Jan 17, 2025
The United States' latest move to expand AI export restrictions on closed general models and high-performance computing chips is set to drive the deployment of smaller, more specialized language models. These restrictions, which limit the export of models exceeding 1 trillion parameters and impose stringent computational constraints, signal a shift in how AI will be developed and deployed worldwide.
The Proposed U.S. AI Export Restriction
The U.S. Interim Final Rule on AI diffusion introduces a structured hierarchy for AI access:
Tier 1 Nations (Australia, Japan, Taiwan, the U.K., and most of Europe) retain unrestricted access but must keep 75% of AI computing within allied countries.
Tier 2 Nations (Israel, Saudi Arabia, Singapore) face initial computational caps that will gradually increase over time.
Tier 3 Nations (China, Russia, Iran, North Korea, and others) are outright blocked from receiving advanced AI chips, model weights, or related technology.
The rules also regulate the transfer of model weights for closed models trained beyond 10^26 operations. Open-weight models remain unrestricted, likely boosting the open-source AI movement.
The proposed US. AI Export Regulation divide the world in 3 hierarchies. Source: DeepLearning.ai
The Changing AI Landscape
If Ilya Sutskever’s claim — pre-training, as we know it, is dead because we have just one internet — was not enough to hint at an LLM plateau, the U.S. government's move puts another friction. As per the new rules, closed general models trained beyond 10^26 computational operations will require export approvals. This means that state-of-the-art AI models like GPT-4 will either become unavailable or severely restricted outside U.S.-allied nations.
A simple computation illustrates this threshold:
High-performance models are typically trained on ~15 trillion tokens. Llama 3.1 405B, Llama 3.3 70B, and DeepSeek-v3 were all trained on ~15T tokens.
The total FLOPs required for each token is approximately 6 times the model size.
Total FLOPs = Number of Parameters × Training Tokens × 6
Thus, any model with parameters beyond 10^26/(6 x 15 x 10^12) ≈ 1 trillion parameters will be subject to export restrictions.
This effectively curtails access to cutting-edge AI outside the U.S. and its closest allies, prompting a strategic shift in how enterprises build and deploy AI.
The Future of LLM Deployment
As we enter 2025, enterprises are moving beyond the GenAI hype and focusing on return on investment (ROI). The excitement of 2023 and 2024 has given way to pragmatism—use cases need to demonstrate real economic value rather than being driven by AI influencers hyping AGI as almost solved. Many enterprises saw their GenAI projects shelved due to an inability to justify cost versus impact.
A fundamental realization is emerging: using massive general models for every task is akin to using a firehose to water a plant. While large models excel in deep reasoning and exploratory applications, they are neither cost-effective nor experienced for domain-specific tasks. Instead, the future lies in a hybrid approach where general models and domain-adapted models work together.
Enter Domain Memory Agents
Domain Memory Agents (DMAs) are smaller, specialized language models designed for specific domain tasks. Unlike stateless general models like GPT-4, which process each query independently, DMAs retain domain-specific memory, allowing them to refine their understanding over time. Think of them as AI systems that have "practiced" a particular task extensively, much like a human expert who learns from experience.
As AI adoption matures, DMAs will play a crucial role in bridging the gap between general-purpose reasoning and real-world enterprise workflows. While large models can still handle broad reasoning tasks, DMAs will be embedded within these workflows to inject domain intelligence, ensuring higher accuracy and cost efficiency.
Consider these examples:
SQL Generation from User Intent: Converting vague user intent into SQL queries is a two-step process. First, the system must interpret the user’s request in the context of the specific database schema, identifying the relevant tables and relationships. Second, it must construct a valid SQL query based on this plan. DMAs excel at the first step—understanding the domain context—while generalist models can efficiently handle the second step, provided scale and cost are not concerns.
Engineering Failure Analysis: Diagnosing engineering failures often requires an expert who can draw on past resolutions and experience to hypothesize potential issues. A DMA can act as this domain expert, identifying likely failure points based on historical data and contextual knowledge. Generalist models can then assist by fetching supporting documentation, querying logs, or summarising potential fixes.
DMAs are designed to learn from iterative feedback, improve continuously, and operate at a fraction of the cost of running massive general models. This shift aligns with the intent of the U.S. policy—if not to stifle innovation, export restrictions will certainly accelerate this transformation. With limits on large models (>1T parameters), organisations will need alternatives. Tier 2 and 3 countries face significant compute restrictions, and will turn more towards DMAs or customised small language models.
Implications for Global AI Development
These restrictions are poised to reshape AI development in profound ways:
Decentralization of AI Development: Countries outside U.S. influence will be incentivized to develop their own AI ecosystems, leading to more localized AI research and development.
Rise of Open-Source Models: With restrictions applying only to closed models, open-weight alternatives will gain traction as enterprises seek flexible, scalable AI solutions.
Increased Investment in Specialised AI: Companies will prioritize fine-tuned models and Domain Memory Agents over generic large-scale models, driving innovation in domain-specific AI applications.
The Road Ahead
While U.S. policymakers aim to safeguard AI leadership and national security, these export controls are likely to accelerate a shift toward domain-specialized AI models. Enterprises and governments worldwide will double down on developing and deploying Domain Memory Agents (DMAs), which provide greater cost efficiency, data control, and superior domain-specific performance.
In the long run, we may see a bifurcated AI landscape: one where general AI models dominate foundational reasoning tasks, and another where globally distributed, domain-specialized AI systems drive industry-specific intelligence. Regardless of geopolitical motivations, the momentum toward DMAs is now unstoppable. Enterprises that fail to adapt risk falling behind, while those that embrace domain-adapted AI will gain a strategic edge in efficiency, accuracy, and control.