LLM 2Q25: Recent Progress of Agentic AI Models

We have discussed GPT5 in our previous article. And here is a more detailed analysis of recent progress of agentic AI models.

We track LLM progress every quarter, and you can also check out our LLM 1Q25 report. It contains a wealth of first-hand research information.

Executive summary

We have analyzed the latest AI model development trends, and our conclusion is as follows: the current progress in coding and agent models, combined with mid-training and reinforcement learning (RL), can support rapid application growth for the next 1–2 years. Therefore, the inference compute demand in 2026 is relatively clear and predictable.

On the other hand, continuing to follow the scaling law in 2027–2028 will require breakthroughs in new model training, which in turn will drive substantial training compute demand during H2 2025–2026. Hence, barring any major macroeconomic disruptions, we believe the compute demand in 2026 is highly foreseeable.

We also discussed how model evolvement would change the battle of GPU v.s. ASIC. With increasing demand for inference, esp. decoding for coding and agentic tasks, we believe ASIC will gain more market shares from GPU. However, with the demand of pretraining and midtraining, GPU will also benefit from the industry beta.