
Tech • IA • Crypto
Nvidia reported a record $81 billion quarterly revenue, up 85% YoY, with $75 billion from data centers alone. Despite the blowout results, the stock fell as investors priced in rising competition. Google’s TPU 8T and 8i chips claim major efficiency gains, signaling a shift toward custom silicon. The market is now betting that hyperscalers will erode Nvidia’s dominance faster than expected.
Google introduced TPU 8T for training and TPU 8i for inference, targeting specialized AI workloads. The company claims nearly 3x better performance per dollar for training and 80% gains for inference. This reflects a broader architectural split between training and inference infrastructure. It also strengthens the case for vertically integrated AI stacks controlled by hyperscalers.
OpenAI launched ChatGPT Workspace Agents, powered by GPT-5.5, to automate multi-step business workflows. These agents can operate across tools like email, calendars, and documents, executing tasks end-to-end. A no-code builder enables non-technical users to deploy custom agents inside teams. The push targets the “middle layer” of enterprise work, where coordination overhead remains high.
OpenAI is redesigning ChatGPT into a full agent-driven platform that executes tasks rather than just answering questions. Deep integration of Codex will allow control over software, workflows, and external tools. The system is moving toward autonomous goal completion with minimal user input. This positions ChatGPT as a central operating layer competing directly with enterprise software ecosystems.
Google’s Gemma 4 (12B) model enables powerful AI to run locally on machines with around 16 GB VRAM. Tools like LM Studio simplify deployment, letting users run models privately without cloud APIs. This reduces cost, latency, and data exposure while enabling offline workflows. The shift marks a growing move toward personal, sovereign AI systems.
Research from Stanford and Tsinghua University shows identical models can perform up to 6x better depending on system design. The concept of harness engineering emphasizes memory, tools, workflows, and verification layers סביב models. Companies like OpenAI and Anthropic are adopting these architectures to improve reliability. Competitive advantage is shifting from model size to system design.
The United Kingdom will require Google to give publishers granular control over AI content usage. Websites can block both training and grounding at page-level precision. New Search Console tools will show how content appears in AI Overviews. The policy could reshape search economics and strengthen publisher leverage across Europe.
The Codeex AI coding agent surpassed 1 million downloads in a week and reached 4 million weekly active users. Teams report 50% more pull requests per engineer and dramatically faster delivery cycles. The system can execute full development workflows across entire codebases autonomously. Its rapid adoption highlights the shift from assistive coding to fully automated software production.