A confluence of factors drives the race to build sovereign AI infrastructure—Generative AI and Large Language Models (LLMs), escalating geopolitical and national security concerns, and NVIDIA’s aggressive expansion into the data center market with its powerful GPU hardware and software stack that fuels the next generation of neoclouds. As a result, “AI Factories” have emerged as a new class of digital infrastructure, attracting investment from governments, telecom operators, and other strategic players.
However, just as these GPU-based infrastructure projects are gaining momentum, the U.S. Bureau of Industry and Security (BIS) has introduced a new wave of proposed export rules — known as the AI Diffusion Rules — that are reshaping the global AI landscape. Our latest Insight Note, “Navigating Risks and Opportunities in Sovereign AI Infrastructure,” explores these regulatory developments and their implications for AI Factory deployments, particularly in Tier 2 countries. You can download the Insight Note here:
Key Takeaways
- Regulatory Quantity Constraints: AI Factories in Tier 2 countries (e.g., Vietnam, Singapore, India, and the Gulf States) as defined under BIS AI Diffusion Rules face a cap on the quantities of GPUs they can acquire. While this cap does not impact current operations, it will limit their future computational expansion and ability to train large language models.
- Shrinking Quantities and Growing Divide: As GPU performance scales up, export quotas could shrink dramatically. For example, Blackwell GPUs, which offer roughly 5x the total processing power (TPP) of H100 GPUs, may reduce the export cap from the equivalent of around 50,000 H100 units to fewer than 10,000 Blackwell GPUs, creating significant constraints for future deployments. Therefore, a growing divide is expected between Tier 1 and Tier 2 countries.
- Risks in Strategic Expansion Plans: While current AI Factories in Tier 2 countries remain within allowed thresholds, ambitious projects like Reliance’s planned 2,000 MW expansion have the potential to support computational capacities far exceeding the limits imposed by BIS rules.
- Opportunity in AI Inference: Model training remains the critical activity for installed GPU clouds, with AI Factories leveraging data privacy, residency, and security to serve national governments and enterprises. Although AI inference remains lagging, telcos are well positioned to play a role in this segment as the market evolves.
- Global Competition and Opportunity: The tightening regulatory environment highlights shifts in global competitive dynamics, where unrestricted access for Tier 1 entities could further challenge Tier 2 ventures in building locally sovereign AI infrastructures.
Investment Risk
Export controls are not the only risk facing AI Factories. One of the most significant risks is financial, stemming from the unpredictable timeline for enterprise AI adoption and the challenge of monetizing consumer AI services and applications. At this early stage of AI adoption, revenue streams remain uncertain. Coupled with high capital expenditure requirements, AI Factories face substantial financial risks.
AI Factories carve out a niche by addressing data sovereignty and security requirements for national governments and large enterprises. Currently, model training serves as the primary revenue driver, making it particularly susceptible to the impact of export controls. Inference, while still underdeveloped, holds significant potential for telcos in particular to capitalize on as the market matures.
To mitigate financial risks, investments are approached cautiously, with many AI Factories starting with a modest deployment of GPUs—typically between 1,000 and 2,000 units. SoftBank, which boasts one of the highest computational capacities with approximately 20,000 GPUs (including H100 and some B100 units), primarily utilizes this capacity for internal use cases, such as training its telco LLM. Telenor, an early adopter of the concept, has secured four customers within a year, while Telus, which entered the market just six weeks after exploring the idea, is still working to establish its position.
Another significant risk is the rapid evolution of AI chip technology, which drastically increases computational capacity while quickly rendering older GPUs obsolete. This steep depreciation rate for AI hardware necessitates a rapid revenue ramp-up to offset costs. However, accelerating revenues presents its own challenges. For instance, in the enterprise segment, one critical gating factor is accessing, cleaning, and organizing data for AI processes—a time-consuming and expensive endeavor.
Concluding Remarks
The rapid emergence of neoclouds and AI Factories brings with it a host of complex decisions for investors looking to capitalize on this next wave of digital infrastructure. From defining the right strategy for each layer of the AI stack to building a differentiated ecosystem that aligns with distinct business models, success hinges on making the right foundational choices early on. Monetizing these ecosystems without clashing with partners, anticipating future technical and regulatory shifts, selecting optimal deployment models (whether edge, hyperscale, or hybrid), and navigating external disruptions like geopolitics and chip supply challenges are all pressing concerns.
P.S. If you’d like to dive deeper into these challenges, feel free to reach out. Our team at Xona Partners has helped clients navigate these complexities and develop tailored strategies that turn ambition into competitive advantage.