The economics of artificial intelligence (AI) models are governed by a complex interplay between training and inference costs. High training costs and rapidly declining inference costs create both challenges and opportunities for developers and users. Understanding these dynamics is crucial for assessing the trajectory of AI development, particularly concerning the competition between closed and open-source models.

Training the most powerful Large Language Models (LLMs) has become extraordinarily expensive, with costs rising into the billions of dollars as models push towards larger parameter counts and more complex architectures. This capital-intensive race may paradoxically accelerate commoditization, leading to diminishing returns as output quality converges across different providers, making sustained differentiation harder.

While training costs soar, the cost of inference, or applying these models, is falling quickly. Improvements in hardware, such as NVIDIA's Blackwell GPU, which consumes 105,000 times less energy per token than its 2014 predecessor Kepler GPU, coupled with algorithmic efficiencies, are driving down inference costs. This decline intensifies competition among LLM providers, focusing pressure not only on accuracy but also on latency, uptime, and cost-per-token.

Lower inference costs are a boon for users and developers, providing access to powerful AI at dramatically reduced unit costs, accelerating the creation of new products and services, and boosting user adoption. However, this also creates monetization and profit challenges for model providers. High training costs combined with low serving costs and decreasing pricing power put business models in flux, prompting questions about the viability of one-size-fits-all LLMs versus smaller, custom-trained models.

As LLMs mature and competition intensifies, open-source models are making a comeback due to their lower costs, growing capabilities, and broader accessibility. Platforms like Hugging Face facilitate easy access to models such as Meta’s Llama and Mistral’s Mixtral, granting startups, academics, and governments access to frontier-level AI without massive capital outlays. China is also playing a significant role in the open-source movement. As of Q2:25, China is leading with the release of three large-scale models in 2025: DeepSeek-R1, Alibaba Qwen-32B, and Baidu Ernie 4.5.

The rise of open-source models, fueled by lower costs, is leading to increased usage among developers. Open-source models are supporting sovereign AI initiatives, local language models, and community-led innovation. Closed models are dominating consumer market share and large enterprise adoption. This divide represents a competition between freedom and control, speed and safety, and openness and optimization. This split influences not only how AI works but also who can wield it.

The global race to develop and deploy AI systems is increasingly a strategic rivalry between the United States and China. The USA has led in model innovation, custom silicon, and cloud deployment, while China is advancing in open-source development and national infrastructure. This competition underscores the importance of sovereignty, security, and rapid innovation.

The future of AI model development will likely be shaped by the interplay between closed and open-source models. Horizontal platforms will push breadth, stitching together knowledge across functions, while specialists will push depth, delivering AI that speaks the language of compliance, contracts, and user intent. The key will be to abstract the right layer, own the interface, and capture the logic of work itself. China is developing local AI platforms, while global platforms dominate elsewhere. As AI continues to evolve, its strategic importance will only intensify.

What does the Stanford AI Index report say about trends in the cost of training notable AI models over time (including any examples like ImageNet models or large language models)?

Related Content From The Pandipedia