O Que É LLM

On August 10, cybersecurity firm Genians disclosed that Kimsuky, a hacker group under North Korea's Reconnaissance General Bureau, has made significant upgrades in generative AI applications—not only using AI for phishing lure creation, but also building an independently operated local large language model (LLM) environment and retrieval-augmented generation (RAG) system. Specific tools include local LLM management tools such as Ollama, GPT4All, Msty, as well as AI code editor Cursor, with the intent to automate intelligence extraction and attack processes. (Yonhap)

According to monitoring by Beating, macro strategist Andreas Steno Larsen recently issued a warning about the latest downward trend of the Silicon Data LLM Token Spending Index. He pointed out that changes in LLM Token pricing are a key indicator determining the entire AI infrastructure cycle. If Token pricing continues to weaken, this investment cycle may come to an end, affecting everything from memory chips to broader hardware and data center construction. The Silicon Data LLM Token Spending Index is a market indicator weighted by spending, used to track the average payment price per million Tokens across the market. Since mainstream service providers generally charge based on Token consumption, this data directly links the actual usage on the software side with the underlying GPU computing power, DRAM memory, and data center construction, making it an important reference for assessing the AI hardware capital expenditure cycle. Data shows that after a significant rise at the beginning of 2026, the Silicon Data LLM Token Index peaked and then fell back by the end of May. This decline in the index is closely related to the change in corporate attitudes towards AI spending. Due to a lack of cost control in the early stages, some tech giants even experienced a blind competition known as 'Tokenmaxxing,' where employees consumed Tokens excessively and meaninglessly to boost internal productivity rankings. With the arrival of hefty bills, companies realized that high Token spending did not yield proportional business returns. Currently, giants like Microsoft and Amazon have begun to tighten internal usage limits on AI tools and even halt related projects. The shift from 'blind accumulation' to 'financial rationality' in enterprises has directly led to the stagnation and weakening of the Token Spending Index.

On June 4, Aethir officially announced the launch of Aethir Mesh, a self-developed open source large model API platform operating on Aethir's decentralized GPU infrastructure. The platform currently offers a unified API endpoint that directly connects to mainstream open source models such as DeepSeek V4, Kimi K2.6, GLM-5.1, MiniMax-M2.5, and Qwen3.6-27B. Unlike aggregation routing solutions like OpenRouter, Aethir Mesh provides inference services directly from Aethir's own computing power, with no resale premiums or third-party routing. Aethir Claw users can directly integrate the Mesh API Key into their agent configurations, achieving a full-stack closed loop from VPS hosting to model inference, with inference data remaining entirely within the Aethir ecosystem. The platform is now live: https://agent.aethir.com/twitterCNmesh