Hexagon Data Center: Saudi Arabia’s Sovereign Compute Anchor
In most national AI strategies, government data infrastructure is a footnote — necessary plumbing maintained by a public sector IT bureaucracy that struggles to keep pace with commercial technology development. Saudi Arabia’s Hexagon Data Center represents a deliberate inversion of that pattern. At 480 MW of total capacity and housing the National Data Lake integrating over 430 government systems, Hexagon is not national IT infrastructure by default — it is national AI infrastructure by design, and its operator, SDAIA, intends it to be the sovereign compute anchor for the kingdom’s most sensitive and strategically significant AI workloads.
Hexagon gets its name from its distinctive architectural footprint: viewed from above, the facility resembles a hexagonal shape, an unusual design choice for data center infrastructure that typically optimizes for rectangular floor plans and modular expansion. Located in Riyadh, the Hexagon DC is operated by SDAIA — the Saudi Data and AI Authority, established in 2019 as the governmental body responsible for data policy, AI strategy, and the national data ecosystem.
Scale and Technical Specifications
480 MW of capacity places Hexagon in a category occupied by only a handful of government-owned data centers globally. Most national government data centers in developed economies operate in the tens of megawatts; the largest US federal data centers (DoD, NSA facilities) are not publicly disclosed but are estimated in the range of hundreds of megawatts. If Hexagon’s 480 MW figure is at full buildout rather than current operational load — and early-stage data centers typically run at a fraction of their rated capacity — it nonetheless represents an extraordinary statement of sovereign intent.
The most technically significant deployment within Hexagon is SDAIA’s AI factory: 5,000 NVIDIA Blackwell GPUs. The Blackwell architecture (GB200/GB300) represents NVIDIA’s most advanced AI training and inference hardware, with each GPU delivering substantially higher performance on transformer-based AI workloads than the previous Hopper generation. A cluster of 5,000 Blackwell GPUs, properly configured and networked, represents a training compute resource comparable to large commercial AI research labs. Saudi Arabia has, in other words, equipped its government data center with frontier-grade AI training infrastructure.
This matters for a specific reason: training large AI models on proprietary government data requires that the compute and the data co-reside in a secure, controlled environment. Sending sensitive government training data to a commercial cloud — even a cloud with Saudi data residency — creates risk vectors that a sovereign compute facility eliminates. The Hexagon AI factory is the answer to the question: where does Saudi Arabia train AI models on its most sensitive government data?
The National Data Lake: 430+ Systems in Integration
The more strategically significant (and less frequently discussed) component of Hexagon’s role is its function as the integration point for the Saudi National Data Lake. This is not a conventional enterprise data warehouse — it is an attempt to integrate and expose data from over 430 separate government systems across ministries, agencies, and public enterprises into a unified data architecture.
The scale of this integration challenge should not be underestimated. Government IT systems in most countries — including highly developed ones — are notorious for incompatible formats, proprietary data standards, political silos, and decades of accumulated technical debt. The Saudi government, despite its relative youth as a bureaucratic apparatus, has the same fragmentation problem that afflicts every large governmental IT estate: healthcare data in one ministry’s system, taxation in another, land registry in a third, all using different data models, different APIs, and different governance frameworks.
Building a National Data Lake that genuinely integrates 430-plus systems into a usable, AI-trainable data asset is a multi-year program of enormous technical complexity. SDAIA has been working on this since its establishment in 2019, with Hexagon DC as the physical infrastructure hub. The degree of integration achieved — whether the 430 systems number reflects genuine data interoperability or simply data ingestion without normalization — is a critical variable that outside observers cannot easily assess.
What is clear is the intended use case: Saudi Arabia wants to train AI models on integrated government data to improve public services, detect fraud, optimize resource allocation, and build Arabic-language AI capabilities grounded in domestic data. The Allam 34B model, SDAIA’s Arabic large language model trained on 8 petabytes of data, is the highest-profile output of this effort.
Allam 34B: Sovereign Arabic AI
The Allam 34B model deserves specific attention as the primary output of the Hexagon AI factory to date. Named for the Arabic word for “knowledgeable” (عَلَّام), Allam is a 34 billion parameter Arabic-centric language model trained on 8 petabytes of multilingual data with heavy Arabic weighting. SDAIA has positioned it as the foundation of Saudi Arabia’s sovereign AI capability — a model that is not dependent on US hyperscalers, not subject to US export controls, and not shaped by the training data biases of models developed primarily on English-language internet corpora.
The 34 billion parameter scale puts Allam in a meaningful performance range for enterprise AI applications — not at the frontier of GPT-4-class models (which run 175B+ parameters or use mixture-of-experts architectures that are equivalent to much larger dense models), but sufficient for many government and enterprise use cases where Arabic language understanding is paramount. Arabic NLP historically suffers from underrepresentation in Western AI training datasets; a model trained specifically on large volumes of Arabic text has structural advantages in Arabic language tasks regardless of its parameter count relative to English-centric frontier models.
The 8 PB training dataset is impressive in scale. For context: 8 petabytes of text data, if it were all books, would represent on the order of 4 billion full-length novels. The quality of that dataset — how much is high-quality Arabic text versus web-scraped noise, how much is Modern Standard Arabic versus regional dialects, how much is domain-specific professional content versus informal social media — determines model quality far more than raw data volume. SDAIA has not published detailed training data composition information, which makes independent quality assessment difficult.
The compute infrastructure for training Allam — and for the planned Allam successor models — lives in the Hexagon AI factory. This creates a strategic feedback loop: SDAIA operates the data lake, trains models on that data, deploys models for government applications, generates more structured data from those applications, and feeds that back into the training pipeline. The flywheel logic is sound; the question is execution velocity.
Sovereign Cloud and the PDPL Connection
Hexagon DC serves a second critical function alongside AI training: it provides the sovereign cloud infrastructure on which Saudi government workloads run under the Personal Data Protection Law (PDPL) and the Kingdom of Saudi Arabia’s data residency requirements (KSA-RoD).
Saudi Arabia’s PDPL, enacted in 2021 and enforced from 2023, requires that personal data relating to Saudi nationals be processed and stored within the Kingdom. This is not a uniquely Saudi requirement — the EU’s GDPR, China’s PIPL, and India’s DPDP Act impose similar residency constraints — but in the Saudi context, it is a foundational regulatory driver of the entire cloud infrastructure buildout. Hyperscalers (AWS, Azure, Google Cloud) building Saudi-based cloud regions are responding, in part, to the PDPL residency requirement.
For government data — which is subject to even more stringent sovereignty requirements than commercial personal data — Hexagon DC provides the gold standard answer: compute infrastructure that is physically located in Saudi Arabia, operated by a Saudi government entity, not subject to any foreign jurisdiction’s legal process or data access demands.
This sovereign compute capability is genuinely valuable in the current geopolitical environment. The EU has grappled with the implications of US cloud providers being subject to CLOUD Act requests even for data stored in European data centers. Saudi Arabia has, in building Hexagon, preempted that problem for its most sensitive government workloads.
The Humain Relationship and Ecosystem Position
SDAIA and Humain occupy different but complementary positions in Saudi Arabia’s AI stack. SDAIA focuses on government AI applications, data policy, and sovereign AI models. Humain focuses on commercial-grade AI infrastructure deployment (GPU clusters, cloud services) oriented toward enterprise and international market customers.
Hexagon DC sits in SDAIA’s domain — it is government infrastructure for government purposes. But its presence shapes the ecosystem in important ways. The existence of a sovereign AI training facility with frontier-grade compute establishes a credible floor for Saudi AI capability: even in a worst-case scenario where US-Saudi AI cooperation is disrupted by geopolitical shifts, Saudi Arabia retains domestic compute infrastructure sufficient to continue AI development on sovereign data.
This is strategically analogous to Saudi Aramco’s insistence on maintaining domestic refining capacity: the kingdom does not want to be entirely dependent on foreign infrastructure for its most critical capabilities. Hexagon is the AI analog of domestic refining.
Execution Challenges and the Measurement Problem
The most honest assessment of Hexagon DC acknowledges a significant challenge: measuring actual utilization and output quality is very difficult from outside the system.
The 5,000 Blackwell GPU deployment is a capital commitment that can be verified (NVIDIA’s supply chain, procurement announcements). The 480 MW capacity is a physical infrastructure metric that can be assessed with moderate confidence. But the National Data Lake integration quality, the Allam model’s actual performance on real Saudi government tasks, and the degree to which Hexagon is delivering AI-enabled government services at scale — these are outcome metrics that SDAIA has not been transparent about.
This is not unusual for government AI programs globally. National AI infrastructure initiatives regularly announce ambitious capacity numbers while the utilization and quality metrics remain opaque. The Saudi context adds additional opacity: SDAIA is not subject to freedom of information requests, its AI model benchmarking results have not been independently peer-reviewed, and its data integration progress is self-reported.
For professional observers of the Saudi compute ecosystem, Hexagon DC should be understood as a genuine and significant sovereign compute investment — the most serious government AI data center commitment in the Arab world and one of the largest globally. Whether it is delivering on its potential for Saudi government transformation is a separate question that the available evidence cannot yet answer definitively.
Security Architecture and Sovereign Data Governance
For a government AI data center of Hexagon’s scale and sensitivity, cybersecurity architecture is as important as computational capacity. SDAIA manages data from 430-plus government systems, including tax records, identity databases, healthcare information, immigration records, and critical infrastructure monitoring. A security breach at Hexagon would represent one of the most consequential data incidents in Saudi history — not merely embarrassing but potentially threatening national security.
SDAIA operates under a cybersecurity framework coordinated with the National Cybersecurity Authority (NCA), which was established in 2017 as the principal Saudi government body for cybersecurity policy and operations. The NCA’s Essential Cybersecurity Controls apply to all government entities; for a facility at Hexagon’s sensitivity level, additional classified requirements almost certainly apply.
The AI dimension of security creates additional complexity: AI models trained on government data can, if extracted or reverse-engineered, reveal sensitive information about the training data even when the raw data itself is protected. Model extraction attacks — techniques used to recover training data or reproduce model behavior without authorized access — are an active research area in adversarial AI. Protecting not just the data but the models trained on that data is a new security discipline that government AI facilities globally are still developing operational practices for.
Hexagon’s security architecture also must address insider threats — the risk that authorized users (government employees, contractors, system administrators) extract data or model weights inappropriately. At the scale of 430-plus integrated systems, the population of people with legitimate access to various components of the National Data Lake is large, and insider threat programs must scale accordingly.
Benchmarking Hexagon Against Global Peer Facilities
To contextualize Hexagon’s capabilities and ambitions, it is worth comparing it briefly against peer government AI infrastructure globally. The US National Laboratories (Argonne, Oak Ridge, Lawrence Livermore) operate AI and HPC clusters of comparable scale; Oak Ridge’s Frontier supercomputer, for example, is an exascale system with AI capabilities. These US systems are, however, purpose-built for scientific computing rather than government AI applications — they run climate models, nuclear physics simulations, and materials science calculations, with AI workloads being a more recent addition.
China’s government AI infrastructure, operated through entities like the National Computing Center network, is considerably larger in aggregate but less centralized than Hexagon’s model. China’s approach is to operate distributed high-performance computing centers in multiple cities, connected by high-speed national networks — a different architectural choice than Saudi Arabia’s concentrated 480 MW facility.
The EU’s planned “AI Gigafactories” — national AI computing facilities being built under the European AI Act and EuroHPC initiative — offer the closest structural parallel to Hexagon: government-funded, nationally operated, oriented toward both government applications and national industrial competitiveness. The fact that Saudi Arabia, a single nation, is building comparable infrastructure to what the EU (with 27 member states and a $20 trillion GDP) is building as a continental initiative says something about the scale of Saudi Arabia’s sovereign AI ambition.
Regional and Geopolitical Significance
Hexagon DC’s significance extends beyond Saudi Arabia’s domestic AI ambitions. As the most capable government AI data center in the Arab world, it positions Saudi Arabia as the natural hub for Arab nations seeking sovereign AI alternatives. UAE has its own AI investments (G42, Falcon, the Mohamed Bin Zayed AI University), but the scale of Hexagon’s compute, combined with SDAIA’s Allam model and the National Data Lake, gives Saudi Arabia a credible claim to being the most advanced Arabic-language AI capability in the world.
For the 22 Arab League member states, most of which lack the capital or infrastructure to build comparable sovereign AI facilities, the prospect of accessing Saudi Arabic AI capabilities through a regional arrangement — analogous to how smaller EU nations leverage French or German AI research infrastructure — is worth watching. Whether Humain, SDAIA, or some other vehicle becomes the platform for Gulf and Arab AI cooperation is an open question, but Hexagon DC is the most credible physical anchor for such an arrangement.