AI Infrastructure Is Exploding: AMD, OpenAI, and Anthropic
AMD, OpenAI, and Anthropic are pushing AI infrastructure to gigawatt scale, while new security vulnerabilities in AI agents are increasing risk in everyday enterprise use.
Inhaltsverzeichnis
Today is not about a new chatbot feature, but about the hard reality behind the AI wave: data centers, chips, power, and contracts worth billions. Anyone who wants to understand where the market is heading needs to look right here – because without infrastructure, all you get in the end are pretty demo videos and very warm rooms.
At the same time, a new security finding shows: the more autonomous AI agents become, the more carefully you need to watch permissions, approvals, and attack paths. In short: AI is getting bigger, faster, and more expensive. And it is not automatically becoming friendlier.
💸 AMD invests up to 5 billion in Anthropic
AMD is making a massive move in the AI race: According to The Decoder, the chip maker plans to invest up to 5 billion US dollars in Anthropic – and in return supply up to 2 gigawatts of AMD MI450 GPUs for training and operating the Claude models. This is not just a deal, it is a statement to Nvidia. With this move, AMD is trying to position itself as a serious alternative in the market for AI chips and inference hardware.
Why this matters: AI is no longer decided solely by models, but by supply chains for compute. Whoever can secure GPU capacity gains influence over the next model generation. At the same time, the question remains how “real” such circular deals are when money and chip orders flow in both directions. Still: for AMD, this is a strategic lever – and for Anthropic, more room for scaling and availability. Original source: The Decoder, additionally reported by The Verge.
🏗️ OpenAI plans a 3.2-GW data center in Georgia
OpenAI is taking infrastructure one step further: with the project “Camellia,” the company is planning a data center in Georgia and has secured a power contract for 3.2 gigawatts, according to The Decoder. For context: this is not a “big server room,” but an energy project on a state-wide scale. On top of that, there are 80 million dollars for the community and 71 million dollars in Codex credits for students – a pretty obvious attempt to secure local acceptance.
Why this is important: the new AI era is increasingly becoming an energy question. Models, training, and inference require not only GPUs, but also grids, water, land, and permits. OpenAI is showing how AI companies will present themselves in the future: as technology providers, but also as infrastructure players with enormous influence over regions. For the industry, this is a signal that “scale” once again outweighs everything else. For municipalities, it is a deal between promise, electricity prices, and patience. Original source: The Decoder.
🧠 Black Forest Labs introduces Flux 3 for image, video, audio, and robotics
With Flux 3, Black Forest Labs is presenting a multimodal foundation model that processes image, video, and audio together and can even generate videos with native sound. According to the company, the model is already said to be on par with or even outperform strong competitors such as Seedance 2.0 in internal tests. That is exciting because BFL is moving beyond classic image generation and clearly thinking in the direction of a “world model.”
Why this matters: multimodal models are the next logical step when AI should not only understand text, but also scenes, motion, sound, and eventually perhaps physical environments. Especially interesting is the connection to robotics: together with mimic robotics, Flux 3 is already being used for such applications. For you, this means: the boundary between content AI and control AI continues to blur. It is not a finished all-rounder yet, but it is a clear sign of where things are heading. Original source: The Decoder.
🔐 Manipulated ChatGPT link can install AI agents in a corporate network
A particularly unpleasant finding comes from Zenity Labs: According to The Decoder, for “AgentForger” a manipulated ChatGPT link was enough to create an autonomous agent in the name of an employee via OpenAI’s Agent Builder. The agent inherited the victim’s identity and access rights and regularly fetched new instructions from the attacker’s mailbox. This is exactly the kind of attack that really hurts in agentic AI: not spectacularly loud, but quiet, clean, and with legitimate permissions.
Why this matters: as soon as AI agents are allowed to carry out actions independently, the security model shifts. Then the question is no longer only, “Is the model secure?” but also: Who is allowed to create what, which approvals apply, how are identities inherited, and how are external inputs validated? For companies, this is a clear reminder not to treat agents as nice productivity helpers, but as potentially privileged software with an attack surface. Original source: The Decoder.
⚠️ AWS infrastructure in Bahrain reportedly targeted in an attack
The geopolitical situation is hitting the cloud directly: heise online reports that Iranian attacks allegedly hit and destroyed AWS infrastructure in Bahrain. Regardless of the exact extent of the damage, the incident shows how vulnerable digital infrastructure is in conflict zones. The cloud is not just “somewhere on the internet,” but depends on very physical places with very physical risks.
Why this matters: for companies with international infrastructure architectures, resilience is no longer a theoretical design question, but an operational necessity. Redundancy across regions, emergency plans, and clear fallback strategies are becoming more important, especially when data centers become part of geopolitical tensions. The cloud is powerful – but it does not stand outside world history. Original source: heise online.
🛠️ Tool tip of the day: monitor AI and cloud workloads better
If you run AI models, inference workloads, or cloud infrastructure in production, you need more than pretty dashboards. A good observability tool helps you detect GPU utilization, latencies, costs, and outages early – before a model turns into a very expensive mystery. This is especially useful in multi-cloud setups and agent systems with many dependencies. #
Don’t want to miss any news? Subscribe to the newsletter