1. Microsoft and Wall Street are converging on the same question: after hundreds of billions invested in AI, how much infrastructure is actually working?

An investigation published on August 17 by The Guardian, based on internal Microsoft documents, points to an apparent gap between the scale of AI infrastructure announced by the company and the number of AI accelerators actually operating. The documents reviewed by the publication indicate approximately 2.2 million AI chips installed, while estimates based on Microsoft’s announced power capacity would suggest a significantly larger number if all of that infrastructure were already operational. Microsoft disputes the publication’s calculations and says they rely on incorrect assumptions. The real constraint, however, may not be chip availability. Satya Nadella has previously explained that the challenge is finding sufficient electricity and data-center facilities ready to connect the accelerators. The investigation comes just as Reuters reports a shift in sentiment on Wall Street: Microsoft and Amazon’s results have eased some concerns about massive AI spending, and investors are now trying to identify which companies will be able to turn those investments into profits. Reuters estimates that hyperscalers — very large cloud and digital infrastructure companies such as Microsoft, Amazon and Google that operate data centers at global scale — could generate around $340 billion more in annual operating cash flow in 2027 than in 2025, while their capital expenditure is expected to increase by roughly $534 billion.

Why it matters: these two developments should be viewed together. The AI challenge is no longer simply “who can buy the most GPUs?”, but who can build data centers quickly enough, secure the required electricity, install the chips, and turn computing power into services customers are willing to pay for. Reuters notes that a data center may require 12 to 18 months to move from construction to revenue generation, while some investors believe hyperscalers such as Microsoft, Amazon and Google hold an advantage precisely because they control both the infrastructure and the customer relationship. This is the logical continuation of a theme discussed in AIdapted’s August 16 edition, where we looked at the effect of massive AI financing on financial markets: the next stage is no longer only about how much capital can be raised, but how much productive infrastructure that capital can actually create.

2. SafePal confirms a data breach affecting nearly 40,000 customers

SafePal, a cryptocurrency wallet provider, announced on August 16 that approximately 39,798 customers were affected by unauthorized access to information associated with their orders. According to the company and information reported by Reuters, the exposed data includes names, addresses and purchase information. The incident does not mean that the crypto wallets of nearly 40,000 people were emptied or that their private keys were compromised; the breach concerns personal data and commercial information connected to customer orders.

Why it matters: databases of this kind can be extremely valuable to attackers even when they do not directly contain wallet keys. Linking a person’s identity, address and the fact that they own crypto-related products can enable much more convincing phishing campaigns and highly targeted attacks. At a time when AI can dramatically reduce the cost of personalizing fraudulent messages, a simple data leak can become raw material for far more sophisticated attacks.

3. Anthropic’s potential IPO is being built around a forecast of $190–200 billion in revenue by 2028

Anthropic expects revenue of approximately $190–200 billion in 2028, according to Reuters, citing people familiar with the company’s finances. The forecast is relevant to discussions about the valuation of a possible IPO — an initial public offering, meaning the first time a company lists its shares on a stock exchange — because bankers and investors must decide what a company is worth today when its infrastructure and development costs are enormous but its growth over the next two years is expected to be exceptional. Reuters reports a current annualized revenue run rate of around $47 billion, which means the 2028 scenario still assumes roughly a fourfold increase. Anthropic has not commented on the report.

Why it matters: an Anthropic IPO would directly test whether public markets are willing to support AI valuations based to an unusually large extent on future expectations. The question is not only whether Anthropic can continue to grow, but whether the entire economy being built around frontier models can generate enough revenue to justify the investment in data centers, energy and chips. The company is also becoming an interesting experiment in what happens when AI agents move beyond chatbots and become systems capable of acting more autonomously — a topic recently examined by AIdapted in its article on the “territorial war” between Claude agents.

4. Restrictions on children’s access to social media are becoming a global trend

Reuters summarized on August 14 the growing number of measures through which governments are attempting to limit minors’ access to social media. Australia has gone the furthest, with a ban for children under 16, while several European countries are preparing or testing their own age-verification and access-restriction systems. France, however, encountered an important legal obstacle: on August 14, the Constitutional Council blocked legislation that would have banned children under 15 from using social media, ruling that the proposed mechanism raised concerns related to freedom of expression and privacy. At EU level, political pressure remains strong for common rules governing children’s access to online platforms.

Why it matters: the debate is no longer only about how much time children spend on TikTok or Instagram. Age verification requires identifying users, which can itself come into conflict with the privacy protections that regulation is trying to strengthen. At the same time, the boundary between social media and AI is becoming increasingly blurred: algorithmic recommendations, generated content and chatbots are being integrated into the same services. Europe is therefore trying to solve two difficult problems at once — how to protect minors from highly persuasive algorithmic systems, and how to do so without creating a permanent identification infrastructure for the entire internet.

5. Microsoft’s infrastructure problem is also a warning for Europe’s ambitions to build its own AI capacity

The Microsoft investigation highlights an essential distinction between announced capacity and computing capacity that is actually available for AI workloads. Building a data center does not automatically mean it can be powered, equipped with accelerators and connected quickly enough to begin serving AI models. The Guardian reports that Microsoft has announced a massive expansion of its power capacity, but some projects may take years to become fully operational. The company itself has pointed to electricity availability and facilities ready to host the equipment as major bottlenecks.

Why it matters for Europe: the continent is increasingly talking about AI sovereignty and building its own computing capacity, but Microsoft’s experience shows how wide the gap can be between a figure expressed in gigawatts and the computing power actually available to customers. AIdapted recently examined this exact bet in its article on Mistral’s plan to reach 1 GW of European computing capacity by 2030. Mistral currently operates around 77 MW and aims to reach 200 MW by the end of 2027 before pursuing the 1 GW target. The lesson Microsoft now offers is equally relevant for Europe: AI sovereignty is not built with models and chips alone, but also with electricity, power grids, land, cooling systems, permits and years of industrial execution.

Sources