DeepSeek choque encore toute l'industrie de l'IA (actu IA)

Item: DeepSeek choque encore toute l'industrie de l'IA (actu IA)
Rating: 4
Author: PostSingularity Institute

Title history

DeepSeek choque encore toute l'industrie de l'IA (actu IA) — observed on 20261028-1917 UTC (YouTube)
DeepSeek lâche une nouvelle bombe (actu IA) — observed on 20261028-1353 UTC (YouTube)

by Eliott Meunier · View the channel on PSI

Formal & Physical Sciences AI — News & Applications UYComputer science UYQArtificial intelligence (AI)

🎙 Eliott Meunier 👥 56K 📅 October 28, 2026 ⏱ 32 min 👁 13K 📄 news review 🧭 2026-10-28

Available in: English (current) Français

Keywords

DeepInferinference optimizationKV cacheMixture of ExpertsAI cost per GPU

Summary

This video reviews the latest developments in AI, focusing on DeepSeek’s release of DeepInfer, its production inference engine, along with detailed operational numbers. The presenter explains the technical challenges of serving large models efficiently, such as distributing weights across GPUs, separating prefill and decode processes, compressing the KV cache, and balancing experts in a mixture-of-experts architecture. DeepSeek’s disclosed figures reveal a revenue of $38 per GPU per hour, leading to a payback period of about two months for a GPU. Independent verification by vLLM confirms these numbers, and the video contrasts this with the costs of renting GPUs from neo-cloud providers. The broader implications for data center investments and AI pricing are discussed, including a potential collapse in inference costs. The video also covers several new model releases: WeVisOne-VL for vision, Kimi K3.1 Pro (competitive and cheaper), MiniMax M3.1 Flash for agents, MemT with memory in weights, a DNA-reading model (Carbon-X), and a Chinese medical model (Ling 3.0 Flash Santé). Each is briefly described in terms of performance, pricing, and licensing. The presentation is clear, well-structured, and backed by numerous links to primary sources.

184 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video offers substantial value by dissecting the economics and technical innovations behind DeepSeek’s inference efficiency, a topic rarely covered in such depth on mainstream tech channels. The argumentation is strong: the presenter builds a coherent case from the open-sourced code and production metrics, cross-validates the $38/GPU/hour figure with an independent test from vLLM, and explains the levers of optimization (KV cache compression to 1/437th, split prefill/decode, expert allocation) in a way that connects directly to cost models. The reasoning is generally sound, though some extrapolations (e.g., global data center oversupply) are presented as plausible scenarios rather than certainties. The use of concrete numbers, such as 155 trillion tokens per day and a 96% cache hit rate, grounds the discussion in data rather than speculation.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high for a YouTube video. The creator cites primary sources, including DeepSeek’s GitHub repository, official API pricing, independent benchmarks from Artificial Analysis, and blog posts from reputable analysts (Simon Willison, Zvi, SemiAnalysis). These sources are diverse and credible. The video also acknowledges uncertainty where it exists, such as the exact number of GPUs DeepSeek operates. The title-content match is strong: the ‘bombe’ refers to the unexpected open-sourcing of DeepInfer and the revealing cost structure, which is indeed the core of the video. The rest of the news items are clearly secondary. No conflicting sources are presented, but the video does not claim to be exhaustive or without bias.

252 words

Title / Content Match

The title accurately highlights DeepSeek's surprising open-source release of its inference engine, which is the central topic; the rest of the video covers additional AI news as implied by the tag '(actu IA)'.

Quality & Reliability

8/10

The video presents a well-researched analysis of DeepSeek's inference engine release, supported by multiple technical sources and independent performance checks. The economic breakdown is consistent with external data, and the video clearly distinguishes between verified facts and speculative extrapolations.

Chapters

Cited Sources

DeepInfer official repository — The open-sourced inference engine and source of production figures.
DeepInfer Q&A documentation — Official details on deployment and optimization choices.
DeepSeek V4.1 Flash model page — Model weights and technical specifications.
vLLM verification blog — Independent performance measurements matching DeepSeek's claimed figures.
DeepSeek February 2025 inference overview — Background on prior inference system architecture.
DeepSeek API pricing — Official pricing used to estimate revenue per GPU.
Artificial Analysis: DeepSeek V4.1 Flash — Independent model performance and cost benchmarks.
Artificial Analysis: Claude Opus 5.5 — Comparison point for Kimi K3.1 Pro.
SemiAnalysis: Data centers 2027 — Analysis of data center projections and demand.
Simon Willison on PHP vulnerability — Coverage of the PHP security issue in the cyber wave.
Zvi's summary of the hacking wave — Overview of the security vulnerabilities discussed.
MemT research paper — Paper on the memory-in-weights model.
Carbon-X model page — DNA-reading model.
Ling 3.0 Flash Santé — Chinese medical model.
MiniMax M3.1 Flash announcement — Official release notes for the agentic model.
Kimi K3.1 Pro announcement — Official announcement of Kimi model.

Concurring Sources

vLLM verification blog — Confirms DeepSeek's claimed $38/GPU/hour performance.
Artificial Analysis DeepSeek V4.1 Flash — Independently measures model efficiency and pricing.

External References

Contribution & Novelties

The video provides a rare, detailed breakdown of DeepSeek’s inference cost structure and the technical innovations that enable it, making a significant contribution to public understanding of AI economics. It connects the open-sourced code to observable performance data, offering a tangible framework for analyzing the profitability of AI inference. The discussion of a potential ‘price collapse’ in AI inference is relevant and timely, given the massive data center investments underway.

Pour aller plus loin :

Mixture of experts — The architecture behind DeepSeek’s efficiency; explains how specialized sub-networks reduce compute.
Transformer (deep learning) — Foundational architecture that includes attention and KV cache concepts.
vLLM blog on DeepSeek V4.1 Flash — Independent verification of DeepSeek’s performance claims and algorithmic details.
DeepSeek DeepInfer GitHub — The actual open-source code for the inference engine, useful for further study.

135 words

Radar Profile

The profile shows a relatively balanced video with high scores in information quantity and technical depth, while quality and reliability are solid but not perfect. The low 'moment' count reflects the chaptered structure. Overall, this is a dense, informative, and trustworthy analysis.

Reliability 8/10