DeepSeek Is Free Now but What’s the Catch?

DeepSeek Is Free Now but What’s the Catch?

Title history
  • DeepSeek Is Free Now but What’s the Catch? — observed on 20261031-2302 UTC (YouTube)

by Kai · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQFGenerative artificial intelligence
🎙 Kai 👥 65K 📅 October 24, 2026 ⏱ 20 min 👁 117K 📄 news review 🧭 2026-10-31
Available in: English (current) Français

Keywords

KV cachecache hit rateDeepInferinference economicsself-hosting

Summary

The video analyzes DeepSeek’s recent release of its inference engine DeepInfer and the associated economics. It explains that the widespread reports of 80% profit margins are based on a misreading of the $38 per GPU-hour figure. The real story lies in the KV cache: DeepSeek serves 155 trillion tokens daily, but only a small fraction are actually computed; the rest are read from a distributed cache with a 96% hit rate. This yields an 11x reduction in compute per input token compared to the 56.3% rate from February 2025. The video details how the cache works, the price gap between cached and uncached tokens (50x), and how the economics only make sense because of this. It also discusses what DeepSeek open-sourced (engine, kernels, model weights) versus what it kept proprietary (the cache management and load balancing). Comparisons with vLLM show it matches DeepInfer’s benchmarks. Self-hosting requires a large cluster (128 GPUs) and substantial SSD storage for the cache. The video concludes with guidance on when to use DeepSeek and the key question to ask any provider: how long they retain your cache and whether it survives across machines.

188 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers a novel perspective on inference economics, shifting the focus from GPU compute to cache hit rates. It provides solid logical reasoning, backs claims with data from DeepSeek’s documentation and benchmarks, and exposes a common misconception. The argumentation is coherent and well-structured, gradually building from the headline numbers to the underlying mechanics.

Scientific Rigor, Source Quality, Title Accuracy

The video cites primary sources including DeepSeek’s GitHub repositories, API documentation, technical reports, and third-party analyses like SemiAnalysis. It critically examines the consistency of the numbers and acknowledges the limitations of open-sourcing parts of the stack. The title and content are well-aligned, as the video indeed reveals the ‘catch’ in the free release.

122 words

Title / Content Match

The title accurately reflects the content, as the video reveals the hidden 'catch' in DeepSeek's open-source release, namely the critical role of cache hit rates.

Quality & Reliability

8/10

The video provides a thorough analysis with clear reasoning, uses primary sources (DeepSeek documentation, benchmarks) and checks the consistency of numbers. It acknowledges limitations and alternative perspectives.

Chapters

Cited Sources

Concurring Sources

  • SemiAnalysis thread — SemiAnalysis agrees with the high profitability of DeepSeek's inference but the video highlights the cache-driven nature.
  • Yuchen Jin's post — Yuchen Jin's initial math shows the revenue but omits the cache role, which the video corrects.
  • OpenRouter providers — Confirms that many providers list the same prices without revealing cache hit rates, supporting the video's critique.

External References

Contribution & Novelties

The original contribution lies in reframing inference economics around the KV cache hit rate rather than raw GPU throughput. The video demonstrates that the advertised $38 per GPU-hour is not a measure of compute but of cache efficiency, and that the 56.3% to 96% improvement in hit rate translates to an 11x reduction in compute per input token. It also differentiates what DeepSeek open-sourced (engine, kernels, weights) from what it kept proprietary (cache management), providing a clear takeaway for evaluating any AI provider’s pricing.

Pour aller plus loin :

  • Mixture of experts — Relevant to DeepSeek’s architecture, which activates only a fraction of parameters per token.
  • Transformer (machine learning model) — Foundational architecture behind KV caches and attention mechanisms.
  • vLLM — Open-source inference engine that matches DeepInfer’s published benchmarks, illustrating the commodity nature of the released engine.

137 words

Radar Profile

The radar profile is consistently high across all dimensions, indicating a well-researched, technically rich, and reliable analysis.

Reliability 8/10

💬 Sur les 30 commentaires analysés, l'accueil est majoritairement positif, avec des éloges pour la clarté et la profondeur des explications, et quelques critiques sur le style de script généré.