Google has launched its flagship AI model Gemini 3.1 Pro, achieving a 77.1% validation score on the ARC-AGI-2 logical reasoning benchmark, with reasoning performance more than double that of its predecessor. The model excels across multiple professional benchmarks: 94.3% on GPQA Diamond for scientific knowledge, 2887 Elo on LiveCodeBench Pro for coding, 80.6% on SWE-Bench Verified, and 92.6% on MMMLU for multimodal understanding. Third-party evaluator Artificial Analysis confirms it as the current highest-performing AI model globally. Google is emphasizing functional outputs through 'smart applications' to enhance its potential in complex tasks such as scientific research and engineering, while providing stronger foundational support for autonomous agent development.

Reflection AI, founded by former Google DeepMind researchers, is seeking to raise $2.5 billion at a $25 billion valuation—nearly 46 times its $545 million valuation from a year ago. The company focuses on developing 'open-weight' AI models, positioning itself as a Western alternative to DeepSeek, serving enterprise and government clients. Investors include NVIDIA (which has invested approximately $800 million), JPMorgan, and Sequoia. Reflection AI is a founding member of NVIDIA's Nemotron Alliance and adopts a Red Hat-like model—freely releasing model weights to build an ecosystem while monetizing infrastructure and services. This move is seen as a strategic U.S. effort to compete with China in open-source AI.

French AI startup Mistral has completed its first debt financing round of $830 million, backed by seven top global banks, to build a data center near Paris. The facility will house 13,800 Nvidia GB300 GPUs with a total capacity of 44MW and is scheduled to go live in the first half of 2026, supporting AI model training and inference services. Mistral also plans to invest over $1.4 billion in digital infrastructure in Sweden, aiming to achieve a total of 200MW computing capacity across Europe by 2027. Recently, the company launched its open-source Small 4 model and enterprise customization platform Forge, with clients including ASML, Ericsson, and the European Space Agency, though its funding remains far below that of OpenAI and Anthropic.

South Korean cloud-native AI inference chip startup Rebellions has completed a $400 million pre-IPO funding round led by Mirae Asset and the Korea National Growth Fund, bringing total funding to $850 million and a $2 billion valuation. The company specializes in high-performance, low-power inference chips, with its Rebel100 NPU based on a chiplet architecture delivering superior performance per watt compared to GPUs. Rebellions follows a 'software-first' strategy, building an open-source, Kubernetes-based cloud-native stack compatible with frameworks like PyTorch and Hugging Face. It has launched vertically integrated products including RebelRack and RebelPOD. The new funds will accelerate U.S. market expansion, with Marshall Choy appointed to lead American operations, and the company is actively preparing for an IPO.

Alibaba has released its full-modal large model Qwen3.5-Omni, achieving state-of-the-art (SOTA) performance on 215 tasks covering audio, video understanding, recognition, and interaction, surpassing Gemini 3.1 Pro. The model supports 113 languages and dialects and exhibits emergent capabilities such as audio-video Vibe Coding. It comes in three versions: Plus, Flash, and Light, with input costs under RMB 0.8 per million tokens—just one-tenth of Gemini 3.1 Pro's cost. The Plus version performs especially well in audio benchmarks. Additionally, Microsoft has released Harrier-OSS-v1, a multilingual embedding model that achieves SOTA results on the multilingual MTEB v2 benchmark.

Insilico Medicine and Eli Lilly have announced a global R&D collaboration. Lilly has obtained exclusive worldwide rights to a portfolio of preclinical oral drug candidates, and Insilico will receive a $115 million upfront payment, with potential milestone payments totaling up to $2.75 billion, plus tiered sales royalties. The collaboration leverages Insilico’s Pharma.AI platform to accelerate the discovery of innovative therapies across multiple therapeutic areas. The deal size has triggered U.S. HSR antitrust review, marking a shift from proof-of-concept to commercialization in AI-driven drug discovery. Lilly’s strong endorsement of the AI platform validates the clinical potential of the candidate molecules.

Google has announced the release of Agent Development Kit (ADK) for Java 1.0.0, marking a major advancement in the ADK multi-language ecosystem. Key features include new tools such as GoogleMapsTool and UrlContextTool; local and cloud code executors; a centralized plugin architecture enabling global execution control and security policies; enhanced context engineering via event compression to manage context windows; Human-in-the-Loop mechanisms allowing agents to request human confirmation before critical actions; session and memory service interfaces supporting integration with Vertex AI and Firestore; and native support for the Agent2Agent (A2A) protocol enabling cross-framework remote agent collaboration.

Google co-founder Sergey Brin stated in an internal meeting that AI agents will be a key direction for the company’s future. The internal AI tool 'Agent Smith,' built on the Antigravity platform, can run asynchronously and automatically perform coding tasks—even continuing work when employees are offline. It can access employee profiles, retrieve documents autonomously, and be operated via internal chat platforms and mobile devices. Due to surging usage, access has been restricted. Brin hinted at ongoing development of tools similar to OpenClaw, and CEO Pichai is pushing to incorporate AI usage into employee performance evaluations. This trend reflects how AI agents are becoming central to automation strategies in tech companies.

According to OpenRouter data, global large model token usage reached 22.7 trillion last week (March 23–29), up 11.2% week-on-week. China’s weekly token usage hit 9.857 trillion, growing 33.94%, surpassing the U.S. (3.007 trillion, up only 1.79%) for four consecutive weeks. The global top four models are all Chinese: Xiaomi's MiMo-V2-Pro ranked first with 3.96 trillion tokens, followed by Jueyi星辰, MiniMax M2.7, and DeepSeek V3.2. China’s National Data Bureau revealed that the country’s average daily token usage has exceeded 140 trillion, more than a thousand-fold increase since early 2024.

According to DigiTimes, TSMC’s 3nm process capacity has entered a rare 'overload' state, triggering intense competition across the semiconductor supply chain. From GPU and CPU design firms to hyperscalers like Amazon and Microsoft, nearly all major players are rushing to place orders with TSMC. Actual capacity falls far short of soaring demand, and the severe imbalance is already affecting product roadmaps. The shortage not only drives up costs but also causes companies to hesitate in accepting downstream orders due to insufficient 3nm capacity. Meanwhile, MSI warns that RAM shortages are reducing GPU supply by 20%, as the three major DRAM manufacturers restrict output, leading to sharp increases in memory prices.