🔧 OpenAI's Jalapeño chip beats Nvidia GPUs by 3.6× 🔧
OpenAI's first custom chip targets inference speed and energy efficiency.

OpenAI has published the first performance benchmarks for Jalapeño, its custom-designed inference accelerator built for low-latency agent workloads. The 700-watt chip delivers up to 3.6× faster responses and 1.9× better energy efficiency compared to Nvidia's 1,200-watt flagship GPUs — a direct challenge to Nvidia's inference dominance. AI itself was used to design circuits and program kernels during an accelerated nine-month development cycle. OpenAI plans to deploy Jalapeño in its own infrastructure by year-end, with two more chip generations in the pipeline through 2027. The chip won't be sold externally — this is about vertical integration, reducing OpenAI's long-term reliance on external hardware suppliers.

🧠 Claude's memory now works across Chat and Cowork 🧠
Anthropic's rollout unifies persistent memory across all Claude platforms.

Anthropic has merged the memory systems of Claude Chat and Claude Cowork, meaning conversations and context now flow seamlessly across both platforms. The feature is enabled by default for all users — a bold UX decision that prioritizes continuity but raises questions about user awareness. Claude now proactively adds topics to memory during conversations, building a persistent knowledge base over time. All memorized information is stored as editable files organized under "Topics" in memory settings, giving users full control to read, edit, or delete entries. This shift toward persistent, personalized AI matters because it locks in user retention through accumulated context, making it harder to switch to competing platforms.

🍏 Apple ships M6 chip in $899 Mac Mini 🍏
Apple's new chips accelerate local AI processing at accessible pricing.

Apple announced the M6 and M5 Ultra chips for the Mac Mini and Mac Studio, positioning the new $899 Mac Mini as its flagship desktop for always-on agentic AI computing. The updated chips handle AI workloads up to 4× faster than their predecessors, significantly expanding the range of models users can run locally. This launch signals Apple's intensifying push into on-device AI, offering developers and power users a capable alternative to cloud-dependent workflows. The M5 Ultra targets the Mac Studio for professional-grade creative and computational tasks. The performance-to-price ratio could accelerate adoption of local inference, expanding competitive pressure on cloud providers.

INTERESTING TO KNOW

🖥️ Perplexity and Nvidia partner on local AI agents 🖥️

Perplexity and Nvidia have launched Portable Computer, a local-first rollout of Perplexity's Computer agent that runs entirely on-device with zero token costs and full file privacy — eliminating ongoing API pricing for agentic workflows. Users choose between Qwen 3.8 27B or PPLX 27B as their local model, with optional access to 15+ cloud models when needed. Currently available for Pro, Max, and Enterprise subscribers on Linux, it requires an RTX GPU with at least 24 GB VRAM, and Windows support is expected in September.

🚪 OpenAI loses data center chief Chris Malone 🚪

Chris Malone, the executive who led OpenAI's data center expansion strategy, departed last week — a notable loss during a critical infrastructure rollout phase. His exit is the latest in a series of high-profile departures that raise questions about leadership stability, with potential impact on timelines for OpenAI's compute expansion. Malone's role was central to securing the physical capacity needed for both model training and the deployment of Jalapeño, OpenAI's new custom inference chip.