Where AI Is Heading in 2025: Key Trends and Predictions for Next Year

1. Application Scenarios

The author believes enterprise AI will see massive growth next year for two reasons: consumer use cases are fragmented and still suffer from information hallucination, making "super AI" applications unlikely, while B‑side enterprises have vertical workflows, stronger payment willingness, and benefit from the Interleaved Thinking mechanism.

For example, Anthropic's new "scaffolding" tools and Skills feature help companies bridge LLM shortcomings in production.

In verticals such as finance and accounting, mature enterprise‑grade AI products are expected to emerge and generate tangible business increments.

2. Infrastructure Construction

GPU and TPU single‑card performance is nearing its limit, prompting a shift toward large‑scale compute clusters. These clusters consume massive power, and internal data‑transfer speeds become a key constraint on performance gains.

Consequently, power generation, energy storage, stable electricity supply, and optical‑module communications are poised for rapid growth, which will also drive short‑term price increases for industrial metals like copper, aluminum, and lithium.

3. Key Competitive Areas

Based on the past three years of AI development, competition will move from a single focus on model capability to a comprehensive contest involving technology stacks, ecosystem building, commercialization paths, and infrastructure.

Compute Dominance: Nvidia has long been the sole GPU provider, but Google's TPU+Gemini integration and Nvidia's new Blackwell architecture are reshaping the landscape.

AI Entry Points: ByteDance's Doubao AI phone assistant initially performed well, but was quickly suppressed by Alibaba and Tencent, illustrating that control over AI entry points equates to rule‑making power.

Flow of Traffic: In the AI era, traffic—who controls the AI usage入口—remains the core competitive factor, mirroring the internet era.

OS vs. Super‑Apps: Operating system vendors hold advantages in security compliance, while internet companies dominate traffic and scenario expansion.

4. Evolution of Technical Paradigms

The prevailing consensus is that World Model represents the future of AI. Current large models rely on textual tokens, creating a large comprehension gap with the physical world.

To achieve breakthroughs in virtual reality, robotics, and autonomous driving, AI must understand real‑world physics. This may involve shifting training inputs from pure text to multimodal "listen and see" (audio/visual) data.

For large‑scale deployment, AI must move from virtual to real environments, requiring long‑term memory capabilities on edge devices, substantial local storage for context retention, and compliance‑ready handling of privacy‑sensitive data.

Conclusion

Next year’s AI focus will concentrate on four pillars: infrastructure (power, compute, communication efficiency), technology stack (World Model, scaffolding tools), commercialization pathways (scenarios, traffic), and edge applications (smart devices, privacy compliance).