced AI chips subject to licensing requirements and end-use verification. Tier 1 countries (close US allies including UK, Japan, South Korea, Netherlands) face minimal restrictions. Tier 3 countries (China, Russia, and others) face near-total prohibition on advanced AI chip exports. Saudi Arabia’s Tier 2 status reflects the US government’s view that the kingdom is a strategic partner in AI development but that some oversight of end-use is warranted.

In practice, the BIS licensing framework for Saudi AI hardware operates through a government-to-government consultation process. Humain’s NVIDIA procurement framework is understood to have involved US government review and approval, not just a commercial transaction between a chip company and a foreign buyer. MCIT’s role as the regulatory counterparty on the Saudi side — managing the international technology agreements that underpin US-Saudi AI cooperation — is central to maintaining the flow of licensed hardware.

The AI Diffusion framework creates structural advantages for companies like NVIDIA, which have the government relations capability to navigate complex export licensing, and structural disadvantages for smaller suppliers that lack the bureaucratic capacity to manage the licensing process for large government-to-government transactions.

The Second-Source Strategy

Saudi Arabia’s AI hardware strategy is explicitly designed to avoid dependence on a single silicon supplier. Humain and SDAIA have both publicly articulated the principle that sovereign compute requires silicon diversification. This reflects two separate concerns: supply concentration risk (what happens if NVIDIA constrains allocation or increases prices) and geopolitical risk (what happens if US-Saudi relations deteriorate and BIS tightens export licenses).

The second-source strategy is manifested in active commitments to multiple alternative silicon providers.

AMD: The JV Partner

Advanced Micro Devices occupies the most strategically significant alternative silicon position. AMD, Cisco, and Humain are co-investors in a joint venture to build 1 gigawatt of AI compute capacity over five years, with AMD supplying MI300 and next-generation Instinct GPU series hardware. The JV structure is meaningful — AMD is not merely a supplier but an equity participant in Saudi AI infrastructure, which aligns incentives in ways that a pure procurement relationship does not.

AMD’s Instinct MI300X is technically capable of most workloads that run on NVIDIA H100, and for memory-bandwidth-intensive workloads like large model inference, AMD’s 192 GB of HBM3 memory per MI300X (versus 80 GB for H100) provides a meaningful advantage. The software ecosystem gap — NVIDIA CUDA versus AMD ROCm — remains the primary obstacle to AMD adoption, but AMD has made substantial investments in ROCm compatibility and major framework support has improved significantly.

The AMD-Humain JV is a credible bet on AMD’s ability to close the software ecosystem gap over the five-year buildout horizon, while giving Saudi Arabia meaningful supply diversification from day one.

Groq: The Inference Specialist

Groq represents a different silicon philosophy entirely. Where NVIDIA and AMD build general-purpose AI accelerators optimized for flexibility across training and inference workloads, Groq’s Language Processing Unit (LPU) architecture is specifically designed for inference — running already-trained models at the highest possible speed with deterministic latency.

The Groq-Aramco Digital partnership, valued at $1.5 billion, is building what Groq describes as the world’s largest inference facility outside the United States. Aramco Digital will deploy Groq LPU clusters to serve inference workloads for Aramco’s internal AI applications and potentially for external customers via a commercial inference service.

The LPU’s technical advantage for inference is genuine: Groq achieves token generation speeds that are 3-10x faster than GPU-based inference for comparable model sizes, with consistent latency that does not degrade under load. For applications where response latency matters — conversational AI, real-time document processing, customer service automation — Groq’s architecture has clear advantages over GPU inference.

The $1.5 billion scale of the Groq-Aramco deal reflects Saudi Arabia’s serious engagement with specialized silicon, not just general-purpose GPU procurement. It also reflects Aramco Digital’s ambition to operate commercial AI infrastructure services, not just internal IT.

Qualcomm: The Edge and Mobile Layer

Qualcomm’s Saudi presence addresses a different point in the AI compute stack: edge inference and mobile AI. Qualcomm’s Snapdragon and Cloud AI chips are optimized for power-efficient inference at the edge — in devices, vehicles, industrial equipment, and small-footprint servers — rather than in large data center clusters.

For Saudi AI applications in smart city infrastructure (NEOM’s pervasive sensing environment), industrial automation (Aramco’s digital oilfield programs), and consumer devices, edge AI is as important as cloud AI. Qualcomm’s AI inference capabilities, embedded in its Snapdragon platform and its Cloud AI 100 inference accelerators, address workloads that do not belong in a 120 kW NVL72 rack.

Intel and SambaNova: The Enterprise Alternatives

Intel’s Gaudi AI accelerator series and SambaNova’s DataScale platform represent enterprise-oriented silicon alternatives that are particularly relevant for Saudi enterprises deploying AI in mixed workloads — combining traditional enterprise computing with AI inference.

Intel’s Gaudi 3 has achieved competitive performance on standard benchmarks for LLM inference and is backed by Intel’s extensive enterprise distribution network. For Saudi enterprises that already run Intel-based server infrastructure, Gaudi 3 offers a lower-integration-cost path to AI inference than deploying NVIDIA or AMD clusters.

SambaNova’s DataScale platform takes a different approach — a dataflow architecture optimized for maximum throughput on specific model architectures, with a software stack designed for enterprise deployment rather than research computing. SambaNova’s smaller scale makes it less relevant for hyperscale AI infrastructure but meaningful for enterprise deployments in financial services, healthcare, and government applications.

Broadcom: The Custom ASIC Enabler

Broadcom occupies a unique position in the silicon supply chain as the primary enabler of custom AI ASIC development. Through its networking silicon (Jericho, Tomahawk) and its ASIC design services (formerly Avago), Broadcom is a critical supplier to companies building custom AI chips rather than buying standard GPU or LPU products.

Hyperscalers — Google’s TPU, Amazon’s Trainium/Inferentia, Microsoft’s Maia — are all custom ASIC programs that use Broadcom networking silicon for inter-chip communication. As Saudi Arabia’s AI infrastructure matures and Humain or ALAT potentially develops custom AI silicon, Broadcom’s ASIC design and networking capability becomes relevant to the kingdom’s long-term hardware sovereignty objective.

Why Chinese Silicon Is Absent

The most conspicuous absence in Saudi Arabia’s silicon supply chain is Chinese semiconductor technology. Despite Saudi Arabia’s active economic engagement with China — Aramco investments in Chinese refinery capacity, BRI connectivity projects, the China-Arab States Cooperation Forum — Huawei’s Ascend 910C AI accelerator and other Chinese AI silicon are not part of the Saudi AI buildout.

The reason is structurally simple: Saudi Arabia’s Tier 2 status under the US AI Diffusion framework is predicated on end-use controls that prevent the diversion of US-licensed AI technology to Chinese entities. Deploying Huawei silicon in the same compute infrastructure that holds US-origin Blackwell GPUs would create compliance risk that could jeopardize Saudi Arabia’s access to the entire US AI supply chain. The US-Saudi AI relationship is the more valuable one, and Saudi Arabia has made the rational choice to preserve it.

This does not mean Saudi-China AI relations are absent — Saudi entities including Saudi Aramco and SABIC have deep commercial relationships with Chinese companies. But at the silicon layer, where US export controls operate, Chinese hardware is excluded from the sovereign compute buildout by the logic of the bilateral US-Saudi relationship.

Silicon Types and Workload Mapping

Training vs. Inference: A Fundamental Divide

The AI compute market is not uniform. Training large language models requires different silicon characteristics than running those models at inference. Training demands maximum memory bandwidth and high-precision floating-point computation — characteristics where NVIDIA’s NVL72 racks excel. Inference demands maximum throughput per watt, consistent low latency, and support for the quantized (reduced-precision) arithmetic that speeds inference without materially degrading model quality — characteristics where Groq’s LPU and optimized inference GPUs (NVIDIA H200, AMD MI300X) perform differently.

Saudi Arabia’s silicon strategy implicitly recognizes this divide. Humain’s NVIDIA GB300 deployment is training-optimized infrastructure for developing large models. The Groq-Aramco facility is inference-optimized infrastructure for deploying those models at scale. The AMD JV provides additional capacity that can flex between training and inference depending on workload requirements.

Custom ASIC: The Horizon Option

The third silicon category — custom ASICs designed for specific workloads — is not yet present in Saudi Arabia’s AI buildout but represents the horizon toward which ALAT is working. Custom ASICs can achieve 5-10x better performance per watt for specific workloads compared to general-purpose GPUs, at the cost of flexibility and enormous upfront design investment.

The question of whether Saudi Arabia will develop custom AI ASICs is open. ALAT’s mandate includes semiconductor R&D, and the combination of TSMC’s (and Samsung’s) foundry capacity and accessible EDA (electronic design automation) tools makes custom chip development more accessible than it was a decade ago. The barrier is not purely technical — it is ecosystem. A custom ASIC is only useful if the software stack supports it, which requires a critical mass of AI developers optimizing for that architecture.

What the Silicon Supply Chain Reveals

The structure of Saudi Arabia’s silicon procurement reveals the architecture of its geopolitical AI position more clearly than any policy document. Saudi Arabia has:

• Made NVIDIA the primary silicon partner and accepted US export licensing oversight as the price of access to the world’s best AI accelerators
• Diversified across AMD, Groq, Qualcomm, Intel, and SambaNova to reduce both supply risk and geopolitical dependency
• Excluded Chinese silicon to preserve US-Saudi AI cooperation
• Pursued a joint venture structure with AMD that gives Saudi entities equity stakes rather than pure procurement relationships
• Committed to ALAT as the long-term answer to hardware dependency, while operating pragmatically within the current US-supply-dependent environment in the near term

This combination — maximum engagement with the US silicon ecosystem while building toward long-term sovereign hardware capability — is the coherent strategy of a government that understands its current dependencies and is working systematically to reduce them. For companies in the silicon supply chain, Saudi Arabia represents not just a customer but a strategic partner that is making bets on which silicon architectures will define the next decade of AI infrastructure. Being on the right side of that bet matters.