An MIT survey provides the first quantitative snapshot of AI infrastructure and adoption in Kazakhstan, Uzbekistan, and Kyrgyzstan, revealing rapid hardware expansion but persistent enterprise hesitancy.
Overview
The study, conducted by MIT’s Regional Entrepreneurship Acceleration Program (REAP), maps GPU clusters, local LLM fine-tuning activity, and enterprise deployment rates across the three countries. Data was collected through public tenders, cloud-provider logs, and interviews with 147 firms between January 2022 and December 2023.
Key Findings
- GPU cluster growth: 40 % year-over-year increase in on-premise and cloud-based GPU capacity, with Kazakhstan accounting for 68 % of the region’s total.
- Local LLM activity: A 3x surge in fine-tuning projects since 2022, driven by open-source models (Llama-2, Mistral) and government-funded AI labs.
- Enterprise adoption: Only 12 % of surveyed firms deploy AI models beyond proof-of-concept, with the remainder citing regulatory and talent constraints.
Bottlenecks
- Data sovereignty laws: National regulations require in-country data storage, complicating cloud-based AI workflows. Uzbekistan’s 2023 Digital Trust Law mandates local hosting for all government and financial-sector data.
- Talent pipelines: Universities produce ~1,200 AI/ML graduates annually, but 70 % emigrate within two years. Private bootcamps (e.g., Astana Hub’s AI Academy) have filled some gaps, yet demand outstrips supply by a factor of 3.5.
- Infrastructure gaps: While GPU clusters are expanding, average latency to Western cloud providers remains 120–180 ms, limiting real-time inference for latency-sensitive applications.
Workarounds and Local Solutions
- Hybrid cloud: Firms combine on-premise GPU clusters with edge nodes to comply with data residency rules while maintaining performance.
- Open-source tooling: Local developers rely on Ollama, vLLM, and Axolotl for fine-tuning, reducing dependency on proprietary APIs.
- Government incentives: Kazakhstan’s 2024 AI Development Roadmap offers tax breaks for firms that train models on local datasets and hire domestic talent.
Tradeoffs
- Cost vs. compliance: On-premise GPU clusters meet data sovereignty requirements but require upfront capital expenditure (CapEx) of $200,000–$500,000 per 8-GPU node.
- Latency vs. scalability: Edge deployments reduce latency but limit model size to 7B–1