Where the GPUs actually are

Headline GPU procurement numbers — Saudi’s 600K NVIDIA pipeline, the UAE Stargate buildout, regional cluster announcements — paper over a more granular reality: GPU clusters at scale require concurrent power, networking, cooling and operational staffing. A cluster announcement is not a cluster; an installed-and-running cluster is. The cluster-level ranking surfaces which deployments have actually solved all four constraints versus which remain announcement-stage commitments.

The MENA region’s GPU cluster topology converges on three primary nodes. Saudi Arabia (Riyadh, Dammam, NEOM-Oxagon) is accumulating cluster count fastest because Humain’s procurement pipeline runs through five operational sites and additional facilities under construction. The UAE (Abu Dhabi G42 facilities, Stargate-aligned sites) leads on operational maturity for clusters already running at hyperscaler-grade specs. Bahrain, Oman, Egypt and Qatar host smaller specialty clusters; their inclusion in the ranking reflects emerging rather than dominant positions. Israel operates AI infrastructure at meaningful scale but is treated separately for geopolitical reasons in this MENA-framed ranking.

Reading the top entries

Humain occupies the top of the cluster ranking because of cumulative commitments rather than current installed base. The November 2025 BIS framework cleared up to 35,000 GB300 Blackwell systems for Humain, with the broader 600K-unit pipeline targeting deployment over a three-year horizon. At full deployment, Humain’s GPU footprint becomes the largest sovereign GPU pool outside the US and China — comparable in scale to a single hyperscaler region. The phasing matters: the first 18,000 systems anchor the initial Riyadh and Dammam cluster build; subsequent tranches expand the footprint as facilities come online and as NVIDIA’s TSMC CoWoS packaging and HBM3e supply allow. By 2027, Humain’s installed base is projected to exceed any non-US, non-China sovereign GPU pool by a wide margin.

SDAIA holds the second slot via the sovereign AI factory partnership with NVIDIA — up to 5,000 Blackwell-class GPUs for government workloads, hosted within the Hexagon government data center. The cluster size is smaller than Humain’s commercial footprint but the strategic weight is higher because SDAIA-controlled GPUs underpin the sovereign LLM (Allam), citizen-services AI deployments and the National Data Lake’s analytical workload across 430+ government systems. The SDAIA cluster is the operational substrate for the Year of AI 2026 ministry-level deployment milestones.

The remainder of the top tier (where data is available) reflects a mix of: Aramco’s Groq LPU cluster supporting the $1.5B Aramco Digital inference partnership; Center3’s hyperscaler-shared cluster capacity in its Riyadh facility; the xAI joint-venture site at the 500 MW Humain campus (Grok inference and training); the AWS Riyadh region’s Trainium and Inferentia-based capacity alongside its NVIDIA hosting; the Google Cloud Dammam hub’s TPU-and-NVIDIA mix; and emerging G42-aligned clusters in Abu Dhabi.

The four-constraint test

A cluster on the ranking has solved — or is credibly solving — four operational constraints that often gate announcement-to-operation transitions.

Power: Sovereign-scale GPU clusters consume 30-50 MW at the small end and 200-400 MW at the hyperscale end. Saudi Arabia’s structural electricity-cost advantage (industrial tariffs in the $20-50/MWh range versus $80-150/MWh in US data center hubs) makes the operating economics work, but power availability still requires coordinated grid connection, often involving Saudi Electricity Company, ACWA Power renewable PPAs and increasingly direct Aramco gas supply for industrial-zone facilities. NEOM-Oxagon facilities are solving the power problem through dedicated renewable buildout (solar + wind + battery storage at scale).

Networking: GPU clusters at training scale require non-blocking fabric across thousands of accelerators. NVIDIA Quantum InfiniBand and Spectrum-X Ethernet are the primary fabrics; Cisco and Arista provide the broader networking backbone. The Cisco-AMD-Humain JV explicitly addresses the networking layer for Humain’s facilities. UAE clusters typically standardize on Microsoft’s Azure-aligned fabric within G42 deployments. Subsea cable landing capacity feeds the inter-region connectivity through the SmartHub Saudi Arabia program and the broader Center3 / stc / Mobily fiber backbone.

Cooling: GB300 systems and equivalent Blackwell-tier accelerators require liquid cooling at densities most legacy data centers cannot support. Saudi Arabia’s climate adds thermal load that drives premium cooling capex versus US benchmarks. The DataVolt-NEOM 1.5 GW facility is being designed cooling-first as a net-zero AI factory; Humain campuses are specifying liquid cooling at the rack level; Hexagon’s design accommodates the next-generation density requirements. Cooling is increasingly a differentiator on the ranking because clusters specified to legacy air-cooling standards cannot host the latest-generation accelerators without expensive retrofit.

Operational staffing: Sovereign-scale clusters require trained operators numbering in the dozens to hundreds depending on scale. Saudi Arabia is closing the operational-talent gap through the SDAIA SAMAI program (workforce-scale AI literacy), KAUST and KFUPM graduate pipelines, and aggressive expat hiring through Humain Ventures’ portfolio companies. Talent is not a binding constraint at the cluster-count level today but becomes a binding constraint as the 600K-GPU footprint scales toward 2030 and the cumulative operator headcount required exceeds what current pipelines produce.

The Saudi-UAE structural difference

Saudi clusters are predominantly housed in PIF-owned or Humain-controlled infrastructure with vertical integration from chips through models. UAE clusters are predominantly housed in G42-operated or hyperscaler-shared infrastructure with horizontal integration through Microsoft and Stargate partners. The architectural difference shapes how each cluster ecosystem evolves. Saudi clusters can be reconfigured top-down through Humain executive decision; UAE clusters are reshuffled through commercial negotiations with Microsoft and OpenAI/SoftBank. Both architectures have legitimate strategic merits and neither dominates on a like-for-like comparison.

Through 2026, Saudi cluster count grows faster (more facilities, more sites, faster announcement cadence). Through 2026, UAE cluster operational maturity remains higher (more tenured deployments, more proven workloads at scale). The convergence point is roughly mid-2027 when Humain’s largest facilities reach full operational stability and the announcement-stage deficit becomes less of an analytical advantage for the UAE side of the ranking.

What the ranking misses

The ranking captures publicly disclosed clusters and ignores undisclosed deployments inside Aramco, SABIC, Saudi banks and government agencies. The aggregated undisclosed capacity is non-trivial and probably exceeds 50,000 accelerators in cumulative deployment across enterprise and government workloads. Some of this undisclosed capacity is older-generation NVIDIA (V100, A100, H100/H200) running production workloads rather than frontier training; the latest-generation deployments are concentrated in the disclosed Humain, SDAIA and hyperscaler regions.

The ranking also undercounts inference-only clusters relative to training clusters. A 5,000-GPU inference cluster supporting Aramco’s industrial AI workload is operationally consequential but produces less press signal than a 5,000-GPU training cluster. The economic split between training and inference is shifting toward inference dominance through the decade as deployed model scale grows; the ranking will increasingly need to differentiate inference-cluster ranking from training-cluster ranking. Qualcomm’s AI200/AI250 deployments and Groq’s LPU cluster footprint are both inference-first architectures that the conventional GPU-cluster framing under-weights.

Forward-looking forcing functions

Three milestones reshape the cluster ranking through 2027. The Hexagon government data center reaches full operational status during 2026 — at 480 MW, it is the world’s largest sovereign DC and its operational ramp materially shifts the SDAIA cluster ranking. The Humain Riyadh and Dammam campus tranches go fully operational on the published schedule, scaling the disclosed Humain footprint substantially. The DataVolt-NEOM 1.5 GW net-zero AI factory delivers its first operational tranche, putting the largest renewable-powered AI facility in the world on the ranking.

Outside Saudi Arabia, the broader UAE Stargate buildout produces additional cluster announcements that will compete with Saudi entries on the regional ranking. Egypt’s emerging cluster positioning at smaller scale provides a fourth-place comparator. Bahrain’s AWS Middle East region remains a structurally important but smaller node. Tencent Cloud’s $150M MENA commitment is small relative to the AWS / Google / Microsoft footprints but signals that Chinese hyperscalers retain a regional role even where they’re excluded from approved AI facilities under the November 2025 framework.

How to read the cluster sizes

A common analytical mistake is to read GPU counts as if they were directly comparable. Different accelerator generations have wildly different effective compute output: a single GB300 system delivers materially more useful compute than an equivalent count of H100 systems on most training workloads, and the training/inference split changes which architectures win. A 10,000-GB300 cluster is roughly equivalent to 25,000-30,000 H100 systems on dense training workloads, so cross-generation comparisons require normalization to FLOPs or to specific benchmarks. The ranking surfaces accelerator generation alongside count where data is available; readers should weight the generation as much as the count when comparing across entries.

The other normalization issue is networking topology. Two clusters with identical accelerator counts can deliver materially different throughput on real workloads if one is fully non-blocking InfiniBand and the other is oversubscribed Ethernet. Humain’s specifications target full non-blocking topology at the largest cluster sizes; UAE Stargate-aligned clusters typically follow Microsoft’s Azure-aligned architecture which is also full-spec at scale. Smaller regional clusters sometimes accept oversubscription to manage capex.

The vendor-mix dimension

Saudi cluster ranking entries differ in their accelerator vendor mix, which has structural implications for ecosystem lock-in and software portability. Humain’s primary clusters are NVIDIA-anchored (GB300 plus successors), with secondary AMD Instinct deployments via the Cisco-AMD-Humain JV providing accelerator diversification. Aramco Digital’s Groq LPU cluster is an inference-first specialty deployment outside the NVIDIA stack. SambaNova’s RDU architecture supports SDAIA-aligned training workloads at smaller scale. Qualcomm AI200/AI250 deployments are emerging for inference workloads requiring specific power-envelope optimization.

The vendor diversification is deliberate rather than accidental. Humain’s multi-vendor stack provides protection against any single supplier’s production constraints (NVIDIA’s TSMC CoWoS bottleneck, AMD’s HBM3e supply, Qualcomm’s manufacturing capacity). It also positions Humain to capture pricing leverage by maintaining credible alternatives. The architectural cost is software fragmentation: workloads optimized for CUDA do not transparently port to ROCm, RDU or LPU architectures, which means Humain’s software engineering team carries cross-platform optimization complexity that single-vendor hyperscalers avoid.

UAE clusters within the G42-Microsoft architecture are more standardized on NVIDIA via Azure, simplifying the software stack but increasing single-supplier dependence. The Stargate-aligned positions add OpenAI-co-developed accelerator considerations through the broader $500B Stargate architecture. Both architectures have legitimate trade-offs.

The next ranking inflection

The ranking is updated quarterly to reflect operational milestones. The next inflection point is the H1 2026 reporting period when the first Hexagon ramp data should be public, the initial Humain GB300 deliveries should be confirmed by NVIDIA as shipped, and the SDAIA sovereign factory’s installed base will surface in Year of AI 2026 disclosures. Watch the trajectory of cluster operational counts through that window — it is the cleanest test of whether the November 2025 framework delivers on its execution promise.

The cooling-architecture differentiator

Cluster ranking entries diverge meaningfully on cooling architecture. Direct-to-chip liquid cooling is becoming the standard for GB300-class systems; rear-door heat exchangers handle the previous generation; chilled-water immersion cooling is emerging at the densest deployments. Saudi clusters built since 2024 are predominantly direct-to-chip-liquid-ready. Legacy facilities retrofitted to host AI workloads sometimes accept thermal compromises that limit their effective accelerator density.

DataVolt-NEOM is being designed for two-phase immersion cooling at meaningful scale, which would put it among the densest deployments globally and would set a precedent for hyperscale immersion adoption that few other operators have committed to at gigawatt scale. The cooling architecture choice cascades into water consumption, secondary loop chemistry, maintenance complexity and uptime characteristics — material operational differences that the headline accelerator count understates.

The MENA ranking versus the global ranking

Read in the global frame, the MENA cluster ranking matters because the region is the third-largest sovereign GPU concentration outside the US and China by 2027 on current trajectory. The US holds the dominant position by a wide margin (hyperscaler clusters at AWS, Microsoft, Google, plus frontier-lab clusters at OpenAI, Anthropic, xAI, Meta and Google DeepMind). China holds the second position with Huawei Ascend and Alibaba/Tencent infrastructure at scale despite export-control headwinds. MENA — combining Saudi Humain plus UAE Stargate and G42 — is plausibly the third concentration by combined installed base, ahead of any individual European or other Asian sovereign program.

The implication for cluster-ranking analysis is that MENA is a structurally important global node rather than a regional curiosity. Cluster decisions made in Riyadh, Abu Dhabi and NEOM-Oxagon during 2026-2028 affect the global AI supply chain because they absorb material fractions of NVIDIA’s TSMC CoWoS allocation and HBM3e supply. A delay at the Humain Dammam campus has implications for the global GPU delivery schedule because shifted allocation changes which other regions receive priority. Reading the MENA ranking as if it were detached from the global picture understates its structural weight.

For the data center capacity side of the picture, see the Saudi MW capacity ranking. For the silicon-side perspective on what’s powering each cluster, see the Saudi silicon supplier ranking. For the broader regional comparison, see the MENA data center investments ranking.

For deeper reading: