AI News- 14th August

Australia discovers how difficult it is to keep children off social media

Meta, Google/YouTube, TikTok and Snapchat have been questioned by an Australian Senate committee amid growing evidence that the country’s ban on social media access for users under 16 is proving much harder to enforce than lawmakers had hoped. Data reviewed by Australia’s eSafety regulator indicates that more than 80% of teenagers who used social networks before the ban continued using them afterwards, while the platforms argue that it is still too early to draw definitive conclusions and that age-verification technologies remain imperfect. Meta, for example, says it has disabled hundreds of thousands of accounts suspected of belonging to minors. (Reuters)

Why it matters: Australia is now providing the first real nationwide test of the idea that the harms associated with social media can be addressed mainly by controlling the age of users. This is precisely the contradiction examined in depth by AIdapted in “24 years after the Persuasive Technology revolution, the first major political response emerges”

: the law focuses on restricting who can enter the platforms without fundamentally changing the mechanisms — infinite scrolling, autoplay or algorithmic recommendations — that helped create the problem in the first place. Australia’s early results make this question increasingly relevant for Europe as well.

Nvidia and Wall Street are trying to build a financial system for AI infrastructure

Goldman Sachs is in discussions with investors, asset managers, banks and insurance companies over a structure through which Nvidia’s initiative could mobilize as much as $500 billion for AI infrastructure. Importantly, Nvidia itself is not investing $500 billion. The idea is to create financing platforms through which institutional capital could fund data centers and computing capacity for Nvidia customers, while Nvidia could guarantee up to 25% of certain projects. This directly continues the financing mechanism AIdapted wrote about on August 12

. (Reuters)

Why it matters: if the model works, AI compute could increasingly be treated as a financeable infrastructure asset, similar to energy, telecommunications or real estate. The industry’s bottleneck is no longer only who can manufacture the GPUs or build the best model, but who can finance the hundreds of billions of dollars required to put all that computing capacity into operation.

Apple built its own AI model for China, with Alibaba’s support

Apple has developed a language model specifically for the Chinese market with support from Alibaba, according to sources cited by Reuters. This represents an important shift from a strategy based exclusively on models supplied by local partners. Apple is attempting to gain greater control over the AI experience on its own devices while still building an architecture that complies with Beijing’s regulatory requirements. The company is also facing mounting pressure in China from domestic competitors such as Huawei. (Reuters)

Why it matters: we may be beginning to see a world in which major technology companies no longer operate a single AI system globally. Regulation, infrastructure and geopolitics could instead produce regional versions of the same product — and China is becoming the first major laboratory for this fragmentation.

Google launches Gemini 3.7 Flash and pushes the AI race toward cheaper agents

Google launched Gemini 3.7 Flash on August 13, a model designed specifically for programming, debugging, software problem-solving and agentic workflows. Through the end of the year, Google is offering it at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens, roughly half the initial price of Gemini 3.6 Flash. At the same time, Google has yet to provide details about its much-anticipated Pro model. (Reuters)

Why it matters: for AI agents that run continuously and execute hundreds or thousands of operations, the price per token can become almost as important as benchmark performance. Google is trying to compete not only on model intelligence, but on the combination of capability, speed and the cost of completing an autonomous task.

Databricks raises $5 billion and reaches a $190 billion valuation. Two of its co-founders have Romanian roots

Databricks announced on August 13 that it had raised $5 billion in new financing, at a company valuation of $190 billion. The two figures should not be confused: $5 billion is the amount of capital raised in the funding round, while $190 billion is the value assigned to the entire company in the transaction. Just six months earlier, its valuation was around $134 billion. Databricks says it has surpassed a $7 billion annualized revenue run rate, with growth exceeding 80% in the second quarter. (Reuters)

There is also an important Romanian connection. Databricks is not a Romanian company, but two of its seven co-founders are Ion Stoica, co-founder and Executive Chairman, and Matei Zaharia, co-founder and Chief Technologist, who created Apache Spark. Both are of Romanian origin. The company emerged from research conducted at UC Berkeley and has since become one of the world’s major providers of software infrastructure for data and AI. Why it matters: the $190 billion valuation shows just how much capital is flowing into the layer that manages enterprise data and AI applications, not only into the laboratories building frontier models. (Databricks)

DeepSeek moves from “very cheap AI” toward a premium model

DeepSeek has officially launched V4 Pro, which it says offers significantly improved capabilities for AI agents. According to figures cited by Reuters from Artificial Analysis, the model costs around $1.32 per million input tokens and $3.96 per million output tokens — roughly nine and fourteen times more, respectively, than V4 Flash under the prices compared by Reuters. The company has separately announced increases in API prices and the introduction of different rates for peak and off-peak hours. (Reuters)

Why it matters: DeepSeek built much of its global reputation on an exceptional performance-to-price ratio. The emergence of a premium model suggests that Chinese AI companies are also trying to turn frontier performance into a higher-margin business — potentially changing one of the Chinese AI ecosystem’s strongest competitive advantages.

Cl0p claims it stole data from nearly 50 companies

The Cl0p cybercrime group claims it extracted large amounts of data from nearly 50 organizations, including Philips, Shell, GE and Fiserv, in a campaign exploiting vulnerabilities in software used by numerous companies. Philips confirmed that an internal server had been compromised and isolated, while saying that customer environments were not affected. Shell is investigating a possible incident, while Fiserv said its analysis had found no evidence that customer data or operational systems were compromised. Reuters has not been able to independently verify the full extent of all the breaches claimed by Cl0p. (Reuters)

Why it matters: this is exactly the type of attack in which a single vulnerability can create dozens of victims. Attackers do not need to penetrate every company individually. They can identify a weakness in widely used software and turn that common dependency into a multiplier for the attack.

French taxpayers’ data stolen in cyberattack

France’s Finance Ministry has confirmed that a cyberattack resulted in the theft of data belonging to both individual and professional taxpayers. At the time of the announcement, authorities had not yet disclosed the full scale of the breach or the precise categories of information compromised. The incident adds to a growing series of attacks targeting databases operated by European public institutions. (Reuters)

This is not an isolated case. AIdapted recently examined the compromise of the UK Police National Legal Database and Liechtenstein’s beneficial ownership register

. Romania experienced its own major incident on July 14, 2026, when a ransomware attack against ANCPI disrupted the e-Terra platform, affecting cadastral operations, notarial authentication procedures and mortgage-related transactions. Why it matters: the danger is no longer limited to the temporary unavailability of public services. Stolen databases can be combined and analyzed with AI to rapidly identify individuals, relationships and high-value targets.

Pony.ai and Uber plan more than 2,000 robotaxis in Europe

Chinese autonomous-driving company Pony.ai and Uber announced on August 14 that they are expanding their European partnership, with the goal of deploying more than 2,000 robotaxis. The collaboration starts from their commercial service in Zagreb and is expected to expand to four additional European cities that have not yet been named. Pony.ai provides the autonomous-driving technology, while Uber supplies access to its network of users. (Reuters)

Why it matters: for Europe, this is no longer simply about testing a few autonomous vehicles. A fleet numbering in the thousands would turn autonomous driving into a visible commercial service and would simultaneously raise questions about safety, legal liability, jobs and dependence on Chinese technology.

The AI boom is starting to show up in the British economy

The UK economy grew by 0.4% in the second quarter, with the information and communications sector accounting for nearly half of that growth. Programming and IT consultancy expanded by 3.7%, while investment in computing equipment and digital infrastructure also increased as companies continued to build capacity for AI-related projects. (Reuters)

Why it matters: this is one of the most interesting macroeconomic signals so far. After years in which AI’s impact was measured mostly through investment announcements, valuations and benchmark scores, data is beginning to suggest that the technology boom may also be becoming visible in GDP, investment and economic output.

Philips and Shell illustrate the European dimension of the Cl0p campaign

In Europe, Philips and Shell are among the most prominent names linked to the campaign claimed by Cl0p. Philips says it isolated an internal server containing company data and that customer environments were unaffected, while Shell is investigating. The campaign appears to involve the exploitation of software vulnerabilities shared across multiple organizations, helping explain why so many potential victims emerged at the same time. (Reuters)

Why it matters: viewed alongside attacks on public-sector databases in France, the United Kingdom, Liechtenstein and Romania, the incident highlights two parallel fronts: compromises of government infrastructure and supply-chain attacks against companies. Data is becoming an even more valuable target because AI tools can make its analysis and exploitation dramatically faster.

Romania prepares the national enforcement framework for the EU Data Act

The public consultation on Romania’s draft legislation for implementing the EU Data Act ended on August 3, and an analysis published on August 13 outlines how the national enforcement system could work. The Data Act already applies directly across the European Union, but Romania still needs to establish the authorities responsible for supervision and the national sanctions framework. The draft designates ANCOM as the competent authority and Data Coordinator, while ANSPDCP would retain its responsibilities for personal-data protection. It defines 77 potential administrative offences, while fines for large companies could reach up to 5% of turnover and, for repeated infringements, up to 10%. These are currently proposed provisions, not penalties already in force. (Wolf Theiss)

Why it matters: the Data Act changes the relationship between manufacturers, users and the data generated by connected products — from vehicles and industrial equipment to IoT devices and cloud services. For companies operating in Romania, the European debate is therefore becoming very concrete: who will enforce the rules and how expensive could non-compliance become?