AI boom on credit: billions, power, and safety brakes
OpenAI, Anthropic, and Nvidia are scaling up AI infrastructure, but power, safety, and regulation are becoming the drag.
Inhaltsverzeichnis
Today it becomes pretty clear what the current AI phase really is: no longer a pure software game, but an infrastructure and power struggle. Between gigantic data centers, new safety brakes, and growing regulation, one thing is becoming clear: whoever controls the chips, the power, and the distribution controls the market.
And yes, the sums now sound more like public finances than startups. But that is exactly why it is worth taking a closer look: the decisions made today will determine how expensive, fast, and safe AI will be tomorrow.
🧮 Nvidia backs OpenAI’s 8-GW data center
OpenAI has signed a 20-year lease in Ohio for an 8-gigawatt data center — a scale at which you no longer really talk about “cloud,” but rather about an “energy project with a chatbot attached.” According to The Decoder, Nvidia is guaranteeing up to $105 billion for the residual value of the facilities and will also be the exclusive chip supplier.
Why this matters: it shows how tightly AI, semiconductors, and infrastructure are now intertwined. Nvidia is no longer just selling GPUs, but is indirectly backing entire data-center deals. For OpenAI, that means planning security; for the market, it also means greater dependence on a single chip giant. And when a data center on the scale of a small power grid is being planned, it is clear: the AI boom is long past being just a software feature and has become a macroeconomic project.
🧠 Anthropic keeps growing: $65 billion annualized
Anthropic reports an annualized revenue rate of more than $65 billion, according to The Decoder — after a sevenfold increase within a year. That is no longer normal startup growth, but a pace at which even Excel eventually gets nervous.
The number is especially interesting because it shows how strongly demand for frontier models and enterprise AI is currently pulling. Anthropic is benefiting in particular from Claude users in developer and corporate settings, where AI is actually being embedded into processes. At the same time, rumors are already circulating about a possible IPO in fall 2026. For you, that means the competition with OpenAI is becoming financially harsher — and the question is no longer whether AI giants will emerge, but who can afford the most expensive seat in the race.
🛑 OpenAI slows model training because of safety risks
OpenAI has temporarily halted training of its most powerful models after the upcoming model “Astra” apparently showed threatening capabilities in the area of cyberattacks. According to The Decoder, a new monitoring system has also been introduced that raises an alarm within 30 minutes if suspicious model behavior is detected.
This is notable because it shows that safety questions are no longer being discussed only after the fact, but are being pulled directly into the development process. Especially awkward: the safety team responsible for catastrophe risks is said to have been dissolved. It feels a bit like, “We’re installing an alarm system after the smoke detector has burned down.” For the industry, this is a signal: cybersecurity and model control are becoming mandatory, not a PR add-on.
🔥 AI could drive new gas power plants
The data-hungry boom also has a pretty unromantic downside: energy. According to heise online, CO2 emissions from U.S. power generation could rise by as much as one third if the planned 99 gas power plants for AI data centers are actually built.
For the AI market, this is not a side issue but a central bottleneck. No power means no data centers, no data centers means no models, no models means no further growth. At the same time, political pressure is increasing: energy policy, grid stability, and climate targets are colliding directly with AI expansion. Anyone planning AI infrastructure therefore has to talk not only about GPUs, but also about turbines, permits, and emissions. It sounds less sexy, but in the end it is the part where projects fail or scale.
🛠️ Tool tip of the day: Claude Code with /design
Anthropic is expanding Claude Code with the /design command: developers can generate visual UI drafts as artboards directly in the terminal before even writing the first line of frontend code.
This is especially useful if you already have an existing codebase and do not want to start from scratch. Claude analyzes the existing style and adapts suggestions accordingly — so less “generic AI layout,” more “fits your project.” For product teams, that can significantly speed up the jump from idea to reviewable design. If you often bounce between terminal, Figma, and half-finished mockups, this is a useful workflow helper. #
⚖️ U.S. Department of Justice investigates Andreessen Horowitz
The U.S. Department of Justice is investigating Andreessen Horowitz over possible market distortion in the AI market. The allegation: partners at the VC firm sit simultaneously on boards of competing data companies such as Databricks and Fivetran. There is also the political dimension: a16z has actively campaigned for AI deregulation and has strong ties in Trump circles.
Why does this matter? Because the AI market is decided not only by models and compute, but also by equity stakes, supervisory boards, and networks. This is exactly where antitrust issues become tricky. If an investor sits in multiple overlapping ecosystems, the question quickly arises: is this smart capital access or already market distortion? For the industry, the investigation could become a precedent — especially now, as capital and control are moving closer together.
📈 Trillion-dollar market with question marks: OpenAI, Nvidia, and others
According to the report, the Wall Street Journal paints a picture of roughly three trillion dollars in AI commitments now on the books of nine tech giants — though often more as future promises than as hard balance-sheet items. In combination with Nvidia’s OpenAI guarantee and Anthropic’s growth figures, it becomes clear how aggressively the industry is betting on expansion.
For you as a reader, this means: we are currently seeing a market built on enormous upfront investment. Data centers, chips, power contracts, and model training are being financed today in the hope that AI will be monetized even more strongly tomorrow. That can work — or become very expensive if demand, regulation, or energy prices do not play along. That is exactly why every billion-dollar announcement also raises the question: who ends up carrying the risk?
Don’t want to miss any news? Subscribe to the newsletter