Why Is DeepSeek’s R1 Losing Users? Inside the Market Shift and Strategy

1. DeepSeek Market Cooling

After a spectacular launch that made DeepSeek R1 the fastest‑growing AI application worldwide, daily active users (DAU) peaked at 22.15 million but have since begun to fall sharply, with overall market share decreasing.

One major reason is that users are moving to third‑party platforms where DeepSeek models see a near‑20× usage increase, while official token consumption continuously drops.

Key pain points on the official platform include:

Higher first‑token latency and slower output speed due to batch processing of multiple requests.

Limited context window of only 64 K tokens, insufficient for large code or document analysis.

Third‑party services such as Lambda and Nebius offer context windows 2.5 times larger, attracting users seeking better performance.

2. DeepSeek’s Market Strategy

DeepSeek deliberately sacrifices user experience to keep computation costs low, prioritizing research resources over commercial revenue. The company focuses on advancing AGI rather than maximizing token sales through chat or API services.

This “lab‑first” approach contrasts with competitors that optimize for user‑facing performance.

3. Anthropic’s Counter‑Strategy

Anthropic, facing similar compute constraints, partners with cloud giants (Amazon, Google) to secure massive compute resources (e.g., 500 k Trainium chips, extensive TPU rentals) and refines its models to use fewer tokens per answer, improving overall efficiency.

4. Domestic User Sentiment

Chinese netizens criticize DeepSeek for slow speed, hallucinations, server overload, and content moderation, while some remain loyal and praise its open‑source ethos.

Others argue that official usage metrics are unfairly low because many deployments run on third‑party hosts.

5. Final Takeaways

The large‑model industry balances compute resources, business models, and technical progress. DeepSeek chooses a path of reduced user experience to maximize R&D capacity, while Anthropic seeks efficiency gains under resource limits. The ongoing race among models like GPT‑5, Gemini 3, and Grok 4 underscores that price wars are only the surface of deeper strategic battles.