Nvidia and Gemini 3 have changed the AI narrative this week. For years, the industry believed that pre-training scaling laws had reached their limit, meaning adding more computing power wouldn’t lead to significant model improvements. However, the launch of Gemini 3, which maintains the same parameter count of one trillion as its predecessor, Gemini 2.5, has proven otherwise. Gemini 3 not only surpassed GPT-5.1 on 19 out of 20 benchmarks but also became the first model to exceed 1500 Elo on LMArena. Oriol Vinyals from Google DeepMind attributed this leap to advancements in both pre-training and post-training techniques, suggesting that the gap between versions 2.5 and 3.0 is unprecedented and indicates no slowdown in progress. On the financial side, Nvidia's recent earnings report underscores a booming demand for AI infrastructure. The company projects $0.5 trillion in revenue through 2026, driven by their strong product lineup and design capabilities. They anticipate a market growth of $3 trillion to $4 trillion in AI infrastructure by the end of the decade. Nvidia's Q3 data center revenue hit a record $51 billion, reflecting a 66% year-over-year increase. Their new Blackwell Ultra GPUs are said to deliver training speeds five times faster than the previous Hopper generation. This suggests that the enhanced power from Blackwell will not only improve cost efficiency but also lead to tangible advancements in model capabilities. The developments around Gemini 3 and Nvidia’s financial performance challenge the notion that AI scaling has hit a wall. By demonstrating that pre-training scaling laws remain effective when combined with algorithmic improvements and robust computing power, these events signal a continued trajectory of growth and capability in AI models. The implications for future AI development are significant, suggesting that both the models and the infrastructure supporting them are evolving rapidly.