Today’s AI day has several major threads at once: more data centers, more chip deals, more infrastructure. In short: the AI world is growing not just in models, but above all in concrete, copper, and megawatts. And while the industry funds its next expansion phase, it is also arguing about security and copyright on the side — business as usual, just with more GPUs.

🚨 AWS infrastructure in Bahrain allegedly destroyed

Probably the harshest news of the day is not a product announcement, but a geopolitical alarm: according to heise, Iranian violence as part of the escalation in the Middle East has also hit AWS infrastructure in Bahrain. If this is confirmed, it would be another example of how physical conflicts directly affect the digital world. Cloud is not “somewhere up there” after all; it runs in data centers, with electricity, network connectivity, and a very real vulnerability.

This matters for companies because incidents like this brutally expose the dependence on individual regions and cloud locations. Anyone operating critical systems should not treat redundancy, multi-region architectures, and disaster recovery plans as a luxury. The sober lesson: digital resilience is now also geopolitics. Or a little less dry: if you store all your data in one basket, don’t be surprised when the basket breaks.
Source: heise

🏗️ OpenAI plans 3.2 GW data center in Georgia

OpenAI is no longer holding back on infrastructure: under the codename “Project Camellia,” the company is planning a gigantic data center in Georgia and has signed a power agreement with Georgia Power for 3.2 gigawatts. This is not just a server room with decent ventilation, but a scale that resembles power plant logic. In addition, there is said to be $80 million for the community and $71 million in Codex credits for students.

Why does this matter? Because it shows how much AI now depends on energy, grids, and political acceptance. The time when you could train models without talking about electricity prices, land use, and local resistance is over. OpenAI is visibly trying to think about infrastructure and site policy together here. That is strategically smart too: anyone who wants to build the next generation of AI needs not only ideas, but above all electricity, permits, and reliable partners.
Source: The Decoder

⚙️ AMD invests up to $5 billion in Anthropic

AMD is moving even deeper into the AI business and wants to invest up to $5 billion in Anthropic. In return, Anthropic is expected to use up to 2 gigawatts of AMD Instinct MI450 GPUs, including the new Helios rack system. This is a deal that works on two levels at once: capital on the one side, secured demand for AI chips on the other. That is exactly how the major alliances in today’s AI industry are formed.

This is especially exciting for the market because AMD is positioning itself directly against Nvidia — not only technically, but also through partnerships. At the same time, one critical point remains: deals like these often look huge on paper, but they can also contain circular money flows. Still, the direction is clear: if you want to keep up in the AI race, you need not only good chips, but long-term ecosystem deals. And that is what the entire infrastructure battle is driven by right now more than by any single model announcement.
Source: The Verge | The Decoder

🧠 Anthropic, AMD and the gigawatt reality

The second report on the AMD-Anthropic deal makes even clearer what this is really about: Anthropic apparently wants to run its Claude models on AMD hardware at scale, and on a gigawatt level. This shows how competition in LLMs is shifting. It is no longer just about “which model is smarter?”, but about “who gets enough compute to keep scaling at all?”

For users, that means the next wave of better AI systems depends more on supply chains, chip availability, and data center capacity than on PR campaigns. For the market, it means infrastructure is the new battleground. And if you thought AI was mainly software — well, the electricity bill has other plans.
Source: The Decoder

🛡️ Cisco brings small open models for security

Cisco is releasing two small open AI models designed to identify vulnerabilities in code. According to its own tests, they find around 150 times more vulnerabilities per dollar than large AI agents, at a fraction of the cost. That is a strong signal for anyone who does not want to rely on the biggest possible models for security, but instead on efficiency and focused tasks.

In the security field in particular, small specialized models are often more practical than all-purpose giants. They are easier to inspect, cheaper to run, and easier to integrate into existing development workflows. That matters for teams with tight budgets, because it makes security analysis more accessible. At the same time, of course, a model is only as good as its integration into review and patching processes. AI finds the gaps — you still have to fix them yourself.
Source: The Decoder

⚖️ Anthropic pays record settlement over book piracy

The legal side remains exciting too: Anthropic is paying $1.5 billion in a settlement with book authors — the largest copyright settlement in a class action so far. But the key point is the context: the allegation concerns downloading around 482,460 works from piracy databases, not AI training itself. That is exactly what makes the case so interesting legally, because a judge had previously assessed the training itself as “spectacularly transformative” and covered by fair use.

For the industry, this sends an ambivalent signal. On the one hand, it is a massive financial blow and a warning shot on data sourcing. On the other hand, the ruling seems to strengthen the position of AI labs when it comes to the actual training process. In short: not everything that goes wrong in the AI ecosystem is automatically a ban on models. But the origin of training data remains a very expensive issue.
Source: The Decoder

🛠️ Tool tip of the day

If you want to test AI-based security workflows, Cisco’s new open models are an exciting starting point. Especially for code scans and prioritizing potential vulnerabilities, such a specialized model can be more useful than a large generalist that tries to do everything. This is particularly interesting for DevSecOps teams that need to keep an eye on cost and speed.

---

Don’t want to miss any news? Subscribe to the newsletter