AI Infrastructure news and analysis on Web Pulse: 34968 curated articles tracking the latest AI Infrastructure developments, tools, and research, updated continuously from vetted sources.
A benchmarking study by the aistack team found that the choice of coding agent harness has less impact than expected on task resolution accuracy, but can double the cost of a single resolved task from $0.35 to $0.7 depen…
ByteDance's cloud division Volcano Engine is enabling next-generation cloud services on Arm's AGI CPU platform and plans to bring the first Arm AGI CPU-powered agentic sandboxes to market, Arm announced at its Sept. 8 Ar…
ZenMux launched a unified AI model gateway that offers built-in insurance compensation when outputs are hallucinated, excessively slow, or low-throughput, according to the company's product page. The platform provides on…
DeepSeek released DeepSeek V4.1 Flash, a 552B-parameter mixture-of-experts model built on a new causal encoder-decoder architecture that natively supports multimodal visual understanding. The model is priced at $0.15-0.3…
DeepSeek formally released its V4.1 Flash model on Sept. 10 and told API users that V4.1 Flash surpassed V4 Pro in performance, cost, speed and total completion time in internal and external testing. Starting Sept. 14 at…
A technical report titled "DeepSeek-v4.1-Flash: Pushing Limits of KV Cache Compression" has been posted as a PDF, presenting work on compressing the key-value cache for the DeepSeek-v4.1-Flash model. The document's body …
DeepSeek released DeepSeek-V4.1-Flash, the smallest model in its new architecture family, now live on the DeepSeek API with native multimodal support under the model name deepseek-flash. The 552B-parameter mixture-of-exp…
A developer building a Cloudflare Worker that polls a status line endpoint every 10-20 seconds discovered that a KV-based rate limiter was exhausting the free plan's account-wide 100,000 daily get() quota, causing every …
Attention cost scales roughly quadratically with context length and the KV cache grows linearly with every retained token, so AI agents that re-send full conversation history each turn pay the quadratic prefill cost repe…
OPAQUE, a confidential computing company that runs AI workloads inside hardware-isolated environments, released an open standard called Weight Custody Manifest that lets AI model builders decide when and where their weig…
Microsoft announced Project Zenith on September 4, a new Windows developer hardware class requiring a minimum 64GB of unified memory and 250GB/s of memory bandwidth to run 30-billion-parameter AI models locally without a…
Researchers introduced Osprey, a target-agnostic pretraining method for speculative decoding drafters that bootstraps them from off-the-shelf pretrained small language models, according to the arXiv paper 2609.09338v1. A…
Researchers proposed X-CoSD, a cross-vocabulary collaborative speculative decoding framework for distributed LLM inference that works with heterogeneous small language model and server LLM vocabularies, according to an a…
A reasoning-aware compression framework that selectively restores the most quantization-sensitive circuits to FP16 achieves Pareto-optimal energy-accuracy points unreachable by uniform quantization, according to an arXiv…
A study of seven experiments on live shared LLM-serving systems found that KV-cache timing side-channel reliability collapses under multi-tenant contention, with mean Cohen's d on a vLLM server running DeepSeek-R1-Distil…
A developer has open-sourced fastapi-multi-llm-starter, an MIT-licensed FastAPI template that unifies streaming access to multiple LLM providers including Anthropic, Google, OpenAI, and DeepSeek behind a single asynchron…
A production playbook for orchestrating AI agents in enterprise .NET applications outlines how a U.S. retail platform cut latency from 1.2 seconds to under 150 ms per SKU and reduced token usage by 40% by adding an orche…
Nvidia said on Thursday it plans to expand data centre computing capacity in Australia to up to 2 gigawatts of AI-related capacity by 2027 through partnerships with local firms Firmus, CDC, NEXTDC and AirTrunk, more than…
A developer built an architecture connecting Google Cloud Run Sandboxes (gVisor) directly to Google Apps Script, enabling deterministic Python and Bash execution inside Google Workspace in 200–450 ms with zero-trust micr…
Kepler Compute, a new memory startup, announced it is developing leadership memory beyond HBM and SRAM that it says can be built without dependence on EUV lithography, with sampling planned for 2026 and production in 202…