Building an 8MW AI Data Center in Kazakhstan's SEZ Alatau

Selecting the right location for an AI-focused data center requires balancing multiple factors: power availability and cost, climate conditions for cooling efficiency, regulatory environment, network connectivity, and proximity to talent. Almaty, Kazakhstan's largest city, scores exceptionally well across all of these dimensions, and SEZ Alatau — the technology-focused special economic zone on the city's southern outskirts — adds fiscal incentives that make the overall proposition compelling. Almaty sits at an elevation of approximately 800 meters above sea level at the base of the Tian Shan mountain range. The continental climate delivers average temperatures of -5C in January and 24C in July, with low humidity year-round. These conditions are excellent for data center cooling — ambient temperatures are below the free-cooling threshold for most of the year, reducing mechanical cooling requirements significantly. The city is also seismically active, which informed our decision to use base-isolated foundations and seismically rated rack systems. SEZ Alatau is a 400-hectare technology park established by the Kazakh government. It provides pre-built industrial infrastructure including roads, power substations, water supply, and fiber-optic connectivity. The SEZ's tax incentives — 0% CIT, 0% property tax, 0% VAT — reduce operating costs by an estimated 15-20% compared to a facility outside the zone. The SEZ also offers streamlined permitting and customs clearance for imported technology equipment, which is critical for timely delivery of GPU hardware.

Our data center is designed to meet the Tier III standard defined by the Uptime Institute and the TIA-942-B telecommunications infrastructure standard. Tier III requires concurrent maintainability — meaning every component of the power and cooling infrastructure can be serviced or replaced without interrupting the IT load. This translates to N+1 redundancy on all mechanical systems, 2N redundancy on power distribution, and multiple independent maintenance pathways. The facility layout follows a spine-and-leaf architecture with two independent power spines feeding four data halls. Each data hall provides 2MW of IT load capacity, for a total facility IT capacity of 8MW. The data halls are configured with hot-aisle containment and rear-door liquid cooling, with a raised floor plenum for cable management and supplementary air distribution. The structural design uses reinforced concrete with seismic base isolation rated for magnitude 7.0 events. TIA-942-B compliance extends beyond the mechanical and electrical systems to encompass physical security, fire suppression, and environmental controls. The facility features seven layers of physical security including perimeter fencing, vehicle barriers, mantrap entries, biometric access controls, CCTV coverage with 90-day retention, and a 24/7 on-site security operations center. Fire suppression uses a clean-agent system (Novec 1230) that protects electronic equipment without the residue or water damage associated with traditional sprinkler systems.

The facility's power system is designed around three Jichai 5MW natural gas generator sets operating in an N+1 configuration. Under normal operations, two generators carry the full 8MW IT load plus facility overhead (approximately 9MW total with PUE of 1.10), while the third generator stands by as a hot spare. The generators are fueled by the municipal natural gas pipeline, with on-site LNG storage providing 72 hours of backup fuel capacity. Jichai generators were selected for several reasons: competitive cost per MW compared to European alternatives, proven reliability in Central Asian climate conditions, excellent local service and spare parts availability, and compatibility with Kazakhstan's natural gas specifications. Each generator is housed in a weatherproof acoustic enclosure with integrated radiator cooling, fuel management, and emissions control systems. The power distribution architecture uses medium-voltage (10 kV) distribution from the generators to step-down transformers at each data hall, converting to 400V three-phase for the IT load. Uninterruptible power supply (UPS) systems using lithium-ion batteries provide 10 minutes of ride-through time on each power spine — sufficient to cover generator start-up in the event of a utility power failure. The municipal grid connection serves as an additional power source and provides power during low-load periods when running generators would be inefficient.

Cooling is the most technically complex aspect of an AI-focused data center, and our system reflects that complexity. The primary cooling method is direct liquid cooling (DLC) using coolant distribution units (CDUs) that circulate a water-glycol mixture through cold plates mounted directly on GPU and CPU packages inside the NVL72 racks. This approach removes approximately 85% of the heat load at the chip level, before it ever enters the data hall air space. The remaining 15% of heat — from power supplies, memory modules, and storage devices — is handled by rear-door heat exchangers (RDHx) mounted on each rack. The RDHx units use the same facility water loop as the CDUs, providing a unified cooling architecture that simplifies operations and maintenance. The combination of DLC and RDHx eliminates the need for traditional computer room air handlers (CRAHs), reducing fan energy consumption by approximately 80% compared to an air-cooled facility. Heat rejection to the atmosphere is accomplished using absorption-based heat management (ABHM) chillers — a technology borrowed from industrial process cooling. ABHM chillers use waste heat from the gas generators' exhaust to drive an absorption refrigeration cycle, producing chilled water without the large electric compressors used in conventional mechanical chillers. This approach converts waste heat into useful cooling, achieving a coefficient of performance (COP) of 1.2-1.4 while consuming minimal electricity.

Power Usage Effectiveness (PUE) is the industry-standard metric for data center energy efficiency, calculated as total facility power divided by IT equipment power. A PUE of 1.0 would mean all power goes directly to computing, which is physically impossible due to cooling, lighting, and other overhead. The industry average PUE is approximately 1.58, and even the most efficient hyperscaler facilities typically achieve PUE of 1.10-1.20. Our target PUE of 1.10 means that for every 100 watts consumed by GPU compute, only 10 watts go to facility overhead. This is achieved through the combination of direct liquid cooling (eliminating energy-intensive CRAH units), ABHM absorption cooling (using waste heat instead of electric compressors), high-efficiency power distribution (medium-voltage distribution minimizes conversion losses), and Almaty's favorable climate (enabling free cooling for 7-8 months per year). The financial impact of PUE 1.10 versus the industry average of 1.58 is substantial. For an 8MW IT load, a PUE of 1.58 means total facility power consumption of 12.64MW, while PUE 1.10 means total consumption of only 8.8MW. The difference of 3.84MW, running continuously at $0.048/kWh, represents annual savings of approximately $1.61 million. Over the facility's 15-20 year operational life, this compounds into tens of millions of dollars in saved energy costs — savings that flow directly to customers in the form of lower GPU-hour pricing.

The data center construction follows an aggressive but achievable 14-month timeline from groundbreaking to first customer load. Phase 1 (months 1-4) covers site preparation, foundation work, and structural steel erection. Phase 2 (months 4-8) focuses on building envelope completion, mechanical and electrical rough-in, and generator installation. Phase 3 (months 8-12) encompasses fit-out of the first two data halls, cooling system commissioning, and power system testing. Phase 4 (months 12-14) is dedicated to IT infrastructure installation, burn-in testing, and customer onboarding. The construction approach uses a modular design philosophy where possible. The CDU assemblies, power distribution units, and rack infrastructure are pre-fabricated off-site and delivered as tested, integrated modules. This reduces on-site construction complexity, improves quality control, and compresses the timeline. The generator sets are delivered as complete containerized units that require only fuel, exhaust, and electrical connections on-site. Key risks in the construction timeline include GPU hardware delivery (subject to NVIDIA allocation schedules), customs clearance for imported equipment, and seasonal weather constraints (Almaty winters can delay exterior construction). We have mitigated these risks through early engagement with NVIDIA's distribution channel, pre-clearance of customs documentation with SEZ Alatau authorities, and scheduling structural work to complete before the first winter season.

Qube Compute has entered into an investment contract with the Government of Kazakhstan through the Ministry of Digital Development, Innovation and Aerospace Industry. This contract, executed under Kazakhstan's investment promotion legislation, provides additional protections and incentives beyond the standard SEZ benefits. Specifically, the contract guarantees stability of the tax and regulatory regime for the duration of the project (10 years), provides preferential access to industrial power tariffs, and grants Qube Compute priority status for government procurement of cloud computing services. The government's strategic interest in the project reflects Kazakhstan's broader ambition to become a regional technology hub. The country's Digital Kazakhstan program identifies data center infrastructure and AI capability as key pillars of economic diversification away from natural resource dependency. Qube Compute's facility is one of the largest private technology investments in the SEZ Alatau's history. The investment contract also facilitates engagement with other government agencies that are critical to the project's success. The Ministry of Energy has committed to ensuring stable power supply to the SEZ, with a dedicated 35 kV feeder line from the regional substation. The Ministry of Industry has expedited building permits and environmental approvals. And the National Security Committee has streamlined the process for importing controlled technology items, recognizing the strategic importance of domestic AI compute capability.