Open-Weight Model Developers

BMO said in a Friday note that concerns about artificial intelligence safety were valid, but leading AI labs showed little sign of slowing development. Citing its AI expert, the brokerage said fears among AI leaders about safety were genuine, given the importance of aligning AI systems with human interests before problems became irreversible, but it questioned how committed developers such as Anthropic and OpenAI were to pacing themselves, pointing to their continued releases of increasingly powerful models. BMO said recent AI-related hacking incidents were largely benign, though it warned that bad actors could increasingly exploit open-access models, noting users had shown an ability to override safeguards and prompt models into misaligned actions. The note followed Anthropic head Dario Amodei's call for a coordinated slowdown in frontier development, which several AI executives supported, with AI safety also discussed during a U.S.-China summit this week. BMO expects the U.S. and China to eventually tighten AI regulation, focusing on open-source and open-access models rather than closed systems, with such rules likely to require approval before broader public release and to strengthen pricing power for closed-source frontier labs including Anthropic and OpenAI. BMO said token prices were expected to keep falling, boosting usage of AI tools, but were likely to bottom out around 2030, after which closed-source providers would raise prices as users grew dependent on their tools. It said Meta Platforms' AI positioning had improved due to better use of its AI harness, though the company remained behind frontier leaders such as Anthropic and OpenAI, and it stayed bullish on AI hardware, citing rising scarcity of computing power. BMO also said white-collar job displacement was likely to become meaningful by 2028, with blue-collar jobs affected later in the 2030s, a delay it attributed to Moravec's Paradox.

Alibaba shares fell 4% to $112.09 on Wednesday after Bloomberg reported that Chinese regulators have opened a probe into AI startups DeepSeek and Moonshot AI over data security concerns, a move that attaches regulatory risk to the AI model layer where Alibaba's Qwen competes. Baidu eased 2% to $90.26 and JD.com fell 1% to $27.08, both shallower declines than Alibaba's, indicating the selling targeted AI model and cloud exposure rather than Chinese technology broadly. The reported probe began after accusations from Anthropic that the two startups had been routing sensitive data through its Claude models; DeepSeek, Moonshot AI and Anthropic are all privately held, and Alibaba is not named in the probe. The decline reverses part of the gain Alibaba shares had made earlier this week after its Apsara conference in Hangzhou, where the company unveiled a new line of AI chips it says are China's most powerful and built for frontier-model training. Alibaba Cloud Intelligence chief technology officer Li Feifei said Wednesday that Alibaba will set up its first cloud regions in Turkey, Finland and the Netherlands over the coming year and expand its data center footprint in Malaysia, Germany, the United Arab Emirates, France and Hong Kong, underpinned by a plan to build a 20-gigawatt global data center network by 2032. Leonid Mironov, a portfolio manager at Gavekal Capital, told Bloomberg that investors fear Beijing may heighten scrutiny of the industry and that sentiment has weakened because the scope of regulation remains unclear and may risk slowing China's model development.

Anthropic and OpenAI announced the launch of new, lower-cost artificial intelligence models amid fierce competition with rivals that tout open-weight models as their selling point. The launches mark the first for the two AI giants since industry experts began warning of the threat AI poses to humanity. The models unveiled include GPT-6 Sol and GPT-6 Luna from OpenAI and Claude Opus 5.5 from Anthropic. GPT-6 Sol and GPT-6 Luna are additions to OpenAI's GPT-6 family, with API prices cut by 50% compared with the promotional pricing of the GPT-5.6 generation. Sol is positioned below Astra and is designed for complex tasks such as coding, while Luna focuses on high-volume work such as data extraction or document summarization. Anthropic, meanwhile, introduced Claude Opus 5.5, the latest model in the Claude 5.5 family, saying it uses tokens more efficiently and costs about 40% less to run than Opus 5. The launches reflect efforts to meet customers seeking more cost-effective models in order to control AI spending amid pressure from Chinese AI companies led by Alibaba, Moonshot AI and DeepSeek, which offer cheaper open-weight models. Dianne Penn, Anthropic's head of product management, research and labs, said the company continues to work on making its thinking and answering processes more efficient so that fewer tokens are used for a given level of effort.

Chinese AI model developers' shares fell after a report that regulators have opened a probe into startups DeepSeek and Moonshot AI over data security concerns. Z.AI Co. and MiniMax Group Inc. shares declined as much as 7.4% and 6.3% in Hong Kong, respectively, following the Information report, which added that the probe began after Anthropic PBC accused DeepSeek and Moonshot AI of routing sensitive data through its Claude models. Alibaba Group Holding Ltd., which develops the Qwen model, slid 4.2%, while Xiaomi Corp. and Tencent Holdings Ltd. each fell about 2.5%. Gavekal Capital portfolio manager Leonid Mironov said investors are afraid Beijing may heighten scrutiny of the industry, with sentiment weakened by unclear regulation that may slow China's AI model development. The development adds to investor concerns about intensified scrutiny of Chinese AI firms amid growing debate over AI safety, ahead of an expected summit between President Donald Trump and Xi Jinping.

China's Cyberspace Administration has launched an investigation into Chinese startups DeepSeek and Moonshot AI over suspicions that state secrets were leaked to U.S. AI companies. The U.S. tech outlet The Information reported on the 22nd that Chinese authorities began the probe amid concerns over "distillation," in which the outputs of high-performance American AI models are used to train Chinese AI systems. According to a report published by U.S. AI startup Anthropic, DeepSeek and Moonshot sent users' questions to Anthropic's cutting-edge Claude model without authorization and used its answers to train their own models. The data involved also included internal information from the Chinese People's Liberation Army, public security authorities, and state-owned enterprises. The Trump administration and U.S. national security and investigative authorities have sharply condemned distillation. Ahead of the U.S.-China summit on the 24th, the issue of data security has come into sharp relief.