According to issue 60 of the Webrazzi newsletter, published on August 28, 2026, AI Week saw simultaneous developments in products, models, cybersecurity, and investments. While companies such as Perplexity, Google, Alibaba, and Z.ai announced new tools and models, research data showed that China-linked hacker groups more than doubled the number of their attacks after integrating open-source models such as DeepSeek into their operations.
Models and Tools Move Toward Practical Use
Perplexity and Nvidia announced Portable Computer, a device that enables AI models and agent workflows to run on users’ devices without relying on the cloud. Google Cloud introduced two specialized solutions: Gemini Enterprise for Legal for contract-review work, legal research, and regulatory monitoring, and Gemini Enterprise for Financial Services for financial-sector research, KYC procedures, portfolio analysis, and bond issuance.
Other developments included Anthropic’s unification of Claude’s memory across its Chat and Cowork environments, and Google’s addition of new voice features to Gemini Live, including the Spark tool, which can perform independent tasks in the background and control Workspace applications hands-free. Meanwhile, OpenAI restored the five-hour usage limit for ChatGPT Plus subscribers using Codex and ChatGPT Work, and enabled stickers generated through ChatGPT Images to be transferred to iMessage and WhatsApp.
At the model level, Alibaba unveiled Wan3.0, capable of producing videos at 1080p resolution and up to 30 seconds long, while converting text, images, audio, PDF files, and presentations into video. Z.ai also launched the open-source GLM-5.3-Flash model after confirming that it was behind the model previously known as Ox Alpha, while Yutori introduced the Navigator n2 model for computer use through clicking, terminal commands, and programming. Google also announced Gemini Omni 1.1 Flash, with video extension, frame interpolation, and 4K upscaling features.
Cybersecurity Tests the Limits of Open Models
Based on findings from security research company TeamT5, the article reported that Chinese state-linked groups favored DeepSeek because of its lower cost and less stringent security restrictions, while researchers have so far found no evidence that more expensive models such as Kimi K3 were used in attacks. This coincides with a warning from the AI Security Institute that models’ capabilities to carry out attacks are roughly doubling every few months.
More than 100 companies, including OpenAI, Anthropic, Google, and Microsoft, called on the public and private sectors to coordinate their response to AI-supported attacks. OpenAI also published a report on a security incident that enabled its models to access Hugging Face systems, explaining how the incident occurred and the measures taken to prevent its recurrence.
AGI Between Expectations and Institutional Constraints
Sam Altman said that a model matching his definition of artificial general intelligence may exist within OpenAI by the end of 2026. Mark Chen said the company had completed 80% of the path toward AGI, while Greg Brockman said future generations may view the current period as the phase in which AGI emerged. Jakub Pachocki was also quoted as saying that the Astra model can perform the equivalent of roughly one week of researchers’ work after being provided with an academic paper.
These statements remain predictions and definitions issued by OpenAI officials, rather than an independent measurement of whether AGI has been achieved. In a different context, Bill Gates said that the world lacks a clear plan for the transformation led by AI, and suggested exploring taxes on AI tokens and robots, as well as creating jobs reserved for humans.
Investment and Computing Advance to the Forefront of the Race
Infrastructure emerged as a parallel focus to model development. An agreement between Anthropic and Nscale worth approximately $45 billion was announced to provide computing capacity, while Nvidia recorded revenue of $96.2 billion in the second quarter of its fiscal year 2027, including $89 billion from data centers, representing a year-over-year increase of 117%.
Nvidia was reported to have agreed to acquire Hugging Face for $12.9 billion, but the two companies had not officially confirmed the agreement as of the date of the article. Perplexity is in talks to obtain an investment from Nvidia in a round that could value the company at up to $30 billion, with expected annual revenue exceeding $750 million. Nvidia also invested in Cloverleaf Infrastructure, while SpaceX plans to build space-based data centers called Starmind around Nvidia’s Vera Rubin NVL72 system, with the goal of launching a lighter version into orbit in the fourth quarter of next year.
These developments show that competition is not focused solely on model quality, but also on where models run, the cost of computing, and their ability to integrate into specialized work. At the same time, unconfirmed acquisition deals, AGI predictions, and methods of using models in attacks remain areas requiring independent verification before they are treated as definitive facts.