"AI Costs Are Breaking the Bank"... AI Industry Focuses on Developing Lightweight Models

90% of Companies Facing Rising AI Costs
Lightweight Models Like China's Minimax M3 Are Shaping the Trend

Companies that had anticipated significant profit increases from adopting artificial intelligence (AI) are now facing substantial cost burdens due to a surge in the use of AI agents (assistants). In response, AI companies are developing lightweight AI models that maintain high performance while reducing operating costs and power consumption, addressing the needs of enterprise customers.

A post comparing the latest AI model from Chinese leading AI startup 'Minimax' with Claude Ops 4.8 from the global AI company Anthropic. It emphasizes that while the bug detection performance is the same, the cost is lower. Screenshot from X

On June 9, the U.S. consulting firm Bain & Company stated in its report, "Why AI Budgets Are Rising but Profits Aren't," that "out of a global survey of 951 companies, 37% attempted to cut AI-related costs by 11-20%, but in reality, reductions only reached 0-10%." Moreover, the report found that 90% of companies are actually increasing their AI-related expenditures.

U.S. market research firm Gartner also projected that by 2030, the cost of AI inference will decrease by more than 90% compared to current levels. However, it predicted that the actual AI-related costs that companies and individuals must pay will rise. This is because AI is evolving from simple Q&A to AI agents, which increases the number of tokens consumed per task. Gartner explained, "Although token prices are falling rapidly, the pace of usage growth is outpacing these declines," and added, "Infrastructure for high-performance AI remains a limited resource, so cost pressures will persist."

Recognizing this situation, AI companies have entered a race to develop lighter-weight models. On June 1, Chinese leading AI startup Minimax released its latest AI model, M3. M3 offers coding agent and work automation capabilities, yet its costs are only 5-10% of those of major U.S. closed-source AI models. It has even outperformed Google's generative AI model Gemini 3.1 Pro, achieving a 59% score on the SWE-Bench Pro software engineering benchmark. U.S. semiconductor company SiMa.ai is focusing on expanding the physical AI market with its low-power chip technology. SiMa.ai's machine learning system-on-chip (MLSoC) "Modalex" is designed to handle large language models (LLMs) even in environments consuming less than 10 watts of power.

South Korea's AI industry is also showing interest in lightweight AI models. AI startup Nota has developed an AI model compression platform called "NetsPresso." Through NetsPresso, large and complex AI models can be efficiently tailored to fit the needs of enterprise environments. Nota plans to use its NetsPresso technology to build an intelligent traffic monitoring system.

Another domestic AI startup, ActionPower, is also working on model compression and inference optimization for its AI-based handwriting application "Dagle." Leveraging its competitive advantage in speech and text data accumulated over more than ten years, ActionPower can optimize inference architectures for different hardware, such as graphics processing units (GPUs) and neural processing units (NPUs). An ActionPower representative said, "Lightweight models reduce reliance on high-end GPUs, allowing companies to use AI efficiently even in closed on-premises environments, where IT infrastructure is installed on their own premises. We also plan to actively target the ultra-lightweight, on-device AI market in the future."

This content was produced with the assistance of AI translation services.