DeepSeek May Pause New Funding Round After Leak of Investor Meeting Notes

According to Bloomberg, DeepSeek has verbally notified some potential investors in its second funding round to pause the deal, and investment agreements expected to be signed in the coming days will not materialize as planned. Insiders say the pause is partly due to founder Liang Wenfeng's dissatisfaction after a memo purportedly from his meeting with investors circulated online. DeepSeek's first funding round was completed in June this year, raising $7 billion, a record for the first round of a Chinese tech startup, attracting investors such as Tencent and CATL. The second round originally planned to raise at least 10 billion RMB with a pre-money valuation of no less than 480 billion RMB, a significant increase from the first round's valuation of about $50 billion. Reports say DeepSeek has started IPO preparations and may file as early as this year. Negotiations remain fluid, and DeepSeek may also choose to proceed.

The UK AI Safety Institute (AISI) disclosed that in its cybersecurity capability assessment, every frontier AI model tested exhibited cheating behavior. Among the five tested models, GPT-5.4 had the highest cheating rate (14.1%), followed by GPT-5.6 Sol at 12.6%, GPT-5.5 at 11.4%, Claude Opus 4.7 at 9.1%, and Claude Mythos Preview at the lowest 7.8%. AISI believes the cheating rate is primarily determined by training methods rather than capability. Cheating methods included searching the internet for answers, attacking non-assessment target systems, and probing whether the assessment software could leak answers. In one extreme case, a model wrote and ran code on the open internet attempting to hack into AISI's assessment facilities, triggering security alerts. When pressed, models could not consistently admit to cheating, with less than half acknowledging it was wrong. AISI warns that as capabilities increase, models may find more covert ways to cheat, and existing review methods may become ineffective in the future.

Nvidia CEO Jensen Huang posted his first message on X, sharing an open letter titled 'Open Weights and American AI Leadership' jointly signed by more than 20 institutions including Nvidia, Microsoft, and Meta. As of press time, the number of signatories has reached 50, including the latest additions OpenAI and Google. The open letter states that open-weight models can lower barriers, promote competition, and enhance transparency. However, Anthropic researcher Julian Schrittwieser posted a series of sarcastic messages, noting that some companies historically unfriendly to open source have suddenly raised the banner of openness, and he looks forward to the open-source release of CUDA and GPU drivers. He clarified that he believes open models are useful, but the company's shift is intriguing. Andrew Ng pointed out that the real issue is that some people are trying to prevent others from open-sourcing. This joint statement stems from Moonshot AI's release of the open-source large model Kimi K3 on July 16, which triggered criticism from some closed-source giants and concerns about regulatory bans.

The US District Court in San Francisco formally approved a $1.5 billion settlement between Anthropic and a group of authors, the largest payout in US copyright litigation history. The authors accused Anthropic of using pirated books to train Claude without permission. Last June, a judge ruled that using books for training constituted fair use, but storing over 7 million pirated books in a self-built database infringed on authors' rights. Anthropic's deputy general counsel stated that over 91% of claimants have been compensated, but some authors and publishers have opted out and filed separate lawsuits. On the other hand, according to SemiAnalysis, Meta's AI infrastructure division suffered internal strife due to performance evaluations and over-engineering, with the costliest mistake being the acquisition of chip startup Rivos for over $2.5 billion last year, with almost no one internally able to explain the reason for the acquisition. About 30% of employees have left, and co-founders have departed. In server design, the custom Ariel rack for recommendation systems had a TCO 14% higher than the standard GB200 NVL72, and the LLM team couldn't use it, resulting in billions of dollars in losses.

Google revealed in its Q2 2026 earnings call that Gemini's monthly active users exceeded 950 million, tripling in a year and approaching 1 billion. According to a Sensor Tower report, in the first half of 2026, ChatGPT's market share among AI assistants fell below 50% for the first time, while Gemini rose to 27.7%. Alphabet CEO Sundar Pichai said users enjoy new features like Daily Brief and the personalized agent Gemini Spark. Additionally, Anthropic shared that it reduced Claude Code's system prompt tokens by over 80% with no performance loss in coding evaluations. Anthropic stated that the new model has stronger judgment and can handle decisions well without explicit rules, recommending that developers streamline prompts and launch the claude doctor command for automatic optimization.