The core of AI progress is a fierce competition between NVIDIA's GPU roadmap (Hopper, Blackwell, Rubin) and Google's custom TPUs. The transition to Blackwell is a monumental engineering challenge involving liquid cooling and massive power increases, which has temporarily given Google a cost advantage in token production.
A clear hierarchy has emerged, with OpenAI, Google's Gemini, Anthropic, and xAI as the four dominant frontier labs. Despite massive investment, other tech giants like Meta, Microsoft, and Amazon have failed to develop competitive internal models, highlighting the extreme difficulty and specialized expertise required.
Previously, AI models were static after pre-training. The introduction of reasoning capabilities now allows for a powerful feedback loop where user interactions and preferences can be used to refine and improve the model, creating a data flywheel similar to those that built internet giants like Google and Netflix.
AI is forcing a fundamental rethink of software business models. Traditional SaaS companies with 70-90% gross margins are struggling to adapt to AI-native models that may operate closer to 40% margins. This transition is a 'life or death' decision for incumbents.
The performance gap between American and Chinese AI is widening, primarily due to hardware access. Chinese open-source models, while impressive, are falling behind as they lack access to the cutting-edge NVIDIA Blackwell chips that will power the next generation of US-based frontier models.
Keep pulling the thread on Gavin Baker.