AI Infrastructure news and analysis on Web Pulse: 17419 curated articles tracking the latest AI Infrastructure developments, tools, and research, updated continuously from vetted sources.
Researchers introduced the Native Differentiable Virtual Machine (NDVM), a runtime structure that blends symbolic structure with numerical computation to accelerate AI model calibration. NDVM achieves about 60x per-lane …
SK Hynix shares jumped nearly 13% in their Wall Street debut, marking the largest initial share sale in the U.S. by a foreign company, as surging demand for AI-powered memory chips drives the company's growth. The South …
DeepSeek releases V4 Pro, a 1.6-trillion-parameter Mixture-of-Experts model with 49 billion active parameters per token, achieving a top open-weight score of 80.6% on SWE-bench Verified. The model requires multi-H200/H10…
Researchers introduced Subspace-Aligned Rewiring (SAR), a post-training editing method for large language models that isolates reasoning-effective components in spectral space, preserving over 99% of performance using as…
South Korean memory chipmaker SK Hynix raised $26.5 billion in its NASDAQ debut on July 10, marking the largest US listing by a foreign company. The IPO was driven by surging demand for high-bandwidth memory chips used i…
An engineer running multi-region AI workloads reports that Chinese LLM APIs from DeepSeek, Qwen, GLM, and Kimi offer dramatically lower costs than US providers like OpenAI, Anthropic, and Google, with DeepSeek V4 Flash c…
Researchers introduced HetDPT, a depth pruning method for Vision Transformers that speeds up DeiT-B by 1.58x and DeiT-S by 1.39x with minimal accuracy loss, and combined with width pruning (HetDPT+) achieves up to 5.19x …
Together AI and AppsFlyer top Green Flag Digital's Q3 2026 ranking of private tech companies, which weighs funding, web traffic, and brand search volume. Together AI leads on the strength of an $800M funding round and co…
Anubis, a proof-of-work system designed to deter AI companies from scraping websites, has been deployed by an unnamed administrator. The system imposes computational costs on mass scrapers while aiming to remain negligib…
A developer detailed the local deployment of GLM 5.2, a 753B MoE model with 40B active parameters requiring 1.51TB of RAM in BF16 precision. Unsloth's dynamic quantizations, such as UD-Q3_K_XL at 343GB, offer a realistic…
Liquefied natural gas (LNG) is regaining strategic importance in South Korea as the government reassesses its energy mix to support massive investments in AI data centers and semiconductor manufacturing. Growing electric…
A developer built a stateful AI agent that survives session kills by using Tensorlake's MicroVM sandboxes instead of containers. The harness preserves running processes, files, and memory across suspend and resume, enabl…
Researchers introduced Lagrangian Reward Augmentation (LARA), a framework that steers frozen language models using safety constraints during inference without repeated weight updates. LARA improves the trade-off between …
A developer proposes replacing LLM API calls with coding agent calls to reduce costs, citing that coding assistants like Codex, Cursor Pro, and GitHub Copilot Pro cost significantly less per million tokens than tradition…
China is advancing its global soft power by releasing open-weight AI models that are widely adopted in Africa and other regions, while the U.S. relies on closed-source proprietary models. Andrew Ng warns that this trend,…
SK Hynix raised $26.5 billion in the largest foreign listing on Wall Street, with shares jumping 13% on Nasdaq debut. Chairman Chey Tae-won said the memory chip shortage will persist beyond 2030, driven by demand from AI…
Real-time AI interaction models like Moshi and Qwen-Omni face a hidden collapse risk due to swelling KV cache, causing sudden performance drops under sustained load. A simple fix—bounding each session's resident state wi…
OpenClaw's GLM-5 inference optimization study found that adjusting parameters like chunked prefill size and request concurrency improved throughput and reduced latency, cutting serving costs by 10.4% per request and 9.6%…
A developer compares Mem0 and TurboMem as memory layers for TypeScript AI agents. TurboMem runs as a native TypeScript library inside the process, avoiding the separate service and infrastructure required by Mem0. The co…
Google DeepMind's Gemma team partnered with Hugging Face in the Fast Gemma Challenge, achieving a 5x inference speed boost for the Gemma 4 model on a single NVIDIA A10G GPU, with peak performance of 491.8 tokens per seco…