When you’d compare alternatives to NVIDIA

No single company has shaped Saudi Arabia’s AI compute buildout more directly than NVIDIA. The May 2025 agreement between NVIDIA and Humain — the Saudi AI national champion backed by the Public Investment Fund — committed 18,000 GB300 Grace Blackwell Superchips in the immediate tranche, with a stated ambition of 600,000 units delivered over three years. That three-year target, if executed at the stated pace, would make Saudi Arabia one of the largest single concentrations of NVIDIA compute capacity outside the United States. The deal was announced alongside a broader $77 billion AI infrastructure commitment, with NVIDIA’s hardware forming the computational core of Humain’s data center program and the foundation upon which Saudi Arabia’s frontier AI training capability is being built.

NVIDIA’s Saudi Compute Score of 7.8 reflects its dominant position across the dimensions that matter most to the current phase of the buildout. On Silicon Access — weighted at 16% — NVIDIA is the defining entity in the Saudi ecosystem: the H100, H200, and now GB300 Grace Blackwell represent the frontier of AI training and inference hardware, and NVIDIA’s CUDA software ecosystem creates deep switching costs that no competitor has yet overcome at scale in production AI deployments. The company’s Capacity score reflects the sheer volume of GPU-hours flowing into Saudi Arabia’s nascent AI infrastructure through the Humain supply commitment. Its Velocity score is high because the Humain deal is not a letter of intent — it is a structured supply commitment with delivery schedules, co-investment agreements, and joint program structures that translate into operational capacity on defined timelines.

The question of alternatives to NVIDIA arises in several distinct contexts that require different analytical frameworks. Portfolio managers tracking Saudi AI investment want to understand whether AMD or Broadcom can capture a meaningful share of Saudi AI infrastructure spend as the program scales beyond its initial procurement phase. Enterprise technology buyers within Saudi entities — at SDAIA, at the major Saudi banks, at Saudi Telecom — are evaluating whether AMD’s MI300X or Broadcom’s custom ASIC programs can reduce vendor concentration risk in their AI infrastructure portfolios. Policy analysts are asking whether Saudi Arabia’s dependence on a single U.S. chip company for its most critical national AI infrastructure creates geopolitical vulnerability — a question that becomes acute if the export control frameworks governing NVIDIA’s sales to Saudi Arabia are tightened, as has occurred in other markets.

Understanding the alternatives is not primarily about finding a replacement for NVIDIA. In the current phase of the Saudi buildout, there is no replacement that matches NVIDIA’s combination of hardware performance, software ecosystem depth, and supply commitment scale. The analysis is about understanding who occupies the adjacent positions in the competitive landscape, where meaningful differentiation exists, and what conditions — technical, economic, or geopolitical — would cause the competitive balance to shift materially over the 2025-2030 period.

How to read the alternative rankings

The Saudi Compute Score evaluates entities across seven weighted dimensions to reflect their overall strategic importance to Saudi Arabia’s AI infrastructure ambitions. Each dimension is weighted to reflect how central it is to the kingdom’s ability to achieve its Vision 2030 AI goals.

Capacity (18%) is the largest weight because deployable compute determines what Saudi Arabia can actually do with AI today. NVIDIA’s GB300 clusters, with their NVLink interconnect fabric and high-bandwidth memory stacks, represent the highest-density training and inference capability available at commercial scale. The weight on Capacity reflects Saudi Arabia’s current priority: getting large AI clusters operational as fast as possible.

Capital (16%) captures the financial resources and investment commitment that can sustain multi-year AI infrastructure programs. Saudi Arabia’s buildout requires partners that can sustain large supply relationships, absorb the capital intensity of advanced semiconductor manufacturing, and co-invest in Saudi-specific programs. NVIDIA’s market capitalization and cash generation make it a credible long-term partner on this dimension.

Silicon Access (16%) is where the chip competition is most directly evaluated. This dimension measures the ability to design, procure, or manufacture frontier AI silicon — and it is the primary axis on which AMD, Broadcom, and Qualcomm are differentiated from each other and from NVIDIA in the Saudi context. NVIDIA’s GB300 leadership and its TSMC allocation priority give it the strongest Silicon Access score in the ecosystem.

Sovereignty (13%) reflects Saudi Arabia’s goal of owning and controlling its AI stack. For silicon suppliers, this translates into questions about whether the kingdom can negotiate preferential supply agreements, co-development arrangements, technology transfer deals, or eventually domestic assembly or packaging. NVIDIA’s current relationship with Saudi Arabia is a procurement relationship, not a technology transfer relationship, which limits its sovereignty contribution.

Geopolitical Resilience (13%) is particularly important for U.S. chip companies operating in the current export control environment. NVIDIA’s sales to Saudi Arabia have proceeded under approved Commerce Department export licenses, but this framework is subject to modification by executive action. The experience of NVIDIA’s sales to China — where export controls progressively tightened from H100 to H800 to A800 and eventually to significant restrictions — is a relevant precedent for Saudi planners.

Velocity (12%) measures execution speed — how quickly announced supply commitments translate into operational clusters that can run AI workloads. NVIDIA’s Humain delivery schedule is moving quickly by the standards of large infrastructure programs.

Execution (12%) looks at whether promised deliveries materialize on schedule and at specified performance levels. NVIDIA’s track record of delivering on its GPU roadmap — from Hopper to Blackwell to the GB300 — earns a strong Execution score.

When the alternatives become preferable

  • When vendor concentration risk drives procurement policy. Saudi Arabia’s Humain is deploying AI compute at a scale where dependence on a single supplier creates material operational risk. A supply disruption, an export license modification, or a manufacturing yield problem at TSMC could impact delivery schedules for the entire Saudi AI program if NVIDIA is the only hardware source. AMD’s MI300X offers a credible second-source option particularly for inference workloads, where memory bandwidth matters more than raw FLOPS per chip and where ROCm’s software maturity has improved substantially through 2024-2025 production deployments. A diversified GPU fleet reduces exposure to any single supply chain disruption or export control event.

  • When custom silicon economics become compelling at scale. Broadcom’s hyperscale ASIC program has demonstrated through deployments at Google, Meta, and ByteDance that purpose-built AI accelerators can achieve 2-4x efficiency gains over general-purpose GPUs for specific, well-defined workloads like recommendation systems, transformer inference, and video understanding. If Humain or SDAIA deploys AI at sufficient scale in a single inference application — and the 600,000-GPU three-year target suggests that scale is achievable — the unit economics of a custom ASIC program become increasingly compelling even after accounting for 18-24 months of upfront design costs and the engineering resources required to optimize model code for a custom architecture.

  • When edge and on-device AI deployment becomes strategically important. Qualcomm’s Snapdragon and Cloud AI platforms address use cases that NVIDIA’s data center GPU stack is not designed for — AI inference at the edge, in smartphones, in industrial IoT sensors, and in automotive systems. Saudi Arabia’s smart city ambitions across NEOM, Diriyah, and the Red Sea development require massive distributed edge AI deployments where Qualcomm’s silicon is the relevant competitive set and NVIDIA’s data center GPUs are not applicable. The edge layer of Saudi Arabia’s AI infrastructure is as important as the data center layer for the practical delivery of AI services to citizens and businesses across the kingdom.

  • When export control frameworks tighten for Saudi Arabia. NVIDIA’s sales to Saudi Arabia operate under Commerce Department export licenses that can be modified by U.S. government action. A shift in U.S. policy — whether driven by AI proliferation concerns, broader geopolitical recalibration, or pressure from domestic constituencies concerned about GPU exports to Gulf states — could constrain NVIDIA supply to Saudi Arabia more sharply than AMD or Broadcom, whose product mixes span different export control tiers and whose governance structures differ from NVIDIA’s pure AI accelerator focus.

  • When software ecosystem portability improves to a tipping point. CUDA lock-in is NVIDIA’s most durable competitive moat and the single most important reason why AMD has not captured more market share despite competitive MI300X hardware specifications. If AMD’s ROCm ecosystem achieves broad, reliable compatibility with the major AI frameworks and the major model serving stacks — PyTorch, JAX, TensorFlow, vLLM, TensorRT-LLM — the marginal cost of choosing AMD for new workloads drops significantly. Saudi entities building greenfield AI infrastructure have meaningfully less CUDA legacy than U.S. hyperscalers, which means the switching cost barrier is lower for them than for AWS, Google, or Microsoft.

The competitive tier breakdown

AMD (SCS 7.8) ties NVIDIA at the composite score with a competitive hardware profile that makes it the most credible near-term alternative for AI accelerator deployments in Saudi Arabia. The MI300X’s 192GB of HBM3 memory per chip is a genuine competitive differentiator for inference workloads that are constrained by memory capacity — large language model serving, retrieval-augmented generation, and multi-modal inference tasks all benefit from the MI300X’s memory advantage. Humain’s deployment plans have publicly acknowledged AMD hardware alongside NVIDIA as part of a deliberate vendor diversification strategy. AMD’s EPYC CPU leadership means it can offer a coherent compute-plus-accelerator solution for AI servers where the CPU and GPU are both AMD, simplifying supply chain management. AMD’s Silicon Access score benefits from its priority TSMC manufacturing relationship, and its Geopolitical Resilience is marginally better than NVIDIA’s because its product mix spans a broader range of applications beyond the most export-controlled AI training accelerators.

Broadcom (SCS 7.8) represents a fundamentally different competitive angle that is best understood as complementary to GPU deployments in the near term and potentially substitutive in the medium term. Broadcom is not primarily selling Humain a GPU to replace an NVIDIA GPU today — it is offering the hyperscale ASIC model for the medium-term efficiency phase: design a custom accelerator optimized for a specific workload, manufacture it at TSMC, and deploy it in a purpose-built cluster with Broadcom’s Tomahawk and Jericho networking silicon providing the interconnect fabric. Broadcom’s VSP (Virtual Server Platform) AI fabric enables the spine-leaf cluster architectures that large-scale training and inference deployments require regardless of whether the compute nodes use NVIDIA GPUs, AMD GPUs, or custom ASICs. For Saudi Arabia’s ambitions, Broadcom is most relevant as a longer-term partner for workload-optimized silicon programs that can materially reduce the per-inference cost of Saudi AI services at national scale.

Qualcomm (SCS 7.5) scores slightly below the other alternatives on the composite due to its more limited presence in the hyperscale data center AI training segment that anchors Saudi Arabia’s current buildout phase. However, Qualcomm is the unambiguous dominant player for the edge and mobile AI segments that will matter enormously as Saudi Arabia’s smart city programs scale from pilot to national deployment. The Snapdragon 8 Gen series powers the vast majority of smartphone AI inference across the kingdom, including applications in Arabic language processing, mobile banking, and government digital services. Qualcomm’s Cloud AI 100 server-class inference chips are gaining traction in telecom edge deployments, which is directly relevant to Saudi Telecom’s AI infrastructure programs. Qualcomm also brings existing commercial relationships in Saudi Arabia through mobile operator partnerships, which provide distribution and execution advantages that AMD and Broadcom, which primarily sell to OEMs and hyperscalers rather than directly to end operators, do not have.

NVIDIA’s structural position

NVIDIA enters the Saudi compute buildout from a position of structural dominance that its alternatives cannot displace in the current phase and will struggle to meaningfully erode before 2027 even in optimistic scenarios. The CUDA software ecosystem representing decades of optimization work, the GB300 performance leadership for training and inference, the NVLink interconnect fabric for tightly coupled multi-GPU training clusters, and the supply relationship with Humain all point in the same direction: for frontier model training and high-throughput inference at the scale Saudi Arabia is targeting, NVIDIA is the answer the kingdom has already committed to.

The 7.8 SCS reflects genuine limitations alongside that dominance. NVIDIA is a U.S. company operating under export license frameworks that are outside Saudi Arabia’s control. It has no manufacturing capacity in Saudi Arabia and limited ability to provide the sovereign control over hardware technology that the kingdom’s long-term AI independence goals require. Its pricing power, while justified by performance leadership, creates cost concentration risk in Humain’s capex plans at the scale of hundreds of thousands of GPUs. And the CUDA monoculture, while currently an asset for NVIDIA, creates systemic fragility in Saudi Arabia’s AI infrastructure that sophisticated planners will seek to address through gradual diversification.

The strategic question for Saudi Arabia’s AI program through 2030 is whether NVIDIA’s role evolves from a supply relationship into a genuine technology partnership that includes co-development, capability transfer, and eventually some form of sovereign manufacturing or assembly presence in the kingdom — or whether it remains a vendor-customer relationship in which Saudi Arabia’s most critical AI infrastructure is permanently dependent on a single foreign company’s hardware roadmap and U.S. government export license continuity.