AI News: July 1, 2026

1. Meituan’s LongCat-2.0 Trains a 1.6T Model Entirely on Chinese Chips

Meituan. The Chinese technology company said it trained LongCat-2.0, a 1.6 trillion parameter model, on a cluster of more than 50,000 domestically manufactured AI ASICs over 35 trillion tokens, without Nvidia hardware. Meituan reports the model outperforms some leading Western systems on benchmarks such as SWE-bench Pro and multilingual tasks while lagging on others, though it was not yet available for independent verification at publication. The result signals that Chinese teams can now train competitive trillion-parameter models on domestic silicon despite US export controls. Source

2. Wayve Runs $85M Employee Tender at $8.5B Valuation

Wayve. The UK autonomous driving startup launched an $85 million tender offer letting staff sell vested equity to investors at the $8.5 billion valuation set in its February 2026 Series D. That $1.2 billion round drew Eclipse, Balderton, SoftBank Vision Fund 2, Ontario Teachers’ Pension Plan, Baillie Gifford, Microsoft, NVIDIA, and Uber, and the company now employs about 1,200 people. Wayve, which uses end-to-end neural networks rather than HD maps, plans robotaxi pilots with Uber this year and integration into Nissan driver-assist systems from 2027. Source

3. Nvidia Challenger Etched Reaches $5B Valuation and $1B in Booked Sales

Etched. The inference chip startup said it has booked $1 billion in orders for full frontier inference clusters after manufacturing its first chip through TSMC, at a $5 billion post-money valuation and roughly $800 million raised to date. Etched sells bundled systems combining specialized chips, custom racks, and software aimed at running frontier models faster, cheaper, and more power efficiently than general-purpose GPUs. Backers include Jane Street, Hudson River Trading, Two Sigma, Ribbit Capital, Stripes, and angels such as Andrej Karpathy, Geoffrey Hinton, and Fei-Fei Li. Source

4. Ex-DeepMind Poker Researchers Raise Series A at $500M for Trading AI

EquiLibre Technologies. The Prague-based lab founded by former DeepMind researchers Martin Schmid, Rudolf Kadlec, and Matej Moravcik, creators of the DeepStack poker AI, reached a $500 million valuation in a Series A led by Creandum, described as the firm’s largest single investment. EquiLibre applies reinforcement learning to algorithmic trading and, in partnership with Tower Research Capital, says its agents trade billions in daily volume across the S&P 500 and Nasdaq. The team claims zero negative months since inception across crypto and equities. Source

5. Taiwan Raids Super Micro Offices in Nvidia Chip Smuggling Probe

Taiwan authorities. Investigators raided Super Micro Computer offices and affiliated companies, including distributor Albatron Technology and data center operator Chief Telecom, over alleged smuggling of Nvidia AI chips to China through Japan using forged export documents. A Super Micro co-founder was indicted, though the company itself faces no charges, and its stock fell 8 percent in US trading. Taiwan does not currently criminalize AI chip exports to China but is weighing alignment with US restrictions, making the probe a signal of tightening enforcement. Source

6. DeepSeek and Peking University Release DSpark for Faster Inference

DeepSeek. DeepSeek and Peking University published DSpark, an open-source inference framework that combines speculative decoding with a confidence-based verification system that adjusts depth based on compute load. The team reports 60 to 85 percent faster per-user response times and throughput gains up to 661 percent in some configurations, released under MIT license alongside DeepSeek-V4-Pro on Hugging Face and GitHub. Faster inference lets teams extract more performance from fewer high-end chips, a meaningful edge under tight supply and export controls. Source

7. California Signs Deal to Use Claude Across State Government at Half Price

State of California. Governor Gavin Newsom announced a partnership giving California state agencies and local governments discounted access to Anthropic’s Claude, with training and support included, for tasks such as document drafting and information analysis. The deal aligns with Newsom’s March executive order on responsible government AI use and marks a divergence from federal policy, which labeled Anthropic a supply-chain risk and chose OpenAI for Defense Department contracts earlier in 2026. It illustrates states carving out independent AI procurement paths. Source

8. Report Says Meta Contractor Secretly Tested Rival Chatbots With Minor-Perspective Prompts

Meta. WIRED reported that Meta contractor Covalen ran a project called “Cannes” in which contractors posed as minors and sent more than 45,000 prompts in a single August 2025 round to ChatGPT, Gemini, and Character.AI, covering self-harm, eating disorders, and drug use, with responses logged into spreadsheets. The tested companies said they were unaware: Character.AI called it a terms-of-service violation, OpenAI opened an investigation, and Google said it did not approve the tests. Meta called the work responsible, industry-standard safety testing and denied using the data to train its models, with testing reportedly active through at least April 2026. Source

9. OKX Launches Marketplace for AI Agents to Hire and Pay Each Other

OKX. The cryptocurrency exchange launched OKX AI, a marketplace letting AI agents autonomously hire one another, settle payments in stablecoins, and build portable on-chain reputations using digital wallets and persistent blockchain identities. Developers integrate agents through Onchain OS, OKX’s toolkit that the company says is compatible with Claude Code, Codex, Hermes, and OpenClaw, with access to its fraud detection and compliance systems. OKX frames the launch as infrastructure for an emerging agent economy it expects to reach trillion-dollar scale within five years. Source

10. Proton Upgrades Its Privacy-Focused Lumo Chatbot

Proton. The privacy company released Lumo 2.0, adding image recognition and generation, user-controlled persistent memory for Projects, a new thinking mode for complex queries, and 76 percent faster responses. Proton says the chatbot uses zero-access encryption in transit and at rest, does no server-side logging of conversations, and never uses customer data for training or shares it with third parties. The update positions Lumo as a functional alternative for users unwilling to trade privacy for capability. Source