- Two SOTA foundation models, Kimi K3 and Qwen 3.8, were launched, challenging top-tier models and proving open models can reach the frontier.
- Foundation models are expensive due to research, compute, and electricity; post-training, inference costs (mainly compute and electricity) dominate.
- Companies vary in infrastructure ownership: leasing data centers (e.g., Anthropic), building data centers (e.g., Meta, Alibaba), or owning power plants and data centers (e.g., SpaceX), affecting cost structures and margins.
- Model-only providers face high risk as success depends on being the best or cheapest, requiring constant innovation or unique products to avoid commoditization.
- Anthropic’s focus on regulation and self-improvement, along with high costs, puts it at risk compared to OpenAI’s product moats and infrastructure investments, increasing unbundling vulnerability.