AI Frontier Daily
Meituan Open-Sources 1.6T Agent Model Trained Without Nvidia GPUs
Today's AI Intelligence Briefing: 7 distinct events rigorously curated from 5 cross-verified sources. Frontier models continue to advance hybrid reasoning and multimodal synthesis, while open-source tooling and autonomous agents accelerate real-world execution. Reports on the same event are merged into a single entry, and covered events do not reappear in later issues.
Meituan's LongCat team has open-sourced LongCat-2.5, a 1.6-trillion-parameter mixture-of-experts model designed for long-horizon multi-step autonomous agent tasks, activating only about 48 billion parameters per token with API support for million-token contexts.
The model was trained entirely on domestic hardware without any Nvidia GPUs, validating the feasibility of a domestic compute stack for ultra-large-scale model training.
Officials say it approaches the level of top Western models such as Claude Opus 5 from several months ago on multiple coding benchmarks, with weights now available for download.
Industry Ecosystem & Compute Economics
4 picksCapital deployment, compute supply chains, and industrial transformation.
Nvidia Board Ousts Jensen Huang as Silicon Valley Executives Petition for His Return
Nvidia's board has swiftly removed CEO Jensen Huang after an investigation, a decision faster than most observers expected. Huang's personal identity is tightly bound to the Nvidia brand, and his departure immediately raised concerns about the company's massive AI ecosystem investments.
According to the Financial Times, Huang has discussed returning as CEO, and allies are forming a Silicon Valley executive 'war room' to lobby the board and the White House with Huang's support.
Tech executives including Elon Musk have signed a petition urging Nvidia to reinstate Huang. Nvidia has not yet responded, and the situation remains in multi-party contention.
Nvidia Accused of Using Circular Financing to Keep GPU Prices High
The Wall Street Journal, citing a report on Nvidia interim CEO Colette Kress, says Nvidia pushed more deals in mid-to-late 2026 to absorb excess TSMC wafer capacity and maintain high GPU prices.
The report links long-standing circular financing concerns to governance changes after Jensen Huang's resignation as CEO, pointing to pricing manipulation disputes in the AI compute supply chain.
Nvidia reportedly offered financing to customers to drive deals and support GPU prices, highlighting potential risks in AI compute supply chain pricing mechanisms and circular financing models.
Firm Tied to Ex-Nvidia CEO Brokered $12 Billion China Chip Orders
According to documents reviewed by The Wall Street Journal, Megaspeed, a company linked to Alice Huang, former partner of ex-Nvidia CEO Jensen Huang, brokered nearly $12 billion in Nvidia high-performance chip orders for Chinese firms between March 2024 and June 2025, involving hundreds of thousands of export-controlled chips.
Megaspeed obtained priority allocations of scarce chips through its status as an Nvidia cloud partner and relationships with major clients such as Alibaba, suggesting much of the China business may flow through indirect channels.
Short-seller estimates suggest over 20% of Nvidia's revenue last fiscal year came from Chinese customers, with a significant portion potentially via indirect channels.
IN BRIEF
- ▪Stanford AI Index Reports Global AI Investment Tops $1 Trillion in 2026— Stanford HAI News
Agentic Systems & Autonomous Workflows
2 picksMulti-agent orchestration, tool usage, and complex task execution.
Northwestern and Google Unveil Self-Improving Agent Harness Method
A joint paper from Northwestern University and Google proposes Harness2, which lets agents automatically optimize their harness (prompts, tools, skills, and memory) on new tasks without ground-truth answers by trial-running candidate modifications and keeping effective changes.
A single-step improvement raised GLM-5 by 11.5 points on WorkBuddy-Bench, surpassing Gemini 3.5 Flash under its native harness.
However, harness edits without trial validation dropped coding scores from 73.4 to 63.2, showing that the validation step is indispensable.
DeepMind Turns Chip Simulation Waveforms into Queryable SQL for AI Agents
Google DeepMind and Columbia University propose BTTF, which compresses over 10 GB of chip simulation waveform dumps into an SQLite database, then uses a multi-agent team to translate natural language questions into SQL and validate design rules.
On 150 expert-annotated validation questions, the Gemini 2.5 Pro-based system achieved 95.33% accuracy, with multi-agent division outperforming a single agent on quality auditing by 17.3% at the cost of 1.96x latency.
The core lesson: for massive logs and trace data, put them in a queryable database rather than stuffing them into prompts.