At the 2026 Cloud Next conference, Google launched its eighth-generation TPU, splitting it for the first time into a training-dedicated TPU 8t and an inference-dedicated TPU 8i. TPU 8t delivers triple the training performance of the previous generation, while TPU 8i offers an 80% improvement in inference performance per dollar. CEO Sundar Pichai reaffirmed a full-year capital expenditure of $185 billion, double that of 2025, with over 50% of machine learning compute dedicated to cloud services. The company also introduced the Gemini Enterprise Agent Platform, Agentic Data Cloud, and a $750 million partner fund. Pichai revealed that 75% of new code at Google is now AI-generated, with AI models processing over 16 billion tokens per minute. Merck announced a partnership with Google valued at up to $1 billion.

On April 22, OpenAI released its new image generation model, ChatGPT Images 2.0, which topped the Arena.ai text-to-image leaderboard with a 1512 Elo score, leading Google's Nano Banana 2 by 241 points. The model achieves over 99% accuracy in text rendering, supports a thinking mode for pre-generation reasoning and web search, and can generate up to 8 coherent images per prompt at resolutions up to 2K. The API price for medium-quality 1024x1024 images is $0.053. DALL-E 2 and DALL-E 3 will be discontinued on May 12. The model excels particularly in dense Chinese text rendering and complex instruction understanding.

On April 22, Ant Group unveiled its Baoling large model Ling-2.6-flash, featuring a MoE architecture with 104B total parameters and only 7.4B active parameters. In Artificial Analysis evaluations, it completed tasks using just 15M output tokens—about one-tenth of comparable models—reducing inference costs by 86%. With four H20 GPUs, it achieves an inference speed of 340 tokens/s and output speed of 215 tokens/s. Previously launched anonymously as 'Elephant Alpha' on OpenRouter, its usage surged over 5,000% within a week, reaching hundreds of billions of tokens daily. API pricing is set at $0.1/million input tokens and $0.3/million output tokens. BF16 and other versions will be open-sourced soon.

Alibaba's Qwen team released Qwen3.6-27B, a 27-billion-parameter dense model that outperforms its own 397B MoE model on agent coding tasks. It scored 77.2 on SWE-bench Verified, 59.3 on Terminal-Bench 2.0 (matching Claude 4.5 Opus), and 87.8 on GPQA Diamond. Key innovations include a 'thought retention' mechanism that preserves reasoning chains across turns and a hybrid architecture combining Gated DeltaNet linear attention (75%) with traditional attention. Natively supporting 262K context length, it can scale up to 1 million tokens. Released under Apache 2.0, it provides BF16 and FP8 quantized versions, enabling deployment on consumer-grade GPUs.

On April 22, OpenAI introduced workspace agents, an upgraded replacement for custom GPTs, allowing users on Business, Enterprise, and Edu plans to create customized AI agents that run autonomously in the cloud. Powered by Codex technology, these agents can automatically generate reports, send Slack messages, draft Gmail emails, support team sharing and collaborative refinement, integrate multiple tools, and request approvals at critical steps. Future updates will allow existing GPTs to be directly converted into workspace agents. This launch is part of OpenAI's 'Release Week' events and positions the company in direct competition with Anthropic's Claude Cowork.

According to the South China Morning Post, despite ample funding, Chinese AI startup DeepSeek has initiated its first external fundraising round, planning to sell no more than 3% equity with a target raise of at least $300 million, valuing the company at no less than $10 billion—some reports suggest over $100 billion. The round aims to establish a valuation benchmark and retain core talent amid aggressive poaching by competitors. Investors include major state-backed funds such as AI-linked entities under the 'Big Fund Phase III'. DeepSeek prioritizes government-guided and industrial investors who can offer strategic resources like AI infrastructure over pure financial investors.

On April 23, Xiaomi announced the public beta launch of its MiMo-V2.5 series, with global open-sourcing imminent. The suite includes base, Pro, TTS, and ASR models. MiMo-V2.5-Pro implemented a complete SysY compiler in Rust within 4.3 hours (672 tool calls, perfect score of 233/233) and autonomously developed an 8,192-line video editor web application within 11.5 hours. MiMo-V2.5 is a native multimodal model supporting image, audio, and video understanding, with improved agent capabilities and approximately 50% lower API costs compared to the previous generation. The entire series reduces token consumption by 42%-50% compared to competing models.

AI startup NeoCognition has secured a $40 million seed round led by Cambium Capital and Walden Catalyst Ventures, with participation from Intel CEO Lip-Bu Tan and Databricks co-founder Ion Stoica. Founded by Ohio State University professor Yu Su, the company aims to address the current ~50% task completion rate of AI agents by building a 'world model' that enables agents to autonomously learn and continuously improve reliability within specific environments. The 15-person team focuses on the enterprise market and plans to embed its technology into SaaS products. The oversubscribed round reflects a shift in capital from foundational models to application-layer innovation.