Sep 6 – Sep 12, 2026

DeepSeek has published benchmarks that it says beat GPT-5.6 Sol and Claude Opus 5, at a cached input price of $0.003 per 1M tokens. The cost structure of the LLM market is being reset. In response, OpenAI is designing Jalapeño, an inference-only custom chip, with Broadcom, turning toward hardware vertical integration. Nvidia, with an ecosystem worth $11 trillion, strengthens its dominance but is also exposed to balance-sheet risk. Anthropic also publicly accused seven Chinese-linked labs of large-scale knowledge distillation attacks. Model IP protection and AI sovereignty are now turning into geopolitical conflict. On regulation, warnings about existential risk have accelerated AI policy discussions in Washington, DC. Companies have begun to treat safety validation as a task as important as performance competition. In the application layer, GPT-6 Astra is reaching into finance, CRM and government. The centre of gravity is clearly moving from general-purpose models to domain-specific ones. This week will be remembered as an inflection point, when competition over model performance widened into a multi-layered contest entangling infrastructure, policy and geopolitics.