AI without Nvidia, Anthropic in court, and Google’s new lab AI

Today makes one thing pretty clear again: AI is long past being just a model game — it’s an infrastructure, legal, and power game. Whoever controls chips, data centers, evaluation, and regulation is increasingly also determining how quickly the market moves.

And yes: while some spend billions on compute, others are trying to make benchmarks watertight. Welcome to the stage of the AI industry where even the side remarks in court rulings and infrastructure contracts suddenly become strategically relevant.

🚀 GLM-5.3-Flash: China’s open-source model without Nvidia

The Chinese company Z.ai has introduced GLM-5.3-Flash, an open-source model with 320 billion parameters that scores just three points behind the larger GLM-5.3 on the Artificial Analysis Intelligence Index — but at about one-seventh of the cost. The real headline isn’t just the price, but the hardware: all inference traffic ran on Chinese AI chips, not Nvidia. That’s more than a technical detail, because it shows that China is increasingly taking its own path in the AI stack. For companies, this is exciting because low inference costs directly determine whether a model stays experimental or becomes production-ready. For Nvidia, it’s a small but clearly audible warning sign. If you’re watching the market, you should pay attention not only to model quality, but to the question: who controls the compute base? Original source

⚖️ Court stops Pentagon’s classification of Anthropic

A federal court in San Francisco has ruled that the Pentagon’s classification of Anthropic as a supply-chain risk was unlawful. According to the report, the Department of Defense allegedly blacklisted the company in retaliation for public criticism of the US government’s AI policy. Formally, the matter is not completely resolved yet, because another proceeding in Washington is running in parallel — but the signal is already strong. This matters for Anthropic because regulatory classifications obviously don’t just cost reputation, but can also affect business, partnerships, and, in a pinch, a path to the stock market. And for the AI industry overall, it’s a reminder: if you speak loudly about politics, you can end up in the legal machine room faster than you’d like. Original source

🔬 DeepMind’s Co-Scientist becomes a research partner

Google DeepMind has evolved Co-Scientist from a pure hypothesis generator into a labor-integrated research system. Built on Gemini, the system delivered experimentally validated results across multiple disciplines — from materials synthesis to the development of a medical AI architecture. The important point is not that AI is “replacing science,” but that it is increasingly being embedded into real research workflows. That’s a big difference from the usual demo videos: here, the systems are not just producing plausible ideas, but contributing to verifiable results in the lab. For research institutions and pharma, that is potentially enormous, because the bottleneck shifts from “finding ideas” to “validating good ideas faster.” That’s exactly where real productivity gains emerge. Original source

💰 Anthropic rents $45 billion in compute capacity

According to Bloomberg, Anthropic has agreed a deal with the British cloud startup Nscale worth around $45 billion. This isn’t the purchase of a few servers, but a clear statement: frontier models are practically unimaginable without massive, long-term secured compute capacity. The scale also shows how capital-intensive the business has become. If you want to play at the top, you not only need to build good models, but also secure compute early, stabilize supply chains, and convince financing partners. For the market, this is a sign that the infrastructure layer is continuing to grow in importance — and that cloud deals are now almost as strategically relevant as product launches. In short: the AI bill is coming, and it’s not a small one. Original source

📱 Google’s memory squeeze hits Android apps

TechCrunch reports that Google wants to introduce new memory limits for Android apps because the AI boom is contributing to hardware bottlenecks. Cheaper smartphones in particular could end up with less RAM as a result. That sounds like a niche platform issue at first, but it matters for the market: when compute resources in data centers become scarcer and more expensive, the consequences show up elsewhere too — for example, in the supply of consumer devices. For app developers, this means more pressure on memory usage and optimization. For users, in the worst case, it means more enforced discipline in everyday mobile life, even though the word “artificial” this time, unusually, doesn’t come from AI. It’s a good example of how AI is changing not just software, but also hardware design and product policy. Original source

🎙️ Gemini 3.5 Transcribe becomes everyday-ready

Google has introduced Gemini 3.5 Transcribe, a new speech-to-text model that recognizes more than 85 languages, filters filler words in real time, and corrects slip-ups. According to Google, the word error rate in streaming is 4.0 percent, and latency is said to be 70 percent lower than Chirp 3. Especially practical: via Function Calling, the model can delegate tasks to other Gemini models. That makes it interesting not just for transcription, but also for meeting workflows, support, media production, and more. The key here is the combination of quality, speed, and multilingual capability — exactly what matters in daily use when languages aren’t perfectly separated, but everything still needs to be usable immediately. For many teams, this could be a noticeable upgrade. Original source

Google DeepMind is experimenting for the first time with a double-blind evaluation for a frontier model. Using “Confidential Space,” the aim is to ensure that neither Google sees the test questions nor the reviewers see the model weights. The pilot project is running with the Singapore AI Safety Institute and could be an important step against manipulation in AI benchmarks. Why does this matter? Because benchmarks are only valuable if they are hard to “game.” That’s been the industry’s problem for years: models are often optimized for tests rather than real robustness. If double-blind methods work, they could become a new standard for more trustworthy AI evaluation. Or, put differently: finally a benchmark where not everyone is secretly peeking at the answers. Original source

🛠️ Tool tip of the day

If you want to systematically monitor AI workflows, models, and product maturity, a tool for continuous research and market tracking is worthwhile. Especially for topics like inference costs, open-source models, and regulation, it can save you a lot of manual busywork. #

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