AI Radar: Nvidia Chaos, Google API & New Data Wars
Nvidia, Google, and Anthropic are shaping AI today: new models, chip power plays, data exports, and a fresh look at real-time speech recognition.
Inhaltsverzeichnis
Today’s AI landscape is intense: it’s about power, markets, and models — exactly the kind of mix where “exciting” quickly becomes “strategically relevant.” Between Nvidia rumors, Google product launches, and new questions around AI training data, it’s becoming very clear where the market is headed: away from pure chatbot hype and toward infrastructure, specialization, and control.
For you, that means: anyone looking to build, buy, or regulate AI should pay especially close attention today. Because today’s news is not just headlines — it’s a small preview of the next phase of the AI race.
🧨 Jensen Huang must appear before the US Congress
Nvidia CEO Jensen Huang is in the spotlight of US politics: House Republicans want to question him behind closed doors about an alleged espionage case involving China. According to the report, Huang is expected to respond to the subpoena by Wednesday evening. Even if the details are still thin, the signal is enormous: Nvidia is no longer just a chip maker, but a geopolitical node in the global AI power struggle. When uncertainty arises around the top leadership, it’s not only investors who nervously watch the calendar — customers, partners, and regulators do too. For the market, this is a reminder of how vulnerable even the most powerful AI players are when technology, politics, and foreign trade collide. Source: TechCrunch
📉 Nvidia leadership wobbles: Kress steps in for now
The situation becomes even more explosive with the report that Jensen Huang has resigned with immediate effect as CEO, president, and board member, according to a company statement. CFO Kress is taking over leadership temporarily. If you had to blink twice just now: yes, this is exactly the kind of news that can shake the entire AI infrastructure sector. For many models and cloud providers, Nvidia is the central hardware layer — so a leadership change there is not just a company detail, but a potential market marker. Especially interesting: whether the strategic direction really changes, or whether management is simply passing the baton formally. At Nvidia, neither is ever entirely harmless. Source: The Decoder
⚙️ Google launches Gemini Decisions for mass decision-making
With Gemini Decisions, Google DeepMind is introducing a new API that is not aimed at fluent text generation, but at fast, closed-form decisions. Up to 1,000 cases can be evaluated in a single pass across text, image, audio, and video — with latencies of 150 to 250 milliseconds. This matters because it highlights a clear trend: AI is increasingly being built as a decision engine rather than a chat engine. For companies, that’s exciting in moderation, scoring, classification, or prioritization — in other words, wherever the most beautiful sentence doesn’t matter, but the fastest usable decision does. Google is positioning itself cleverly against the industry’s “generate everything” reflex. And yes: not every AI needs to write a novel; some just need to finally say what’s what. Source: The Decoder
💸 US data brokers are selling training data to China
An investigation by ChinaTalk suggests that US data brokers such as Mercor, Surge AI, and AfterQuery have sold AI training data worth around half a billion dollars to Chinese labs. Particularly sensitive: this allegedly includes RL environments originally developed for OpenAI or Anthropic. The issue is so explosive because data is no longer just raw material, but strategic infrastructure. If export controls barely catch these flows, it reveals a core problem in AI regulation: code can be protected, compute can be limited — but data often takes the smarter route. For global competition, this means access to high-quality training data remains one of the underrated levers in the AI race. Source: The Decoder
🧠 Anthropic bypasses Nvidia’s CUDA advantage
Anthropic claims that, with the help of its own AI tools, it ported Claude to AMD GPUs in just three days — with almost no human intervention. That’s more than a technical footnote: it directly challenges Nvidia’s biggest moat, the CUDA stronghold. If model porting becomes significantly faster and cheaper, dependence on a single chip ecosystem decreases. For the industry, that could be a turning point, because hardware decisions would then depend less on developer convenience and more on cost, flexibility, and supply-chain risk. Whether Anthropic is an exception or a harbinger remains to be seen. But the trend is clear: AI is not only eating work — it now seems to be eating away at some hardware lock-in too. Source: The Decoder
🎙️ ElevenLabs Scribe v3 Live makes conversations more robust
With Scribe v3 and Scribe v3 Live, ElevenLabs is launching new speech recognition models for louder, messier conversation settings. Especially interesting: the real-time version is supposed to stop “losing” speakers halfway through a sentence, meaning it transcribes more accurately when multiple voices, background noise, or interruptions are present. That’s relevant for call centers, meeting tools, dictation apps, and voice agents — anywhere speech-to-text is not just nice to have, but business-critical. According to measurements by Artificial Analysis, Scribe v3 ranks just ahead of competitors like OpenAI, Google, and Microsoft in accuracy. The catch: top precision comes at a cost. For the market, this means audio AI continues to mature — and that quality is increasingly being judged by latency, robustness, and price, not just demo polish. Source: The Decoder
🛠️ Tool tip of the day
If you’re experimenting with multimodality or want to quickly build decision workflows, it’s worth taking a look at a platform for API testing and prototyping for AI models — especially if you’re moving back and forth between text, image, and audio. For early product ideas, that’s often the difference between “interesting demo” and “actually works.” #
Don’t want to miss any news? Subscribe to the newsletter