AI Infrastructure

NVIDIA released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed hybrid MoE LLM that achieves 2.03x server throughput over Nemotron-3-Super at matched user throughput. The model uses iterative puzzle compression to reduce pa…

Nvidia CEO Jensen Huang said his software engineers prefer building AI agents over writing Python code, arguing that the shift creates jobs rather than eliminating them. Huang positioned the trend as a promotion for engi…

Cerebras Systems announced plans to deploy 200 megawatts of AI compute capacity across Europe by 2027, with initial capacity online by late 2026. The build-out, spanning France, Norway, and Finland, will partly serve Ope…

A rare bacterium, Cupriavidus gilardii, was found in wastewater discharged during construction of Meta's $800 million AI data center in Cheyenne, Wyoming, prompting a months-long cleanup and new environmental regulations…

Budget smartphones face a severe threat as surging RAM prices, driven by AI data center demand for memory chips, push costs to unsustainable levels. Memory now accounts for up to 64% of the bill of materials for sub-$99 …

AWS has launched a self-hosted control plane called Claude apps gateway that centralizes access, spending, and governance for Claude Code and Claude Desktop. The gateway replaces per-developer cloud credentials and manua…

A developer recounts their experience at the Databricks Data and AI Summit 2026, highlighting key announcements and trends in data engineering and artificial intelligence. The event showcased new features in Databricks' …

Hugging Face Spaces are experiencing a service disruption where Space-to-Space requests to *.hf.space domains return 503 errors from an AWS ELB since July 8. The issue affects all traffic between Spaces regardless of SDK…

Anthropic launched new free courses in its Anthropic Academy, covering API development, enterprise deployment, and individual project use of its Claude AI model. The courses aim to help developers and organizations build…

A user reports persistent rate limiting errors when downloading the DeepSeek-V3-0324 model from Hugging Face, despite trying Git over HTTPS, SSH, and the Hugging Face CLI with authentication tokens. The issue prevents su…

SK Hynix, the South Korean memory chipmaker, is preparing to list on the Nasdaq through American depositary receipts, targeting approximately $28 billion in proceeds. The offering is already more than seven times oversub…

Sales of refurbished tech devices like Xbox, iPhone, and MacBook are surging as new device prices spike due to memory chip shortages driven by the AI buildout. Apple, Dell, and Microsoft have raised prices, pushing consu…

Bain Capital confirmed that a special-purpose vehicle it set up for SK Hynix still holds a 14% stake in Kioxia Holdings, as Bain itself fully exited its position in the Japanese NAND flash memory company. The exit caps o…

ChangXin Memory Technologies (CXMT) is raising $4.3 billion in its Shanghai STAR Market IPO, the largest A-share listing in China this year, driven by surging AI memory demand that boosted its first-quarter net profit to…

A developer debugging a Gemini on Vertex AI deployment encountered six distinct bugs, each initially appearing as a 401 Unauthorized error. The issues ranged from a PM2 configuration typo causing a 502 Bad Gateway to the…

A developer successfully ran GLM 5.2, a 744B-parameter Mixture-of-Experts model, on a laptop with 32GB RAM by streaming routed experts from disk. The project, called Colibrì, is a single C file engine that activates only…

DNotifier provides observability into multi-agent systems by capturing internal states, communication logs, and decision paths of interacting AI agents. The platform offers a unified SDK to track execution graphs, monito…

Francisco Javier Roldán Velásquez, founder and CEO of Ethos Engine, is seeking pre-seed funding and Silicon Valley connections to scale his deterministic AI safety infrastructure. The system uses proprietary mathematical…

Chinese AI labs, including DeepSeek and Zhipu AI, are developing custom chips to reduce long-term operating costs and improve hardware-software synergy, but the strategy carries significant risk due to heavy upfront inve…

Google announced the general availability of AlloyDB AI functions, which allow developers to call large language models directly within SQL queries, along with two acceleration techniques: smart batching and proxy models…