Humain Campus — Riyadh: Saudi Arabia’s Sovereign AI Headquarters
The HUMAIN Riyadh campus is the anchor deployment of Saudi Arabia’s AI infrastructure buildout and the most strategically significant data center under construction in the Gulf region. At 200 megawatts for Phase 1, the Riyadh campus is the first physical instantiation of HUMAIN’s mandate to build a national AI factory ecosystem — the compute substrate on which Saudi Arabia’s AI programs will run. The choice of Riyadh as the primary campus location, the MIS design-build contract that is its construction mechanism, and the 18,000 GB300 Grace Blackwell AI supercomputers that constitute its initial GPU loadout collectively define what Saudi Arabia means when it commits to becoming an AI superpower.
Understanding the Riyadh campus requires understanding what it is not. It is not a colocation facility — HUMAIN is not renting space to tenants in the conventional sense. It is not a hyperscaler’s private cloud region — though Google, Microsoft, and AWS have adjacent relationships within the HUMAIN ecosystem. The Riyadh campus is Saudi Arabia’s national AI factory: government-affiliated compute infrastructure owned and operated through the Public Investment Fund’s HUMAIN vehicle, designed to serve national AI programs, attract strategic partnerships, and establish Saudi Arabia as a genuine AI compute destination rather than merely a consumer of foreign AI services.
Phase 1 Hardware: 18,000 GB300 Grace Blackwell Supercomputers
The Phase 1 hardware specification is the result of the NVIDIA-Humain partnership announced in May 2025. HUMAIN will deploy 18,000 NVIDIA GB300 Grace Blackwell AI supercomputers as the first installation at the Riyadh and Dammam campuses, with the Riyadh campus absorbing a significant portion of the initial allocation.
The GB300 is NVIDIA’s most capable AI training system available at the time of deployment. Each GB300 Grace Blackwell supercomputer unit combines 72 Blackwell B200 GPU chips — configured as 36 B200 dual-chip modules — with NVIDIA’s Grace Arm-based CPU via NVLink-C2C at 900 GB/s bidirectional bandwidth. The NVLink-C2C interconnect eliminates the PCIe bottleneck that limits conventional add-in GPU cards: the CPU and GPU share a unified memory address space with NUMA-aware access, enabling AI workloads that move data between host and device memory to run at significantly higher throughput than PCIe-attached GPU architectures.
At cluster scale, the 18,000 GB300 units are connected via NVIDIA InfiniBand — as specified in the NVIDIA partnership agreement — providing the non-blocking GPU-to-GPU communication fabric that distributed training requires. The aggregate compute capacity of 18,000 GB300 units is measured in exaflops for AI-precision operations: enough to train frontier-class Arabic language models, run large-scale multimodal AI programs, and serve inference at the scale HUMAIN’s commercial ambitions require. The memory capacity — 192GB HBM3e per GB300 unit, 3.46 petabytes across the 18,000-unit Phase 1 cluster — is sufficient to hold models with hundreds of billions of parameters entirely in GPU memory across the cluster.
The power requirements for 18,000 GB300 units are substantial. Each GB300 supercomputer carries a maximum thermal design power of approximately 10 kW, putting the 18,000-unit Phase 1 cluster’s power draw at approximately 180 MW at full utilization — nearly the entirety of the 200 MW Phase 1 campus power budget. This near-total GPU utilization of available power is intentional: the Riyadh campus is an AI compute facility, not a mixed general-purpose data center with spare power capacity for non-GPU workloads. The remaining power budget covers networking, storage, cooling overhead, and facility management systems.
MIS Contract: SR1.88B Design-Build Partnership
The design and construction of the Riyadh campus is being executed under a landmark contract awarded to Al Moammar Information Systems (MIS), a Riyadh-based IT infrastructure and systems integration company listed on the Saudi Exchange (Tadawul). The contract’s disclosed scale is striking: the value exceeds 155% of MIS’s 2024 revenues of SR1.21 billion, implying a contract value of approximately SR1.88 billion — roughly $500 million at current exchange rates. A single contract worth 155% of the contractor’s annual revenue is an extraordinary concentration of work that underscores both the urgency of HUMAIN’s construction timeline and Saudi Arabia’s commitment to directing major infrastructure work to domestic companies.
The MIS design-build contract encompasses civil engineering and construction management, mechanical and electrical systems integration, technology infrastructure deployment, and project delivery against aggressive timelines. MIS brings the local contractor relationships, Saudi construction regulatory experience, and existing technical relationships with Saudi government and quasi-government clients that make it the appropriate primary contractor for a facility of this strategic sensitivity.
International technology vendors — NVIDIA for compute hardware, Cisco for networking, cooling systems specialists for the thermal infrastructure — supply the specialized technology components under separate supply agreements. MIS coordinates their installation within the Saudi-built facility, serving as the systems integrator that ensures all components arrive, fit together, and are commissioned on schedule. This is the most complex role in data center construction: not any individual technology component is the hard part, but the coordination of dozens of parallel workstreams in the right sequence.
Vision 2030 local content requirements apply to HUMAIN’s construction program. As a PIF subsidiary, HUMAIN is subject to the Iktva (In-Kingdom Total Value Added) local content framework that Saudi Arabia applies to government-linked procurement programs. The MIS contract, worth nearly SR2 billion for a Saudi-headquartered company, is a direct expression of this local content obligation — keeping a significant portion of the construction program’s economic value within the Saudi economy rather than flowing to international construction management firms.
The +50 MW/Quarter Expansion Cadence
HUMAIN’s expansion approach at the Riyadh campus follows a phased cadence that is structured around power capacity additions. The +50 megawatts per quarter expansion rate, if sustained from the initial 200 MW Phase 1, would bring the Riyadh campus to 400 MW within four quarters and toward the 500 MW potential campus capacity within six to eight quarters of Phase 1 completion.
This cadence has important infrastructure planning implications. Power infrastructure — the utility transformers, 380kV/132kV substations, switchgear, and distribution systems that connect grid power to the data hall — has construction and delivery lead times of 12-24 months for large components. At a +50 MW/quarter addition rate, HUMAIN must be commissioning power infrastructure for quarters five and six while still completing the Phase 1 installation. The phased approach is therefore not as simple as sequential construction blocks; it requires parallel infrastructure provisioning across multiple expansion phases simultaneously.
GPU hardware availability is a second dimension of the expansion cadence. NVIDIA’s Blackwell supply chain — dominated by TSMC’s CoWoS advanced packaging capacity and HBM3e supply from SK Hynix and Samsung — is constrained globally, and HUMAIN’s expansion rate is partly limited by available hardware allocation. HUMAIN’s long-term partnership with NVIDIA and its status as one of NVIDIA’s largest global customers provides allocation priority, but absolute supply constraints remain relevant. The +50 MW/quarter cadence should be understood as HUMAIN’s target rate, subject to power infrastructure completion and GPU hardware availability.
The financial model benefits from the phased cadence. Rather than deploying the full 500 MW capital program upfront, phased deployment allows HUMAIN to commission revenue-generating capacity in stages, using early phases’ commercial operations to partially fund subsequent phases. This cash flow profile is more favorable than a single large upfront investment, particularly for a new commercial entity establishing market relationships.
Sovereign AI Location Logic: Why Riyadh Specifically
The Riyadh location for Saudi Arabia’s primary sovereign AI campus reflects optimization criteria that are not purely technical — if technical optimization were the only consideration, the Eastern Province would be preferred for its lower power costs and Aramco proximity. Riyadh is the primary campus because the Saudi national AI program is, at its core, a government program, and government programs concentrate around government capitals.
SDAIA’s headquarters, the Ministry of Communications and Information Technology, the National Data Bank infrastructure, and the primary government agency IT operations are all Riyadh-based. The AI programs that will be the most demanding and strategically significant workloads on HUMAIN’s infrastructure — Allam Arabic LLM training, National Data Bank AI pipelines, government service AI applications — all originate from Riyadh-based programs. Physical proximity between the AI factory and the program teams running AI development reduces operational friction and enables the informal collaboration between engineers and infrastructure operators that accelerates AI program execution.
Network latency is not the primary argument — fiber latency between Riyadh and Dammam is measured in single-digit milliseconds, not meaningful for most AI workloads. The institutional proximity argument is more important: Saudi government agencies operating classified AI workloads on HUMAIN infrastructure want to be able to physically inspect, audit, and access that infrastructure under controlled conditions. A sovereign AI factory in the national capital satisfies this requirement in a way that a remote or internationally located facility does not.
PDPL data residency compliance is most straightforwardly demonstrated when data that originates from Saudi government systems in Riyadh is processed in Riyadh and never leaves the metropolitan area. This residency chain — from government database to AI training pipeline to trained model, all within Riyadh — creates the simplest possible PDPL compliance demonstration for the National Data Bank’s AI programs. Track progress across all HUMAIN campus deployments at Infrastructure.
Operational Milestones and Capacity Ramp
The operational milestone timeline for the Humain Riyadh campus reflects the phased deployment model described above. Phase 1 first power-on — the moment the initial GB300 racks receive power and begin running test workloads — represents the transition from construction to operational status. This milestone is typically followed by a commissioning period of 4-8 weeks during which each rack, network port, and storage system is validated and certified before entering production use.
The first production AI workload on the Riyadh campus is a significant milestone beyond its operational significance. It marks the beginning of Saudi Arabia’s national AI factory era — the moment when HUMAIN transitions from infrastructure developer to AI compute operator. SDAIA’s Allam training programs, HUMAIN’s commercial AI services, and the first sovereign Arabic AI inference serving to Saudi government applications all depend on this transition. The timeline from first power-on to full Phase 1 operational capacity at 200 MW is typically 3-6 months for a facility of this scale, reflecting the parallel commissioning of power, cooling, compute, and network systems across 200 MW of infrastructure.
Liquid Cooling Infrastructure: Engineering for Saudi Climate
The thermal engineering requirements for a 200 MW AI factory in Riyadh are substantially more demanding than for an equivalent facility in a temperate climate. Riyadh experiences summer ambient temperatures exceeding 45°C regularly, and the design dry-bulb temperature for mechanical systems sizing is typically 48-50°C. At these outdoor air temperatures, conventional air-cooled data center cooling — which relies on rejecting heat to outdoor air — requires large chillers operating with poor efficiency coefficients of performance (COP).
HUMAIN’s Riyadh campus design addresses this challenge through direct liquid cooling (DLC) as the primary thermal management approach for the GB300 Grace Blackwell AI supercomputers. Each GB300 unit’s GPU and CPU chips are directly cooled by liquid coolant circulating through integral cold plates, with heat transferred to a facility-level cooling water loop that can be rejected to outdoors at significantly higher temperatures than data hall supply air. The facility cooling water, running at 40-45°C supply temperature, can be rejected to outdoor dry coolers or cooling towers even in Riyadh’s summer ambient — where chilled air systems would struggle to operate efficiently.
The power efficiency implications of liquid cooling for HUMAIN’s operational economics are substantial. A conventionally air-cooled 200 MW IT load facility in Riyadh’s climate would require 60-80 MW of additional cooling power, yielding a PUE of 1.3-1.4 at best in summer months. Liquid cooling can achieve PUE of 1.15-1.2 year-round, saving 30-40 MW of cooling power relative to air cooling — equivalent to enough power for approximately 7,500-10,000 additional GB300 AI supercomputers. The liquid cooling investment pays back in operational savings and enables higher compute density within the same 200 MW power budget.
Security Architecture and Sovereign AI Protection
The Riyadh campus, as Saudi Arabia’s primary sovereign AI facility, requires physical and logical security measures that exceed the standards of commercial colocation facilities. The compute running in this facility processes national AI programs, sovereign data, and government AI workloads that are of the highest strategic sensitivity in the Kingdom.
Physical security encompasses multi-layer perimeter control with biometric and card access at each zone boundary, 24/7 on-site security personnel, comprehensive CCTV coverage with appropriate retention, and anti-tamper protection for compute hardware that prevents physical access to GPUs, storage, and networking without detection. The facility design separates different security zones — government-classified workload zones, HUMAIN commercial AI zones, and facilities management zones — with physical barriers and access controls that enforce the security classification boundaries.
Network security on HUMAIN’s Cisco infrastructure provides the logical layer of this protection. Cisco Hypershield’s distributed enforcement architecture, combined with CiscoXDR monitoring, creates a fabric-level security posture that catches insider threats, network anomalies, and unauthorized data movement before they can result in exfiltration of training data, model weights, or inference results. The combination of physical and logical security creates the security architecture that Saudi Arabia’s sovereign AI programs require for HUMAIN infrastructure to be eligible for the most sensitive government AI workloads.