Baidu officially launched the ERNIE large model 5.1 on May 9, leveraging multi-dimensional elastic pretraining technology. The total parameters are compressed to about one-third of ERNIE 5.0, and activated parameters to approximately half, reducing pretraining costs to just 6% of industry models at comparable scale. The model ranks fourth globally and first in China on the Arena Search Leaderboard with a score of 1223, becoming the only domestic model on the list. It outperforms DeepSeek-V4-Pro on agent evaluation benchmarks such as τ³-bench and SpreadsheetBench, achieving a 99.6 score in the AIME26 math competition—approaching the level of Gemini 1.5 Pro. Technically, it adopts a decoupled fully asynchronous reinforcement learning architecture and a four-stage post-training pipeline using multi-teacher online policy distillation. The model is now available for experience via the Qianfan platform and ERNIE Bot official website.

DeepSeek has initiated its first external funding round, targeting up to 50 billion RMB (approximately $7 billion USD), with a valuation reaching 350 billion RMB (~$50 billion). Founder Liang Wenfeng personally contributed around 20 billion RMB, accounting for 40% of the round. Tencent, Alibaba, and the National Integrated Circuit Industry Investment Fund are all engaged in discussions. This marks DeepSeek’s transition from a 'no-funding, no-commercialization' research lab model to a capital-intensive AI company. Funds will accelerate model iteration and commercialize enterprise products. The V4.1 model is scheduled for June release, supporting MCP protocol and multimodal capabilities for images and audio. DeepSeek's valuation previously surged from $10 billion to $50 billion within 21 days.

NVIDIA has released Star Elastic, a post-training method that embeds three inference models of 30B, 23B, and 12B parameters into a single checkpoint, enabling zero-shot slicing extraction. Based on the Nemotron Nano v3 hybrid architecture, it uses learnable routers and Gumbel-Softmax joint training, saving 360 times token cost compared to full pretraining. The elastic budget control strategy (small model thinking + large model answering) improves accuracy by 16% and reduces latency by 1.9x in the 23B→30B configuration. NVFP4 quantization compresses the 30B model to 18.7GB, allowing the 12B version to run on an RTX 5080 and achieving 7426 tokens/s throughput on the RTX Pro 6000. All variants are publicly available on Hugging Face.

Anthropic's AI model, Claude Mythos Preview, discovered thousands of zero-day vulnerabilities during controlled testing, including 271 flaws identified in a single Firefox scan, as well as vulnerabilities in the 27-year-old OpenBSD and a remote execution flaw in the 17-year-old FreeBSD system. The model has not been publicly released but is being offered through Project Glasswing to around 40 tech companies for early access to assist in remediation. Federal Reserve Chair Jerome Powell and the Treasury Secretary have urgently convened major bank CEOs to discuss risk mitigation. Anthropic warns adversaries could replicate this capability within 6 to 12 months. OpenAI has responded by launching GPT-5.4-Cyber. The cybersecurity community views AI-accelerated vulnerability discovery as an acceleration of trends rather than a disruption.

NVIDIA, AMD, and Intel—the three major chipmakers—have rarely joined forces to invest $100 million in the RadixArk team, developers of the open-source inference engine SGLang and reinforcement learning framework Miles. SGLang has been deployed across over 400,000 GPUs, serving companies like Google and xAI, with Day-0 compatibility for new models. RadixArk achieved full-stack support for inference and RL training on the same day DeepSeek-V4 was released. The three hardware vendors have differing motivations but share a unified goal: building an open AI infrastructure layer not locked to any single cloud provider or chip giant. NVIDIA aims to strengthen its ecosystem, while AMD and Intel hope to break CUDA's dominance through a fair, open layer.

Airbnb CEO Brian Chesky revealed during the May 8 earnings call that nearly 60% of code last quarter was generated by AI tools, significantly accelerating feature development. AI also resolved 40% of customer support issues without human intervention. Chesky noted that AI enables individual engineers, under supervision, to accomplish work previously requiring a 20-person team. An increasing number of design and engineering managers are returning to coding or adopting AI tools like Claude Code. Financially, Q1 net profit reached $160 million, revenue grew 18% to $2.7 billion, and booking nights totaled 156.2 million. Chesky did not confirm potential layoffs but acknowledged current challenges in applying AI to the travel domain.

Google's AI drug discovery subsidiary, Isomorphic Labs, is in advanced discussions for a new funding round aiming to raise over $2 billion, led by Thrive Capital with participation from Alphabet. The company spun out from Google DeepMind in 2021, with CEO Demis Hassabis continuing in his role, leveraging AI technologies like AlphaFold to accelerate drug development. It has established partnerships with major pharmaceutical firms including Johnson & Johnson, Eli Lilly, and Novartis. The new capital will upgrade its drug design engine and expand global operations. Despite past setbacks in Silicon Valley's attempts in healthcare, AI-driven drug discovery continues to attract significant investment.

In May 2026, SK hynix faces severe shortages in high-bandwidth memory (HBM) chips driven by AI demand, prompting several global tech giants to proactively offer funding for new production lines and ASML extreme ultraviolet lithography equipment to secure supply. SK hynix's capacity is nearly maxed out, with clients' HBM demands over the next three years far exceeding production capabilities. The company holds 57% of the HBM market and reported a 72% operating profit margin in Q1. Despite ample capital, SK hynix is carefully evaluating partnership proposals to avoid dependency on a single buyer. Industry analysts expect memory shortages to persist until 2030. Sony and Nintendo have already stated that rising memory prices are impacting their gaming businesses.

Xiaomi has launched the MiMo-V2.5 native multimodal large model, supporting unified understanding of text, images, video, and audio, with strong agent capabilities. It employs a Mixture-of-Sparse-Experts (MoSE) architecture with 310B total parameters and 15B activated parameters, supporting up to 1 million tokens of context length. Core technologies include hybrid attention mechanisms, a 729M-parameter vision encoder, a 261M-parameter audio encoder, and a multi-token prediction module. The model was trained on approximately 48T tokens using FP8 mixed precision and optimized for agent tasks via SFT, reinforcement learning, and knowledge distillation. It has been released on Hugging Face and supports deployment via SGLang and vLLM frameworks.