Tencent's Xiaowei, highlighted in its latest earnings report, is an AI agent focusing on user privacy and efficiency. A 24-hour trial showed its potential in automating daily digital tasks, though with occasional frustrations.
French startup Kog is developing software optimization techniques to achieve extremely fast AI inference on standard datacenter GPUs, aiming to target professional workflows and larger models.
The cost for businesses to run AI models hit a yearly low between US$1.16 and US$1.18 per million tokens, driven by a global price war and Chinese open-source tools like DeepSeek, according to Jefferies and Silicon Data.
Breakthroughs in cheap Chinese open-weight AI models have spooked US investors, though analysts argue lower costs will supercharge global demand and accelerate AI adoption.
Advanced Micro Devices has agreed to acquire Toronto-based startup Taalas to boost its AI inference capabilities with customized, hard-wired chips, expanding beyond general-purpose GPUs.
AI infrastructure company Runware has launched its 'Sonic Inference Pod,' a modular, transportable data center designed for flexible AI compute, offering higher quality inference at a lower cost compared to serverless platforms and GPU clouds.
AI chip startup Etched, founded by three Harvard dropouts, has closed a $300 million Series C funding round, valuing the company at $10.3 billion. This marks a doubling of its valuation in seven months, following successful chip manufacturing and $1 billion in orders.
A Hong Kong High Court has ordered the release of 14 defendants, including 12 men and one woman, charged with the alleged murder of former triad gangster Luk Wing-ho five years ago. The court ruled that prosecutors failed to prove Luk's fatal head injuries resulted from deliberate assaults rather than boxing bouts at his gym, directing the jury to acquit all defendants due to insufficient evidence regarding the 2021 incident.
Peking University researchers have created an all-optical interconnect system that links standard electronic chips, accelerating AI inference speeds by over 100 times and reducing computational resource needs by nine times. Published in National Science Review, the system uses FPGA chips and custom hardware components for signal conversion.
Europe must accelerate AI development in sovereignty, scale, security, and sustainability to remain competitive. Key strategies include a hybrid approach to tech sovereignty, supporting smaller businesses, shifting from reactive to proactive security, and adopting a frugal approach to AI's energy consumption.
French AI startup ZML has launched ZML/LLMD, an inference server designed to optimize large language models across various chips, including Nvidia, AMD, and Apple. The software aims to break vendor lock-in and reduce AI costs by enabling efficient use of diverse hardware, potentially disrupting Nvidia's market dominance.
Chinese AI startup DeepSeek is reportedly developing its own AI inference chip, aiming to reduce reliance on Nvidia and potentially challenge Huawei. This move aligns with Beijing's push for domestic alternatives amid US export controls.
China's DeepSeek is developing its own AI chip for inference, the process of generating responses to user queries. Previously relying on Nvidia and Huawei chips, the company began work on its own solution about a year ago and is in early discussions with partners. The move comes amid growing demand for inference chips as AI becomes more widespread.
Prince Harry's final lawsuit against Associated Newspapers, publisher of the Daily Mail, has been dismissed. A judge ruled he failed to prove his privacy invasion claims, stating there was a possibility the news came from legitimate sources. The decision ends Harry's legal battle against tabloid publishers.
AI chip startup Etched announced $1 billion in contract orders for its "frontier inference clusters," systems designed to accelerate AI model inference. The company also revealed it has raised $800 million to date, including a $500 million round closed in December.
Chinese food delivery giant Meituan has open-sourced LongCat-2.0, an AI model with 1.6 trillion parameters. It claims to be the country’s largest AI model trained entirely on domestic hardware for both pre-training and inference, marking a significant step towards China's self-sufficiency in chip technology.
OpenAI has revealed Jalapeño, its first custom-built inference processor, created in partnership with Broadcom. Designed for OpenAI's AI models, it aims for better performance-per-watt than current alternatives and could reduce reliance on Nvidia GPUs.
OpenAI and Broadcom announce Jalapeño, an ASIC designed for large language model inference in data centers, promising better performance per watt than current solutions, with deployment expected by year-end.
Hong Kong, a global AI hub, faces significant energy constraints due to its electricity deficit. To sustain its AI ambitions, the city needs to integrate energy-aware planning and leverage the Greater Bay Area for compute infrastructure, focusing on high-value, low-energy segments like financial AI and regulatory technology.
AI inference company Baseten is reportedly close to securing a $1.5 billion funding round at a $13 billion valuation, a significant jump from its previous $5 billion valuation just five months ago. The round is described as split-priced, with some investors valuing the company at $13 billion and others at $11 billion.
Chinese AI models are competitive with US counterparts, but hardware lags. While domestic chips are used for inference, pre-training still relies on foreign silicon. China is pushing for self-sufficiency in AI hardware across all development stages.
Xiaomi achieves AI inference speed record with MiMo-V2.5-Pro-UltraSpeed, processing over 1,000 tokens per second on a single 8-GPU node, outperforming competitors with software innovations.
Chinese researchers, led by Huawei and multiple institutes, successfully conducted full-parameter post-training on a 1.6 trillion-parameter AI model using over 1,000 Huawei chips, enhancing China's AI self-reliance.
Perplexity CEO Aravind Srinivas announced a new hybrid local-server AI inference orchestrator at Computex 2026. The system automatically decides where to run AI tasks, balancing local processing for privacy with cloud power for complex computations, aiming for optimal "token value per watt."
