The infrastructure supporting AI development is consolidating around whoever controls the compute, the data, or the distribution channel, while the actual safety and misuse problems remain unsolved and largely unmonitored. SpaceX's acquisition of Cursor signals that coding AI isn't a standalone business anymore but a feature to be bolted onto existing power structures with capital and scale. Anthropic's watermarking announcement addresses a symptom rather than the disease: Claude can now stamp its outputs, but the real question is enforcement and whether a watermark survives editing or copy-paste, and the company is already navigating the gap between what it can technically implement and what it can actually prevent. Meanwhile Amazon is harvesting Twitch streams for training data on an opt-out basis, which inverts the default assumption about consent and reveals how quickly data becomes a resource to be extracted rather than a choice to be made. The CSAM abuse case using Grok and the financial advice study showing blind spots in ChatGPT, Claude, and Perplexity both point to the same friction: these tools are deployed at scale without adequate friction for misuse or error, yet the industry response defaults to technical markers and voluntary opt-outs rather than architecture that prevents harm. Malaysia's emergence as a data center hub completes the picture. The real competition isn't about which model is smarter or safer. It's about who owns the compute, who controls the data pipeline, and who can move fastest to put tools in users' hands before accountability catches up.
Sloane Duvall