3Before9

Employees across the United States, India, Canada and Mexico woke on 31 March to termination emails from "Oracle Leadership" sent at 6 a.m. local time with no prior warning from managers or HR. Investment bank TD Cowen estimates the cuts hit between 20,000 and 30,000 staff - roughly 18 per cent of Oracle's 162,000-strong workforce - making it the largest layoff in the company's history. The restructuring is expected to free up between $8 billion and $10 billion in cash flow, money Oracle needs to cover an estimated $50 billion in capital expenditure on AI infrastructure this fiscal year alone. Oracle's total debt now exceeds $124 billion and its share price has more than halved since September 2025, wiping roughly $463 billion off its market capitalisation.

Five years ago almost no accident and health insurance claims at Ping An Insurance Group were processed without a human in the loop. Today nearly 60 per cent are fully automated, with some settled in as little as 51 seconds. The Chinese insurer, the country's largest private insurance group, is preparing to launch an AI-powered gateway in early April that links its 500 services across banking, insurance and healthcare. AI agents will handle underwriting and claims while boosting cross-selling across the platform, and the company's systems have been trained to recognise dozens of Chinese dialects. Ping An's overall headcount has dropped by 118,000 - roughly 30 per cent from its 2018 peak.

Alibaba's Qwen team released Qwen3.5-Omni on 30 March, a fully multimodal large language model that processes text, images, audio and video simultaneously and responds in real time across 36 languages. The model handles over 10 hours of audio input and more than 400 seconds of 720p video, with speech recognition covering 113 languages and dialects - up from 19 in the previous generation. On multilingual voice benchmarks, the top-tier Qwen3.5-Omni-Plus variant surpassed Google's Gemini 3.1 Pro in audio understanding, reasoning and translation, while matching it on audio-visual comprehension. New features include voice cloning from a single sample and semantic interruption, which lets the model distinguish between a filler "uh-huh" and an actual attempt to cut in.