Why DeepSeek beat ChatGPT in the App Store, plus Privacy, Data Center Investment, AI Acceleration

DeepSeek’s sudden rise to the top of the App Store is tied less to marketing and more to two product choices that make the model feel more controllable: it exposes reasoning in a way users can edit, and it delivers a free, widely available “reasoning model” experience that pulls in non-technical users. The reasoning display functions like a user-facing control surface—people can adjust prompts and immediately see what the system is doing. That transparency also appears to be feeding back into the broader industry: OpenAI is reportedly using DeepSeek’s openly shown reasoning outputs for model distillation, turning a UI feature into training signal.

A second driver is distribution strategy. By offering R1 for free in the App Store, DeepSeek reaches the “autocomplete crowd”—people outside tech who often dismiss chatbots as shallow, generic response generators. The reasoning-first interface undermines that critique by making outputs feel more grounded and inspectable. The result is a product that’s easy to evaluate in real time, which helps explain why the app climbed without relying on “gaming” the algorithm.

The momentum, however, comes with a sharp privacy and legal warning. DeepSeek’s terms of service are described as “creepy and concerning,” including claims that users can only seek redress through Chinese courts, that account deletion may not remove data, and that the company keeps a monitoring table for undefined “illegal activities.” The terms also reportedly don’t clearly grant users rights to model outputs. More troubling, the app is said to log keystrokes—an issue the transcript argues is easy to dismiss only if someone plans to run models locally. For the vast majority of users, the app route means accepting those data practices, and the concern is framed as worse than typical social-app tracking because it captures direct thinking and model outputs that may be difficult to reuse safely.

Beyond user impact, the competitive response in the US is portrayed as a catch-up sprint shaped by economics. Model makers are investing heavily in chips for next-generation systems, but the transcript distinguishes training from inference: most purchased compute goes to serving models to users, not training them. The claim is that DeepSeek’s earlier drop was tied to insufficient chips for inference at scale, and that industry defensiveness is understandable because serving demand is expensive.

Under-discussed—but presented as potentially transformative—is how DeepSeek’s approach is being replicated. The transcript argues that “reasoning models” can be trained more effectively now because large amounts of reasoning behavior have entered public datasets. Techniques involving group policy reinforcement and extracting reasoning from the training data were attempted earlier and failed; they now work because there are enough examples—cited as on the order of hundreds of thousands of responses—to reach a critical mass. The long-term implication is a feedback loop: models learn reasoning from data, produce better reasoning outputs, and those outputs become training material for the next generation, accelerating development toward something closer to self-improvement.

In short: DeepSeek’s App Store win is attributed to transparent reasoning and free access, while the broader industry is reacting on three fronts—privacy scrutiny, inference-heavy investment, and rapid replication of reasoning training methods that may speed up the next wave of model progress.