The GPU Compute Tightness Index fell to 43.8/100 as of June 29, 2026, entering Loose territory after a monthly drop of 13.7 points, indicating abundant idle AI infrastructure and supply exceeding near…
GPU utilization in production Kubernetes clusters averages just 5%, according to Cast AI's 2026 State of Kubernetes Optimization Report, while hardware costs continue to rise. The GPU Cost Optimizatio…
AI chip startup Etched raised $800 million and secured $1 billion in sales contracts for its Sohu ASIC, a transformer-specific chip that claims to outperform Nvidia's H100. The company, founded in 202…
LLM inference optimization can reduce serving costs by 5-10x and latency by 3-5x, often determining whether an AI feature ships. The bottleneck is memory bandwidth during autoregressive decoding, and …
HIVE Digital Technologies trained large language models on older Nvidia A40 GPUs in Paraguay and matched the performance of Nvidia's H100 chips after optimization with Columbia University researchers.…
Revolut has developed PRAGMA, a transformer-based banking foundation model trained on 40 billion events from 25 million users, achieving a 64.7% improvement in fraud recall and up to 130% uplift in cr…
The launch of Z.ai's GLM 5.2 and DeepSeek V4 Flash has created a 50x price gap between open-weight APIs and closed frontier models, reshaping the build-versus-buy calculus for developers. While open-w…
IBM unveiled a 0.7nm chip technology using a nanostack architecture that doubles transistor density over its 2nm design, packing nearly 100 billion transistors on a fingernail-sized chip. The chip is …
OpenAI has built its first custom AI chip, named Jalapeño, in collaboration with Broadcom. The chip is designed for inference, not training, meaning it handles model responses after training is comple…
Google's TPU v5p and upcoming Ironwood TPU7x offer higher raw matrix throughput and simpler interconnects than NVIDIA H100 GPUs for large-scale AI training, but GPUs remain superior for inference with…
Nvidia's advanced AI chips, including B200, H100, and H200, are selling at up to 50% premiums on China's black market, with over $1 billion worth smuggled in three months after US export controls tigh…
Cohere's 30-billion-parameter coding agent, with only 3 billion active parameters, achieved a score of 0.6 on a benchmark, outperforming models four times its size while running on a single H100 GPU.…
The AI hardware supply chain faces severe constraints, with TSMC holding 72% of advanced chip manufacturing and ASML monopolizing EUV lithography. CoWoS packaging capacity is sold out through 2026, an…
NVIDIA released Qwen3.6-35B-A3B-NVFP4, a post-training FP4-quantized variant of Alibaba's 35B MoE model that fits on a single H100 by reducing VRAM from ~71 GB to ~23 GB. The quantization targets weig…
Nvidia H100 GPUs cost $30,000 to $40,000 to buy in 2026, with cloud rental rates ranging from $1.03 to $12.29 per GPU-hour depending on the provider. Neo-clouds offer the cheapest spot instances, whil…
A developer open-sourced AutoMegakernel, a tool that compiles any HuggingFace model into a single persistent megakernel, reducing overhead by launching one kernel per forward pass. It includes a stati…
The AI boom is pulling GPU vendors away from double-precision (FP64) hardware essential for scientific HPC, as NVIDIA, AMD, and Intel prioritize low-precision AI cores. New chips like NVIDIA's B200 an…
Sarvam raised $234 million at a $1.5 billion valuation in a Series B extension led by HCLTech, which contributed $150 million as the lead strategic investor. The Bengaluru-based AI startup, which has …
Cohere released North Mini Code, a 30-billion-parameter open-source coding agent with a Mixture-of-Experts architecture that activates only 3 billion parameters per token, enabling it to outperform mo…
A developer traced a hybrid Mamba-Transformer MoE inference run and found that MoE all-to-all collective stalls dominate the tail latency, with a 69x tail ratio, despite dashboards showing 96% GPU uti…