Uzbektelecom AI Data Centers Push National Compute Beyond Tashkent
Uzbektelecom opened two AI-ready data centers more than 300 kilometers from Tashkent, shifting part of Uzbekistan’s computing capacity beyond its capital. The Uzbektelecom AI data centers in Bukhara and Kokand provide cloud services and dedicated GPU servers through the national UzCloud platform.
The launch matters because geographic distribution changes more than the location of the equipment. Uzbekistan is attempting to combine regional computing, sovereign cloud services, and a future photonic network into one national infrastructure layer. That approach contrasts with a model in which advanced computing remains concentrated in one city or depends heavily on foreign cloud regions.
The facilities create a credible foundation, but the announcement leaves important questions unanswered. Uzbektelecom has not disclosed GPU models, accelerator counts, power capacity, pricing, utilization targets, or measured workload performance. The story is therefore about a change in infrastructure strategy, not proof that Uzbekistan already has a mature AI compute market.
Uzbektelecom AI Data Centers Add Regional GPU Capacity
The immediate change is the arrival of operational GPU-backed cloud facilities in two regional cities.
Uzbektelecom commissioned the Bukhara and Kokand sites on September 22 during ICT Week Uzbekistan 2026. Uzbekistan’s minister of digital technologies, Sherzod Shermatov, and Japan’s ambassador to Uzbekistan, Kenji Hirata, participated in the opening.
The government’s data center announcement says the facilities support artificial intelligence, cloud services, big-data processing, and high-performance computing. Both operate within UzCloud, the national cloud platform developed by Uzbektelecom.
UzCloud’s listed services include infrastructure as a service, platform as a service, software as a service, S3-compatible object storage, and backup services. Infrastructure as a service gives customers rented computing resources without requiring them to own the underlying servers.
The AI portion centers on two offerings called AI Cloud and GPU as a service. GPU as a service lets customers rent access to graphics processors, which can execute many AI calculations in parallel, rather than purchasing and maintaining those processors themselves.
Uzbektelecom says the servers can support model training, inference, machine learning, image recognition, video analytics, large-data processing, and digital twins. A digital twin is a software representation of a physical asset or process that updates using operational data.
These use cases cover several distinct computing profiles. Training requires sustained processing and fast communication between accelerators. Inference focuses on running trained models, often with stricter response-time or cost requirements. Video analytics can combine high data volumes with near-real-time processing.
The company has not published enough technical detail to determine how well the facilities handle each profile. Buyers do not yet know the available memory per GPU, interconnect design, storage throughput, scheduling policy, or supported AI software environment.
That missing information does not erase the change. Regional organizations can now seek domestic computing capacity without building complete server rooms or sending every workload to Tashkent. The launch turns a multi-year infrastructure project into services that customers can begin evaluating.
The two projects received Tier III Certification of Design Documents from Uptime Institute. The independent certification registry lists Uzbektelecom Cloud Hub Bukhara 1 and Cloud Hub Kokand 1.
Tier III design certification evaluates whether the documented facility design supports concurrent maintenance. Operators should be able to service planned infrastructure components without shutting down the entire environment.
That designation is useful, but it has a defined scope. A design certification does not report GPU performance, service availability, customer support quality, security outcomes, or operating efficiency. It also should not be confused with proof that every operational process has been independently tested.
Uzbektelecom says both sites have redundant cooling and power systems, additional communications channels, and connections to its backbone network. They also connect to TAS-IX, Uzbekistan’s national internet exchange.
The company reports that its broader infrastructure follows ISO/IEC 27001, the international standard for information security management systems. Uzbektelecom also plans to pursue PCI DSS compliance for its data centers, which would address controls for payment-card data.
Together, these elements make the facilities more than isolated collections of GPU servers. They are regional access points inside a national cloud and telecommunications environment. The harder question is whether geographic distribution will translate into measurable reliability and useful capacity.
Moving Compute Beyond Tashkent Changes the Reliability Model
Bukhara and Kokand matter because physical distance gives Uzbektelecom options that another facility near the capital would not provide.
Both cities sit more than 300 kilometers from Tashkent. According to the ministry, they are Uzbektelecom’s most distant data center locations from the capital.
Geographic separation helps limit correlated failures. A serious power, network, or environmental disruption affecting one area is less likely to disable infrastructure hundreds of kilometers away. Customers can place replicas or backup workloads at another site instead of relying on two rooms within the same metropolitan region.
Uzbektelecom says workloads can shift between locations when one facility experiences a failure. That is the central promise behind its distributed design, but the practical result depends on software and networking.
A second building does not automatically provide disaster recovery. Customers need replicated data, compatible application environments, tested recovery procedures, and clear recovery-time objectives. They also need enough spare capacity at the receiving site to absorb displaced workloads.
AI adds further complications. Large model checkpoints can occupy substantial storage, while training datasets may be much larger. Moving those assets between cities can take too long if bandwidth, storage throughput, or orchestration tools become bottlenecks.
Inference applications introduce another concern. A public service, factory analytics system, or financial application may tolerate only a short interruption. Restarting the same workload elsewhere requires more than copying files. Identity controls, network routes, application state, and monitoring must also follow.
The regional sites can still improve resilience even before full workload mobility arrives. Organizations can choose a facility closer to their operations, keep backups outside Tashkent, or divide applications across locations.
Regional placement can also reduce the institutional concentration of computing. Universities, local authorities, and businesses outside the capital gain a clearer route to national cloud infrastructure. They no longer need to treat advanced compute as an exclusively capital-based resource.
Latency benefits are less certain. Physical proximity can reduce network distance for some users, but the result depends on local access networks and traffic routing. Uzbektelecom has not published latency measurements from regional customers to either facility.
The company’s backbone position gives it a structural advantage. Uzbektelecom can coordinate data center capacity with national transport networks and local connectivity. An independent cloud operator might need to negotiate those components separately.
That advantage also creates pressure. Customers will judge the service as one system, not as separate achievements in telecom, facilities, and cloud software. An available server provides little value when provisioning is slow, storage is constrained, or regional connectivity is inconsistent.
Uzbektelecom’s approach therefore pressures the capital-centric infrastructure model rather than one named competitor. The company is betting that distributed national capacity offers more strategic value than concentrating every advanced service around Tashkent.
The outcome will depend on whether users can operate across the sites without adding excessive complexity. If replication, monitoring, and failover remain manual, geographic diversity will look stronger on a map than inside a production application.
GPU as a Service Tests Uzbekistan’s Sovereign Cloud Strategy
The strategic contest is between accessible domestic compute and the scale advantages associated with larger external cloud regions.
Sovereign cloud infrastructure keeps selected data and computing under domestic jurisdiction and operational control. Governments often value that model for regulated records, public services, security-sensitive workloads, and applications subject to local data rules.
Uzbektelecom’s position connects sovereign infrastructure with consumption-based access. Organizations can rent compute, storage, platforms, and GPUs instead of financing complete facilities themselves.
This arrangement can lower the entry barrier for local AI projects. A university team could request GPU resources for a research run. A government agency could test document processing without purchasing accelerators. A retailer could evaluate video analytics before committing to a larger deployment.
Those possibilities do not guarantee adoption. GPU availability must be paired with development tools, supported frameworks, data pipelines, technical assistance, and predictable provisioning. Many AI teams need complete environments rather than access to raw accelerators.
Foreign hyperscale platforms generally offer broader catalogs, mature developer tooling, and large pools of computing resources. Their scale can make it easier to combine model hosting, databases, observability, security, and managed AI services.
UzCloud’s potential advantage is different. It can offer local hosting, integration with national connectivity, closer institutional relationships, and infrastructure aligned with domestic requirements. Those qualities matter most when data location and local support outweigh access to the largest global service catalog.
The correct comparison is therefore not GPU against GPU. Buyers must compare complete operating environments, including transfer time, compliance, support, failure recovery, software compatibility, and the effort required to move workloads later.
Uzbektelecom also needs a workable allocation system. Scarce accelerators create operational choices about reservations, queues, priority customers, and maximum job duration. Those policies influence whether small organizations receive meaningful access or wait behind larger users.
Pricing will matter, although the launch announcement provides no commercial figures. Consumption-based access sounds flexible, but customers still need predictable billing. Model training can generate sudden demand across compute, storage, and networking.
Hardware refresh cycles present another test. AI accelerators evolve quickly, and software support often follows specific hardware generations. A national provider needs a procurement and replacement strategy that avoids leaving customers on aging systems.
The earlier foundations of this project reach back several years. In January 2023, Toyota Tsusho, Internet Initiative Japan, NEC, and NTT Communications announced an infrastructure contract with Uzbektelecom.
That project covered data centers in Tashkent, Bukhara, and Kokand, along with transport, data, and international communications networks. Toyota Tsusho acted as coordinator, while IIJ supplied containerized data center modules and cloud-platform expertise.
NEC’s role included optical communications equipment and training. NTT Communications supported the international data network. Financing involved the Japan Bank for International Cooperation, Nippon Export and Investment Insurance, and MUFG Bank.
The 2026 openings are therefore not a sudden response to AI enthusiasm. They are the visible result of a broader telecommunications modernization program that began before today’s services reached customers.
AI has changed how that infrastructure is presented and potentially how it earns revenue. General cloud capacity can now be packaged as domestic access to scarce computing resources. Whether that shift succeeds depends on actual customer workloads, not the number of possible use cases listed at launch.
A Photonic Network Could Turn Separate Sites Into One Compute Fabric
Uzbektelecom’s larger mechanism is to connect distributed facilities so they can behave less like isolated data centers.
The company plans to commission another Tier III-designed facility in the Zangiata district of the Tashkent region. The ministry described it as the fifth site in Uzbektelecom’s geographically distributed network.
Zangiata will also host an All-Photonics Network pilot involving Uzbektelecom, Toyota Tsusho, NTT DOCOMO BUSINESS, and NEC. An All-Photonics Network, or APN, carries signals optically across more of the transmission path, reducing repeated optical-to-electrical conversions.
The pilot belongs to NTT’s Innovative Optical and Wireless Network concept, commonly called IOWN. The program targets high-capacity, low-latency communications that can connect computing resources across different locations.
Published APN pilot details say the first phase will link Uzbektelecom facilities in Ohangaron and Zangiata during 2026. The target is up to 800 gigabits per second and latency below one millisecond across routes of up to 100 kilometers.
Those targets come from the project participants and require operational validation. The report does not provide independent test results from the Uzbekistan deployment.
Plans for 2027 and 2028 would extend the network to four data centers, adding Bukhara and Kokand to two Tashkent-area sites. The proposed architecture includes redundant and protected communication paths.
If delivered, that network could make distributed compute more practical. Large datasets could move between facilities faster. Operators could place inference closer to users while retaining shared storage or training resources elsewhere.
It could also support remote GPU access. An application might send data to accelerator capacity in another facility while maintaining acceptable response times. That model becomes important when every site cannot justify an identical pool of expensive hardware.
NTT and DOCOMO have tested related ideas in Japan. A March 2026 remote GPU demonstration connected distributed accelerator resources and a 5G network through IOWN APN for low-latency video analysis.
The companies said that experiment met latency requirements assumed for remote robot control. It was a controlled demonstration, not evidence that Uzbekistan’s planned network will deliver identical results under commercial traffic.
Uzbektelecom must address longer distances as it expands toward Bukhara and Kokand. The first pilot’s sub-millisecond target covers routes up to 100 kilometers, while the regional sites are more than 300 kilometers from Tashkent.
Physics, routing, equipment, and operational overhead all affect end-to-end latency. A strong result on the initial link does not establish performance across the eventual national topology.
Bandwidth also needs context. An 800-gigabit connection is substantial, but shared capacity can fill quickly when organizations transfer training data, replicate storage, or stream video from many sources. Operators need traffic controls and capacity planning alongside headline bandwidth.
The project’s strongest potential lies in resource pooling. A connected network of facilities can shift selected workloads, share specialized hardware, and place services according to demand or availability.
That would give regional data centers a role beyond backup. They could become active parts of a national computing fabric. The distinction matters because idle disaster-recovery sites deliver less economic value than facilities handling daily production work.
The Missing Numbers Define the Risk
Uzbektelecom has established location, certification, and service scope, but it has not disclosed the figures needed to judge AI capacity.
The announcement contains no GPU inventory. It does not identify the accelerator vendor, hardware generation, memory configuration, cluster topology, or expected performance.
Power capacity is also absent. Data center operators commonly describe facilities through total electrical capacity or the power available to IT equipment. Without those figures, readers cannot compare Bukhara and Kokand with other regional or international projects.
The same gap applies to efficiency. Uzbektelecom has not published power usage effectiveness, water requirements, renewable-energy sourcing, or expected energy consumption for AI workloads.
These omissions matter because GPU servers place unusually high demands on power delivery and cooling. A facility designed for conventional enterprise servers may need further changes before supporting dense accelerator clusters at scale.
Tier III design certification addresses maintainability and redundancy in the facility design. It does not answer how much AI hardware is installed or whether customers can obtain it promptly.
The company also has not released utilization goals. Low utilization would suggest that domestic demand remains limited or that onboarding is difficult. Extremely high utilization could produce queues and constrain access.
Customer evidence is another missing layer. Uzbektelecom names model development, video analysis, image recognition, and digital twins as supported scenarios, but it has not identified production users at the two new facilities.
Early reference customers would help distinguish general-purpose infrastructure from an active AI market. Government agencies may become anchor users, but private-sector and research adoption will show whether demand extends beyond state-backed projects.
Operational transparency will be important. Customers need service-level commitments, incident reporting, support procedures, and clear explanations of where data resides. Regulated buyers will also want evidence about access controls and auditability.
Cybersecurity deserves equal attention. Connecting multiple facilities can improve resilience, but it also broadens the system that defenders must monitor. Identity, management interfaces, orchestration tools, and inter-site traffic all require protection.
Uzbektelecom says its infrastructure follows ISO/IEC 27001 and plans PCI DSS certification. Those frameworks can support disciplined controls, but customers still need service-specific documentation and tested response procedures.
Vendor concentration represents another uncertainty. Japanese partners supplied important infrastructure, networking technology, and operational knowledge. Long-term success requires local teams that can maintain systems, troubleshoot failures, and evolve the platform without constant external intervention.
The 2023 project included training for Uzbektelecom personnel, which addresses part of that issue. The next evidence should come from operating results, uptime records, service adoption, and successful capacity expansion.
There is also a commercial risk in building ahead of demand. AI infrastructure requires continued capital spending, while hardware can depreciate quickly. Uzbektelecom must attract workloads before equipment becomes less competitive.
The opposite risk is undersupply. If demand grows faster than procurement, limited capacity could favor a few large customers and weaken the promise of wider access.
For now, the responsible conclusion is narrow. The facilities increase Uzbekistan’s domestic AI infrastructure footprint and improve geographic diversity. Public evidence does not yet establish their scale, economics, or real-world performance.
Three Signals Will Show Whether the Strategy Works
The next phase should be judged through disclosed capacity, operational network tests, and visible customer adoption.
The first signal is technical disclosure from Bukhara and Kokand. Uzbektelecom should identify GPU configurations, available capacity, supported frameworks, storage performance, and service-level commitments.
Those details would let developers and enterprise buyers assess workload fit. They would also reveal whether GPU as a service is a specialized pilot or a platform intended for sustained commercial use.
Clear provisioning information matters as much as raw hardware. Customers should know whether they can reserve capacity, how jobs enter a queue, what data-transfer controls apply, and which tools monitor consumption.
Substantive disclosure would strengthen the case that Uzbektelecom is selling an operational AI environment. Continued silence would leave the market dependent on broad service descriptions.
The second signal is measured performance from the APN pilot. Uzbektelecom and its partners have specified ambitious bandwidth and latency targets for the Ohangaron-to-Zangiata connection.
Readers should watch for end-to-end test methods, sustained throughput, latency under load, failover behavior, energy measurements, and the applications used during validation. A demonstration that includes remote GPU workloads would directly support the distributed-compute thesis.
The more difficult milestone will arrive when the network extends toward Bukhara and Kokand. Those routes will test whether the architecture retains useful performance across much longer distances.
Successful expansion would strengthen Uzbektelecom’s claim that separate sites can function as a coordinated national resource. Delays or sharply weaker performance would make the facilities look more like independent regional clouds.
The third signal is customer adoption. Named production workloads, repeat usage, private-sector participation, university projects, and published utilization data would provide evidence that the infrastructure solves real problems.
The strongest examples would involve workloads that benefit specifically from local hosting or regional distribution. Public-service continuity, industrial video analysis, locally governed AI systems, and cross-site disaster recovery would all test the strategy.
Generic experiments are less informative. A short demonstration can show that software runs, but it does not prove economic value, operational reliability, or sustained demand.
Enterprise teams evaluating the platform should document requirements before selecting a location. Data residency, recovery time, accelerator memory, model portability, network latency, and support coverage can matter more than the AI label.
They should also preserve their research and operational decisions in a searchable AI knowledge base. Infrastructure choices become difficult to revisit when benchmarks, contracts, compliance notes, and incident findings remain scattered.
Uzbektelecom’s AI data centers have already changed the map of Uzbekistan’s computing infrastructure. They moved GPU-backed cloud services into Bukhara and Kokand and created a path toward a connected national compute network.
The next question is measurable: will customers receive enough capacity, reliable cross-site operations, and useful performance to treat that network as production infrastructure? Watch the hardware disclosures, APN test results, and customer workloads. Those three signals will show whether regional sovereign compute becomes a working platform or remains an ambitious infrastructure plan.
