Hong Kong sto2026-10-20 05:19:41Hong Kong-listed large model stocks fall, with Zhipu slipping below HK$500Large model-related stocks in Hong Kong moved lower, according to market data cited by Odaily. Zhipu, trading under 02513.HK, fell more than 5% and dropped below HK$500. MINIMAX-W, listed as 00100.HK, declined 6.64%. The move affected the two companies described in the report as Hong Kong’s "large model duo." No additional details on trading drivers were provided in the source brief. The update was published by Odaily as a short market item based on quoted market data.40
AI Talent2026-10-12 09:44:10AI graduate pay surges as top Chinese tech firms reportedly offer up to 8 million yuan to new PhDsChina’s race for elite artificial intelligence talent is pushing entry-level compensation to unusual levels, according to a MarsBit report citing recruiters and market chatter. A widely shared post on social media said Nvidia offered a Peking University PhD graduating in 2026 a first-year pre-tax package of about 2.9 million yuan for an AI computing chip R&D role, including a signing bonus and performance bonus, with a chance to relocate to its Silicon Valley headquarters after two years. Recruiters quoted in the report said such numbers are no longer isolated. They described a market in which fresh PhDs from top schools such as Peking University and Tsinghua University can receive annual packages above 2 million yuan, with stronger candidates reaching 5 million to 6 million yuan and the very top tier climbing to 8 million or even 9 million yuan. The report also mentioned a 6 million yuan offer from Xiaohongshu to a master’s graduate in large models and an 8 million yuan offer allegedly from JD.com. The same pressure is showing up in embodied AI. Recruiters said compensation differs by specialty, with some embodied algorithm roles around 1 million yuan and world model positions generally above 1.5 million yuan. The report tied the salary inflation to rising AI spending by ByteDance, Alibaba and Tencent, and to the belief that top researchers can materially cut model training costs through better engineering and algorithm choices.460
Infinigence A2026-10-12 04:22:07Infinigence AI closes 2.3 billion yuan Series C, total funding reaches 4.3 billion yuanTsinghua-linked AI infrastructure company Infinigence AI has completed a 2.3 billion yuan Series C round, bringing its cumulative funding to 4.3 billion yuan, according to ChainCatcher. Founded in 2023, the company does not design chips or build its own data centers. Instead, it uses software to integrate computing power from different chip platforms, including Huawei Ascend, NVIDIA, and AMD, and offers large-model training and inference services to clients. Developers can access models directly through APIs without deploying them on their own infrastructure. Bloomberg reported on Oct. 1, citing people familiar with the matter, that Infinigence AI had confidentially filed for a Hong Kong IPO. The report said the company could go public as early as the first half of 2027, with a pre-listing valuation said to be about 14.3 billion yuan.260
OpenAI2026-09-28 08:44:09OpenAI and Anthropic Nearly Reached a Deal to Stress-Test Each Other’s ModelsOpenAI and Anthropic were close to signing a formal agreement earlier this year that would have let the two rivals test each other’s commercial AI models through API access, according to The Information. The proposed arrangement went beyond informal cooperation: lawyers reportedly drafted contract terms covering what could be tested, how testing would work, and limits on data retention. The talks also sat within a broader discussion about AI safety rules before training and before release, and about creating a standards body with participation from Google to audit frontier models. The report ties that push for outside scrutiny to a growing sense inside leading labs that internal monitoring is not keeping pace with model capability. It points to OpenAI’s delayed internal recognition of the July Hugging Face breach, details about autonomous behavior by OpenAI agents, and recent public warnings from Anthropic CEO Dario Amodei. Even if finalized, the proposed deal would only have covered commercial APIs, not unreleased systems such as the models cited in this year’s incidents. Both companies have since said they are willing to open more of their development environments to outside evaluators.400
Alibaba2026-09-22 04:15:38Alibaba says Qwen4 and next-generation video model are still in trainingAlibaba disclosed fresh progress on its large-model work at the 2026 Apsara Conference, according to a Techub News brief citing QbitAI. The company said its multimodal model Qwen4 is currently in training, and that its next-generation video model is also still being trained. The update was limited to that status report and did not include a launch date, technical specifications, or performance metrics. The announcement points to continued development across both multimodal and video-generation model lines inside Alibaba’s AI stack. Techub carried the item as a short conference update.490
Sun Wukong Ec2026-09-15 06:29:23Sun Wukong Ecosystem to Host X Space Roundtable on the Shift Toward AI AgentsThe Sun Wukong Ecosystem said it will hold an X Space roundtable at 8:00 p.m. on Sept. 15 focused on the next phase of AI competition as model capabilities continue to improve. The session is titled: "AI enters the Agent era: as model capabilities keep advancing, where will competition move beyond building stronger models?" According to the announcement, the discussion will center on how competitive dynamics are changing as foundational large models see broader reasoning capabilities. The event will examine the shift from comparing model parameters to competing on Agent-based automated execution, computing power orchestration, and on-chain interaction ecosystems. Invited industry KOLs will also discuss how Agents are moving from passive response to autonomous decision-making and execution, and what that change could mean for the connection between application layers and infrastructure.710
JD Cloud2026-09-09 05:28:59JD Cloud plans 100,000-GPU domestic computing cluster with Moore ThreadsJD Cloud said on Sept. 9 at the 2026 JD Global Tech Explorers Conference that it plans to work with Moore Threads to build a 100,000-card full-function GPU cluster, positioning it as a large-scale domestic AI computing infrastructure project. The cluster is set to focus on large model training, inference and embodied intelligence, with computing resources to be opened to users across industries. According to the company, the two sides will coordinate across chips, cloud platforms and model training to support iteration of JD’s JoyAI large model and speed up model training, simulation and high-quality data generation. The project is aimed at supply chain AI, embodied intelligence, integrated large-model training and inference, and industrial applications. JD Cloud said it had previously built a domestic 10,000-card cluster with Moore Threads and other partners, while Moore Threads said the planned AI factory would cover large model training, token production and agent training. Moore Threads also disclosed that its revenue in the first half reached 1.736 billion yuan, up 147.42% year over year, with cloud products contributing 1.693 billion yuan, or about 97.5% of total revenue.940
DeepSeek2026-09-08 13:22:36DeepSeek said to target a team of about 1,000 this year as hiring broadens beyond model researchDeepSeek is aiming to expand its workforce to about 1,000 people this year, according to BlockBeats, which cited a frontline institutional investor. The company has not publicly confirmed that figure, but its hiring push is already under way. In June, DeepSeek said every department would at least double in size, and this week it opened roughly 150 positions focused on backend and agent infrastructure. The latest hiring round shows the expansion is no longer centered only on model research. Open roles now span large-model research platforms, agent frameworks, APIs, online services, data engineering and elastic computing. The report also cited Cui Tianyi, head of the Harness team, who said data volumes, machines, training and evaluation workloads, agent environments and user requests are all rising quickly. That increase is putting pressure on DeepSeek’s existing backend architecture, which he said needs continuous upgrades and, in some cases, a full rewrite.820
