Chinese Robots Beat Human Records in Running

In Beijing, a humanoid robot developed by Chinese company X-Humanoid ran 100 meters in 9.39 seconds, beating the world record of 9.58 seconds set by Usain Bolt in 2009. Another robot achieved a high jump of 2.88 meters, surpassing the human record of 2.45 meters. The performances were recorded at the World Humanoid Robot Games, now in its second edition, with more than 2,000 robots competing in 51 events, ranging from athletics and football to table tennis. These results follow the demonstrations in Beijing that AIdapted reported on August 20, when China’s robotics industry was talking about the approach of a “ChatGPT moment” for humanoid robots. AIdapted – China and the “ChatGPT moment” for robots

Why it matters: The record does not mean that robots are already better than humans at real-world tasks. Running on a specially prepared track is very different from moving around a factory, handling objects, or working in an unpredictable environment. What matters, however, is the pace of progress: the record-setting robot shows that mechanics, balance control, and autonomous software can evolve together, while China is trying to turn these demonstrations into an industrial advantage.

Cyberattack Shuts Down British Power Plant for Four Days

A small power plant in the United Kingdom was shut down for approximately four days following a cyberattack attributed by the British press to hackers with links to Iran. The incident took place in July, but the information became public on August 23. The British government said the facility was small, that there were no interruptions for consumers, and that the attack did not pose a threat to the operation of the national grid. The National Cyber Security Centre is continuing to assess the incident, while authorities have discussed additional protection measures with energy-sector operators.

Why it matters: This case is significant because the cyberattack had a concrete physical effect: an energy facility was taken offline. This was not simply about compromising a database or shutting down a website, but about interfering with infrastructure that produces electricity. The case should also be viewed in the context of warnings about AI agents capable of operating autonomously on the internet; AIdapted has already reported on a test in which an AI agent attempted to introduce malware into a real-world project. AIdapted – AI Agents Created Fake Identities and Tried to Convince a Developer to Accept Malicious Code

Alibaba Raises $10.2 Billion for the AI Race

Alibaba announced on August 23 an operation to sell 710 million new shares in Hong Kong, for approximately HK$80 billion, equivalent to $10.2 billion. The company said all of the capital raised will be directed toward developing its AI capabilities, including “full-stack” capabilities covering chips, computing infrastructure, and AI models. It is the largest primary share offering ever conducted by a company listed in Hong Kong. The offering was oversubscribed, prompting Alibaba to increase the initial size of the deal.

Why it matters: Alibaba is taking on very high costs to build its own AI ecosystem at a time when China’s access to the most advanced Western chips remains a strategic issue. The company recently reported a 75% decline in quarterly profit, largely as a result of rising AI-related spending. The message, however, is clear: in the AI race, Chinese companies are willing to sacrifice short-term profits for infrastructure, models, and computing capacity that could generate revenue in the years ahead.

Nvidia Warns That AI Servers Could Become More Than 15% More Expensive

Some of Nvidia’s largest customers have been informed that prices for servers equipped with its AI chips could rise by more than 15% for many configurations, starting with systems delivered in early 2027. The information was first reported by Bloomberg and subsequently reported by Reuters; the increases will vary depending on the chip generation and the amount of memory used. Systems based on the Vera Rubin and Grace Blackwell generations are among those affected, while server manufacturers working with operators such as Microsoft, Google, and Oracle have already informed customers. Nvidia has not officially confirmed the report.

Why it matters: The issue shows that AI infrastructure is beginning to hit a less visible bottleneck than processors: memory. Modern AI chips use very large amounts of HBM (High Bandwidth Memory – extremely fast memory designed to feed AI accelerators), and demand for this component has risen sharply. If memory becomes more expensive, the effect reaches server prices and, ultimately, the cost of building and operating AI models.

Nvidia Invests Billions in Poolside to Build an American Rival to Chinese Models

Nvidia has reached an agreement worth approximately $6 billion with U.S. startup Poolside to license the company’s technology and develop an open-weight AI model (a model whose weights are made available so developers can use and adapt it). Nvidia will separately invest approximately $1 billion in Poolside and plans to bring on more than 100 of the startup’s employees to strengthen its Nemotron project. The goal is to develop a powerful American model capable of competing with Chinese open-weight models such as DeepSeek and Kimi.

Why it matters: Nvidia is visibly moving beyond its role as a chip supplier. The company already controls a critical part of the infrastructure on which AI runs, and through Poolside it is seeking more direct influence over the models that generate demand for that infrastructure. It is also a response to the competitive shift coming from China: open-weight models can be distributed, adapted, and run more easily than fully closed systems, which can accelerate their adoption.

Nvidia Discusses Investment in Perplexity at a Valuation of More Than $30 Billion

Nvidia is discussing a potential multibillion-dollar investment in Perplexity in a funding round that could value the startup at more than $30 billion, according to The Information, as reported by Reuters. That valuation would be more than 50% higher than the $20 billion valuation established in its previous funding round. Perplexity has reached annualized revenue of more than $750 million, up from less than $250 million at the beginning of the year, with a significant portion of the growth attributed to Perplexity Computer, an AI agent designed to automate professional tasks.

Why it matters: Perplexity occupies a strategic position between a search engine and an AI agent. Users no longer receive only a list of results, but a synthesized answer and, increasingly, the ability to delegate tasks to the system. For Nvidia, investing in such a company would extend its influence from the “hardware on which AI runs” to the application through which users interact with AI. It is further evidence that the boundaries between infrastructure, models, and AI products are beginning to disappear.

Anthropic Recruits Google TPU Chip Veteran

Anthropic has recruited Amir Salek, the former head of Google’s custom chip organization, for its compute team, according to reports published on August 21. Salek spent nearly a decade at Google and was involved in the development and expansion of Tensor Processing Units (TPUs), the specialized chips Google built for AI workloads. The recruitment comes as Anthropic develops its own hardware design team, although the company continues to use processors from Nvidia, Google, Amazon, and AMD. The recruitment has not been publicly confirmed by Anthropic and should therefore be treated as a media report rather than an official announcement.

Why it matters: If confirmed, the recruitment shows how important control over hardware has become for frontier AI laboratories. Anthropic spends enormous sums training and running its Claude models, while chip availability and computing costs could become constraints just as important as model performance. A chip designed specifically for its own models could eventually give the company greater control over costs and performance. The AI race is therefore becoming increasingly vertical: model, software, chip, memory, and infrastructure.

Velatir Raises €5 Million to Control AI Use in Companies

Danish startup Velatir, based in Odense, has raised €5 million in a seed round led by Spintop Ventures and Ugly Duckling Ventures, just six months after its pre-seed round. The company is developing an infrastructure layer that allows organizations to see which AI tools employees and agents are using, what data is being sent to these services, how much the usage costs, and what rules can be applied. The platform includes a catalog of more than 4,000 AI tools and can enforce security policies and usage limits. Velatir says its platform is built on infrastructure that it owns and hosts in Europe, without relying on U.S. hyperscalers.

Why it matters: Companies are discovering a new problem: employees are adopting AI much faster than organizations can establish rules governing its use. The phenomenon is known as shadow AI — the use of AI tools that have not been officially approved or monitored by a company. For a business, the issue is not just security; confidential data, costs, regulatory compliance, and the ability to determine which processes are already being automated are also at stake. Velatir is interesting precisely because it turns this problem into a new commercial category: a kind of “administrative control layer” for an organization’s entire AI activity.

Sources