Meituan's LongCat-2.0: A New Chapter in AI Sovereignty
In what may be the most significant development in China's push for AI self-sufficiency, Beijing-based tech giant Meituan on June 30 open-sourced LongCat-2.0 — a 1.6-trillion-parameter large language model trained entirely on domestic semiconductor hardware. The model was developed on a 50,000-card domestic computing cluster using AI ASIC superpods, and supports a context window of 1 million tokens.
What makes LongCat-2.0 genuinely historic is that it's the first trillion-parameter model to complete both full pre-training and inference on home-grown Chinese chips. Previous Chinese frontier models, including DeepSeek-V4-pro, used domestic chips only for inference while relying on Nvidia hardware for training. Meituan has crossed that line entirely.
The company open-sourced the model weights, echoing the strategy that propelled DeepSeek to global attention earlier this year. Performance-wise, LongCat-2.0 sits on par with DeepSeek's current flagship, which launched in April. For the U.S. chip export control regime, this is an uncomfortable data point: Chinese companies are demonstrating they can build competitive frontier models without access to Nvidia's latest GPUs.
OpenAI Previews GPT-5.6 Sol and Unveils Custom Jalapeño Chip
OpenAI had a landmark week with two major announcements. On June 24, the company unveiled Jalapeño — its first custom AI inference chip, co-developed with Broadcom in a record-breaking nine-month development cycle. The chip is specifically designed for inference workloads and delivers what OpenAI describes as substantially better performance per watt than current state-of-the-art. Initial deployment is targeted for late 2026, with plans to scale to gigawatt-level data centers alongside Microsoft.
Two days later, on June 26, OpenAI began a limited preview of its GPT-5.6 model family: Sol (the flagship), Terra (balanced for everyday work), and Luna (fast and affordable). Sol sets new benchmarks on Terminal-Bench 2.1 for complex coding workflows and introduces a new ultra mode that orchestrates sub-agents for complex tasks. However, access remains tightly restricted — only about 20 government-vetted partner organizations can currently use the models, with broader rollout expected in the coming weeks.
Together, the two announcements signal OpenAI's ambition to own the full stack — from silicon to frontier models — reducing its dependence on Nvidia and building a vertically integrated AI infrastructure.
Google Delays Gemini 3.5 Pro as Talent Exodus Deepens
Google's next frontier model, Gemini 3.5 Pro, has officially missed its June general availability deadline. The company is pushing the launch to July 2026, citing the need to refine coding performance, token efficiency, and long-horizon agentic capabilities based on early tester feedback. A Google spokesperson declined to comment on the revised timeline.
The delay comes at a particularly sensitive moment. Google has lost six senior AI researchers in five months, including Denny Zhou, who departed for Meta on June 30, and Gemini co-lead Noam Shazeer, who left for a competitor. In response, Sergey Brin circulated an internal memo urging the team to "bridge the gap in agentic execution" with Anthropic's capabilities, and Google has expanded its AI coding strike team to include the midtraining phase.
Meanwhile, the current Gemini 2.5 Pro with Deep Think, launched June 22, has posted impressive benchmark results — 82.4% on GPQA Diamond and 94.1% on HumanEval Plus — providing some competitive reassurance even as the next-gen model slips.
Alphabet Raises $84.75 Billion in Record AI Infrastructure Financing
In the largest AI infrastructure financing in corporate history, Alphabet raised $84.75 billion through a combination of a public offering, an at-the-market program, and a notable $10 billion private placement from Berkshire Hathaway. Warren Buffett's investment represents a significant strategic bet on Google's AI infrastructure at a time when the company faces talent challenges and model delays.
The massive capital raise underscores a broader industry trend: the AI race is increasingly becoming an infrastructure arms race. Between SoftBank's €75 billion commitment to French data centers, Amazon's custom AI chip business hitting a $20 billion annual run rate, and now Alphabet's record raise, the companies positioning to lead in AI are making bets measured in tens of billions.
EU AI Act Enters Final Countdown to Full Applicability
With just over a month to go, the EU AI Act is set to become fully applicable on August 2, 2026. The upcoming deadline brings into force the Act's transparency obligations under Article 50, including requirements that AI chatbots identify themselves as machines, that synthetic media (audio, images, video, and text) be clearly labeled as AI-generated, and that deepfakes carry mandatory disclosure labels.
Companies deploying AI in Europe must now ensure compliance with these transparency rules, though a grandfathering provision gives generative AI systems already on the market until December 2, 2026 to implement watermarking requirements. The European Commission has published draft guidelines and is developing a Code of Practice to help organizations navigate the new rules.
Across the Atlantic, the regulatory picture is more fragmented. Colorado's AI Act, originally set to take effect June 30, was significantly scaled back and delayed to January 1, 2027 after Governor Polis signed a replacement bill on May 14. The revised version drops the original risk-based framework in favor of narrower transparency and disclosure requirements — a sign that comprehensive U.S. AI regulation remains elusive even at the state level.