English BTUAI.ge Version Is It Worth Building AI Infrastructure in Georgia?

Georgia should build AI infrastructure, but in phases and as a shared platform. A 64-GPU centre would require an estimated GEL 9.3 million in initial investment and, at maximum continuous load, about 0.97 GWh of electricity a year – roughly 0.007% of Georgia’s 2025 domestic consumption. The binding constraint is utilisation, not national energy volume. Under the scenario used here, local ownership becomes cheaper than comparable foreign cloud capacity at around 45% sustained utilisation. A shared 64-GPU pilot is therefore defensible; a hyperscale project is not, unless international customers and new power supply are secured first.

Infrastructure Means More Than GPUs

AI infrastructure includes accelerators, high-speed networking, storage, resilient power, cooling, cybersecurity, workload orchestration, data governance and a specialist operating team. A missing layer can turn expensive hardware into underused capacity.

Three scales must be separated. An eight-GPU node can support research and pilots. A shared 64–256 GPU centre can serve universities, startups, government and companies. A 50–100 MW hyperscale data centre is a different industry: it requires international anchor clients, new generation and long-term contracts. Treating all three as the same investment question produces the wrong answer.

Why Demand Is Emerging

In the second quarter of 2026, 47% of companies surveyed by the Business Association of Georgia were researching or planning AI, 24% were piloting it and 13% had fully integrated AI into at least one process. Sixty-five percent expected to raise AI budgets over the following one to two years. This is not a GPU-demand forecast, but it shows a transition from curiosity to organisational use.

Half of surveyed firms still relied on ready-made global platforms. Cloud remains rational for small and volatile workloads. Local capacity becomes valuable when demand is recurring, sensitive data should remain under local control, guaranteed access matters or several organisations can aggregate demand.

Geostat reports that 94.9% of enterprises had internet access in 2025, but connectivity does not equal AI readiness. Structured data, digitised workflows and skilled staff may be more restrictive than bandwidth. Infrastructure cannot manufacture demand where usable data and business cases are absent.

The 64- and 256-GPU Scenarios

The model uses public benchmarks. A listed Supermicro eight-GPU H100/H200-class system is about USD 328,657. A 35% allowance is added for high-speed networking, storage, cooling, deployment, transport and contingency. At the National Bank of Georgia’s official exchange rate on 27 August 2026, one deployed node is roughly GEL 1.16 million. Eight nodes, or 64 GPUs, total about GEL 9.3 million; 32 nodes, or 256 GPUs, about GEL 37.2 million.

These are planning scenarios, not procurement quotes. Actual cost depends on GPU generation, export rules, software licensing, network topology, storage, warranty and whether an existing facility can be used.

Source and status: scenario calculations by BTU researchers using public technical and pricing inputs.

Energy Is Manageable at This Scale

NVIDIA specifies maximum system power of 10.2 kW for an eight-GPU DGX H100/H200-class node. With a target PUE of 1.35, 64 GPUs would consume at most about 0.97 GWh annually and 256 GPUs about 3.86 GWh.

Georgia’s domestic electricity consumption was about 14.3 TWh in 2025. The scenarios would represent roughly 0.007% and 0.027% respectively. Using the 2026 public-supply tariff for 35/110 kV customers, annual electricity cost at continuous maximum load is about GEL 293,000 and GEL 1.17 million. Resilient feeds, UPS, backup generation and reliability controls remain essential.

Hyperscale changes the equation. A continuous 50 MW load consumes about 438 GWh a year, or 3.1% of national consumption; 100 MW consumes 876 GWh, or 6.1%. Such projects require additional generation and grid reinforcement rather than relying on the existing system.

Cloud Versus Local Ownership

AWS publishes a Capacity Blocks price of USD 37.76 per hour for an eight-H100 configuration. At that benchmark, continuous use of 64 GPUs would cost about GEL 6.93 million annually. The local 64-GPU ownership scenario – five-year depreciation, electricity, maintenance and a six-person core team – costs about GEL 3.15 million a year. Approximate break-even appears at 45% average utilisation.

At 25% utilisation, cloud is about GEL 1.42 million cheaper. At 50%, local ownership is ahead by roughly GEL 313,000; at 75%, by about GEL 2.05 million. Scale lowers the 256-GPU break-even to about 43%. A stress case with higher build costs, a PUE of 1.54 and a higher electricity tariff moves the 64-GPU threshold to about 51%.

The comparison is not perfectly like-for-like. Cloud can be cheaper under reserved or interruptible contracts but may add storage and data-transfer charges. Local ownership offers control and guaranteed capacity but carries technology obsolescence, low-utilisation and staffing risk.

A positive local advantage means local ownership is cheaper. The comparison uses AWS published Capacity Blocks pricing.

What Local Compute Should Do

Georgia should not build a global foundation model from scratch merely because it owns GPUs. A 64–256 GPU platform is better suited to adapting open models for Georgian, training domain models, secure inference on local data, university and startup research, cybersecurity, financial modelling, video analytics and testing agentic systems.

The case improves when capacity is shared. A single university, ministry or company may struggle to sustain 45% utilisation. Aggregating demand from universities, banks, telecoms, health, manufacturing, government and startups can lift use and lower unit cost. That requires transparent quotas, pricing, data isolation, independent security assurance and clear service levels.

The Largest Risk Is Empty Capacity

The costliest mistake is buying excess capacity early. GPUs depreciate rapidly as new generations deliver more performance per watt. At low utilisation, fixed cost per GPU-hour rises sharply. The essential KPIs are therefore used or sold GPU-hours, availability, repeat demand and full cost per GPU-hour – not the number of GPUs purchased.

Human capital is the second constraint. Data-centre operations, networking, cybersecurity, MLOps and domain research are all required. Vendor concentration, export rules and spare-parts availability create further risk, so the architecture should support hybrid operation and switching where feasible.

BTU Researchers’ Assessment

According to BTU researchers, Georgia should pursue demand-led rather than maximum-scale infrastructure. Stage one combines an eight-GPU node with cloud capacity to measure demand. Stage two moves to a shared 64-GPU centre only when anchor users support at least 50% expected utilisation over 24 months. Expansion to 256 GPUs should occur only after the existing centre sustains 65–70% utilisation for six months and a queue of unmet demand emerges.

The state or a university should not be only an owner. An independent operator, separate commercial and research tariffs, audited utilisation data and private co-investment are needed. The goal is to serve a Georgian AI ecosystem, not to create an expensive symbol.

Key Findings

• A shared 64-GPU centre is estimated at GEL 9.3 million initially; a 256-GPU platform at GEL 37.2 million.
• Energy demand is small at 64–256 GPUs; sustained utilisation and human capital are the binding constraints.
• Cloud is better for small and volatile demand; local ownership becomes competitive around 43–46% sustained utilisation.
• A stress case moves the 64-GPU threshold to about 51%, making anchor commitments essential.
• A 50–100 MW hyperscale centre would consume roughly 3.1–6.1% of Georgia’s annual electricity and requires new generation.
• The realistic opportunity is Georgian-language and domain AI, secure inference, research and agentic systems – not a race to train a global frontier model from zero.
• The recommended path is hybrid and phased: cloud plus eight GPUs, then 64 GPUs after proven demand, and 256 GPUs only after demonstrated use.

Why This Matters for Georgia

Local compute does not create full technological independence, but it creates choice: where sensitive data is processed, who can access research resources and how exposed local projects are to foreign pricing and availability. That matters for Georgian-language AI, public data, cybersecurity, finance, health and education.

Without financial discipline, the same infrastructure becomes an expensive idle asset. Investment should therefore be tied to committed GPU-hours, real data projects, energy availability, talent and explicit expansion thresholds.

Conclusion

It is worth building AI infrastructure in Georgia if the country begins with a shared hybrid platform around or below 64 GPUs and expands only after demand is proven. Energy does not constrain this scale; economics depends on utilisation above roughly 45–50%, competent operations and multiple users. A hyperscale centre is justified only with international anchor customers, additional power and long-term contracts. Georgia does not need to choose between infrastructure and cloud. It needs a disciplined combination of both.

Data and Main Sources

• Business Association of Georgia, Q2 2026 AI adoption survey.
• Geostat, ICT Use in Enterprises 2025 and national energy statistics.
• GNERC 2025 annual report and 2026 non-household electricity tariffs.
• NVIDIA DGX H100/H200 technical and data-centre planning specifications.
• AWS Capacity Blocks pricing and Supermicro published server benchmark.
• National Bank of Georgia official USD/GEL rate, 27 August 2026.

Authorship

Prepared by the academic team of Business and Technology University and the BTUAI Research Team, Tbilisi, Georgia.