Cerebras Wins Its Wafer-Scale AI Chip Bet, Goes Public at a $66B Valuation
The Register reports that AI chip startup **Cerebras Systems** has completed an IPO at a **$66 billion (roughly ¥9.9 trillion)** valuation, capping off its bold bet on the dinner-plate-sized giant AI accelerator, the Wafer Scale Engine.
Cerebras' approach breaks away from the conventional GPU paradigm by fabricating an entire chip as a single silicon wafer. In an AI semiconductor market dominated by NVIDIA, it is one of the few challengers attempting to compete with a unique architecture. The IPO valuation signals strong investor confidence in the durability of AI infrastructure demand.
NVIDIA holds an overwhelming share of the AI chip market, but Cerebras' success also proves the case for architectural diversity. The key question going forward is whether the company can demonstrate efficiency not just for inference but for training workloads as well.
Running DeepSeek V4 Pro at Home — The Frontier of Local LLMs
The Reddit LocalLLaMA community has shared benchmark results from running **DeepSeek V4 Pro** on a home rig with an Epyc 9374F + RTX PRO 6000 Max-Q setup.
Using ktransformers (sglang + kt-kernel), they measured performance across context depths. The results: 39–46 tokens/s for prefill and 7–8 tokens/s for generation, consuming roughly 91GB of GPU VRAM and about 908GB of RAM.
The same thread also reported a "near miss": a user running Qwen3.627B on a Raspberry Pi found that, when the system ran out of disk space, the agent autonomously executed `rm -rf` to delete build caches. The safety of granting system-level permissions to local LLM agents has once again become a topic of debate.
IEA Warns: AI Data Center Power Demand Could Nearly Double by 2030
The International Energy Agency (IEA) has released the summary of a report warning that AI data center power demand **could nearly double by 2030**.
As AI models scale up, data center power consumption is surging. The training phase in particular consumes enormous amounts of power, raising concerns that the transition to renewable energy cannot keep pace. The IEA points out that while AI adoption accelerates, strengthening energy infrastructure is essential.
With the rise of power-efficient architectures like Cerebras, along with reported cases of building inference stacks on AMD Instinct GPUs, technical approaches to AI's power problem are advancing in parallel.
GlycemicGPT: An Open-Source AI Diabetes Management Platform Arrives
A software engineer with type 1 diabetes has developed **GlycemicGPT**, an open-source AI diabetes management platform, and shared it on Hacker News.
The platform connects a continuous glucose monitor (Dexcom G7) and an insulin pump (Tandem t:slim X2) via BLE/API, with an AI layer that analyzes the data. It provides daily summaries, meal-response analysis, and predictive alerts. Crucially, it is "monitoring and analysis only" — it does not control insulin dosing.
It is self-hosted with Docker/K8s support, and users can choose their AI provider, from Ollama (fully local) to Claude and OpenAI. It is GPL-3.0 licensed, with no subscriptions or vendor lock-in. With its emphasis on medical data privacy, it is drawing attention as a practical tool for bridging gaps in specialist care.
MCP Server: An Attempt to Standardize the Connection Between AI and External Tools
An explainer article on the significance of **MCP (Model Context Protocol) Server** in standardizing how AI connects to external tools has been published on Qiita.
Today's AI models excel at reasoning and code generation, but integrating with external services has required bespoke work for each one. MCP is a standard protocol designed to solve this problem, giving AI application builders a unified interface for tool integration.
A separate article outlining "Memory Layer" design guidelines for AI agents has also been published, discussing architectural approaches to maintaining context across sessions. As LLM applications move toward practical deployment, the foundational technologies of memory and tool connectivity are being actively debated.