Anthropic is exploring designing its own AI chips through 'extreme co-design' to cut costs, reduce latency, and break free from Nvidia dependence, though it would rely on manufacturers like TSMC or Samsung for production.
Snap's eyewear subsidiary, Specs, partners with Qualcomm to power future AR glasses with Snapdragon technology, promising true spatial AR capabilities that accurately overlay digital content onto the real world — setting it apart from Meta and Google's competing devices ahead of a consumer launch later this year.
A San Francisco startup is growing AI chips from living human brain cells that train models 3x faster while using 66% less energy, backed by $25 million in seed funding and already in talks with major AI labs.
Snap's fifth-generation AR Spectacles impress with crisp visuals, gesture controls, and real-time multi-user AR experiences powered by four cameras and dual Snapdragon processors, but a bulky design and $99/month developer-only price tag stand between these glasses and mainstream adoption.
Meta doubles down on AI dominance, committing an additional $21 billion to CoreWeave on top of a prior $14.2 billion deal, securing Nvidia GPU-powered cloud capacity through 2032 as part of a staggering $115–$135 billion capital expenditure plan this year.
Amazon surges to $717 billion in revenue with AWS AI income exceeding $15 billion annually, as the tech giant unleashes a $200 billion spending spree on AI infrastructure, robotics, drone delivery, and satellite broadband, cementing its dominance across cloud, grocery, and logistics.
Alibaba and China Telecom launch a massive AI data center in Guangdong powered by 10,000 homegrown Zhenwu chips, signaling China's aggressive push toward semiconductor self-sufficiency as U.S. export restrictions cut off access to advanced Nvidia hardware.
TriAttention, a new open-source method, compresses KV cache by 10.7x and doubles inference speed on long reasoning tasks while maintaining full attention accuracy, using trigonometric frequency-domain compression and integrating seamlessly with vLLM for easy local deployment on memory-constrained GPUs.
A new open-source toolkit called Gemma Multimodal Fine-Tuner now lets developers fine-tune Google's Gemma AI models on text, images, and audio directly on Apple Silicon Macs, requiring no NVIDIA hardware and supporting terabyte-scale cloud datasets while keeping data fully private on-device.
Generalist AI's new GEN-1 robot model is achieving a 99% success rate on physical tasks like folding clothes and packaging items, up from 64%, while operating three times faster and requiring only one hour of training data per task on any hardware.