US companies continue to lead in AI model releases, but China's output is closing the gap. According to research institute Epoch AI, US organizations released 50 notable models in 2025, while China's industry released nearly as many, at 87. Industry now accounts for over 90 percent of notable AI models, up from just under 50 percent in 2015.
Nvidia has benefited most from the growth in global AI compute capacity, which increased 3.3 times every year since 2022. Total AI compute has grown 30-fold since 2021, with Nvidia accounting for over 60 percent of total capacity.
Training large language models can generate enormous carbon emissions, with estimates ranging from 5,184 tons (OpenAI's GPT-4) to 140,000 tons (Epoch AI's estimate of xAI's Grok 4). Emissions from model inference are also increasing, with some models consuming up to 23 watts per prompt.
The capabilities of AI models have improved rapidly, with multimodal LLMs conquering benchmarks nearly as quickly as they can be invented. However, models still struggle with simple tasks like reading analog clocks and understanding calendars.
