OpenAI unveils its first custom AI inference chip, 'Jalapeño,' built with Broadcom, delivering up to 4x better performance-per-watt than Nvidia's Blackwell across key benchmarks, with plans to scale production into 2027 to cut compute costs and broaden AI access.
NVIDIA launches the DGX Station, a deskside AI supercomputer powered by the GB300 Grace Blackwell Ultra Desktop Superchip, delivering 20 petaFLOPS of compute and 748 GB of memory to run trillion-parameter AI models locally on enterprise desktops.
NVIDIA launches the DGX Spark, a compact desktop AI supercomputer powered by the GB10 Grace Blackwell Superchip, capable of running 200 billion parameter AI models locally with 1 petaFLOP of performance and 128 GB of unified memory, bringing cloud-level AI power to the desktop.
Mistral AI and HUMAIN are launching a multi-hundred-million-euro partnership to build sovereign AI infrastructure across Saudi Arabia and the Middle East, developing localized Arabic language, cybersecurity, and voice AI models for regulated industries including finance, telecom, and the public sector.
Nvidia's linear math breakthrough slashes AI costs, Hugging Face eyes a $13B acquisition deal, and London startup Inherent stuns the industry as its lean 27-billion-parameter AI agent Faraday outperforms OpenAI and Anthropic in scientific research replication.
Anthropic hires Amir Salek, founder of Google's TPU program and former Nvidia chip leader, to build a custom in-house AI silicon operation designed specifically for Claude's workloads, while maintaining existing hardware partnerships with Nvidia, Google, and Amazon.
A mysterious free AI model called Ox Alpha, built by an anonymous developer, shatters OpenRouter's all-time launch record by generating 11.6 trillion tokens in just three days, more than doubling the previous record, while amassing 327,000 unique users — yet its creator, architecture, and data policies remain completely unknown.
OpenAI slashes GPT-5.6 prices by over 20% across multiple model tiers, dropping its top model to $4 per million input tokens to undercut Anthropic and DeepSeek as enterprises pull back on frontier AI spending.
Thomson Reuters launches 'Thomson,' its first proprietary AI model trained on decades of Westlaw, Reuters, and financial data, costing just $40 million to develop while reportedly matching frontier model performance — now live in CoCounsel Legal.
Liquid AI and Artificial Analysis launch Pipette, an open-source benchmarking platform testing 30+ AI models across 1,000+ edge device configurations, delivering real-world performance data on devices like the iPhone 17 Pro and MacBook Pro M5 Max through native apps and an interactive public dashboard.