Ant, Baidu, China Telecom and Zhipu Advance China's Enterprise AI Agent Push

Leiphone reported that at the forum on September 16, Shen Dou, president of Baidu Intelligent Cloud, said AI is still in an early infrastructure phase, comparable to the first month after the invention of the electric light. He calculated that global AI capital expenditure over the past year was about $500 billion, or 0.4 percent of global GDP, while investment and consumption in the power system are close to 5 percent of global GDP. Baidu divides enterprise tasks into three categories: about 20 percent are daily operations such as weekly reports, PowerPoint presentations and resource coordination; about 30 percent are professional tasks such as marketing, data analysis and coding; and the remaining 50 percent are core tasks involving proprietary production processes, process parameters, business data and know-how.

Baidu's response includes general agents, specialized agents and customized agents. Its general agent, Baidu Dazi, acts as an entry point; at CNPC, more than 5,000 employees have used it more than 700,000 times. Specialized agents include Miaoda, Shengsuan, Famou and Yijing. Auto design company Alte used Famou to run 120 autonomous iterations for wind resistance design, cutting hours of evaluation to seconds, and China Southern Power Grid used agents for alarm analysis and work order generation. Customized agents are supported by the Baidu Dazi development platform. Taikang has deployed more than 3,500 agents and consumes 20 billion tokens per day, including in compliance review.

Baidu also described an industry agent operating system connecting tasks, knowledge, data, permissions, tools and processes to cloud computing, storage and security resources. It said more than 150 models are available, agent long-chain inference performance has tripled in half a year, and its Agent Cloud handles task startup, permissions and traceability. Baidu's Kunlun chips support all-domestic clusters, and its Tianchi supernode has completed thousand-cabinet deployment. Baidu said data centers use standardized prefabricated modules, shorten construction by more than 30 percent and can reach an annual average PUE of 1.08. In Baidu's second-quarter results, AI cloud infrastructure revenue rose 50 percent year on year, GPU cloud revenue rose 283 percent, and GPU cloud has grown by triple digits for four consecutive quarters.

Leiphone reported on September 17 that Ant Group has connected Qwen Office as its company-wide intelligent office Agent base. Qwen Office uses Kubernetes private deployment, runs entirely on Ant's intranet, supports access to Ant's own models, keeps data within the domain, makes logs auditable and conclusions verifiable. It will provide intelligent office services to all Ant employees, integrate Ant's own Agent capabilities, digital employees, Ant Ding and project management, and connect with Ant's internal AI security control foundation. Luo Ji, president of Ant Group's platform technology business group, said Ant is advancing deep AI transformation of organization and office models and that the collaboration with Qwen Office will produce an integration model for large enterprises with complex office scenarios. Qwen Office, described as the industry's first enterprise-level general Agent product, had more than 30 million users one month after launch, with enterprise users accounting for more than half; customers include Changan Automobile, CIMC Enric, Huifu, Transfar Group and Laoxiangji.

According to ifanr, Zhipu GLM chief scientist Tang Jie said on X that GLM-5.3-Flash was optimized for inference on more than 100,000 domestic AI accelerators. It took two weeks from first run on the domestic accelerators to carrying all production traffic, and end-to-end throughput increased 3.2 times. An Infra Agent powered by GLM-5.3 participated in much of the optimization, which Tang described as a model helping optimize the service that runs itself. Zhipu said this is still far from full recursive self-improvement but an early form has appeared. The deployment faced limits in domestic chip memory and interconnect bandwidth, new model architecture adaptation, 1 million-token long context and multimodal requests. Zhipu used ReplaySSM to trade computation for memory, intra-node tensor parallelism, INT8, FP8 and BF16 mixed-precision caching, and an Encode-Prefill-Decode disaggregated architecture. It said end-to-end service performance improved about three times and hardware utilization and per-token cost approached mainstream NVIDIA GPU levels.

Zhipu said GLM-5.3-Flash ran anonymously as Ox-Alpha on OpenCode and OpenRouter and became one of the most used models on both platforms within a week, processing more than 62 trillion tokens in six days. The company developed a dense feedback mechanism to give the Agent information closer to engineering judgment. The Agent located a precision issue in the KDA context parallel path caused by TF32 rounding errors accumulating with sequence length; the fix was merged into Flash Linear Attention PR #1180. It found KV Transfer and DeepEP Dispatch did not effectively overlap, with transfer overhead exceeding 30 percent in some cases, traced the issue to a Python GIL release problem in the intra-node path, and reduced overhead to below 1 percent. It also reorganized a Decode Kernel that repeated normalization calculations four times, yielding a 1.71 times performance gain. Tang said human engineers still set goals, build feedback environments and review high-risk changes, but their role is shifting toward designing feedback systems.

QbitAI reported that IDC released China's first Enterprise General Agent Product Technology Assessment. Rather than using a conventional large-model benchmark, IDC used nearly 100 non-public tasks to test whether an Agent could complete work in real office environments, including email, calendars, documents, long context, multi-step loops, complex PowerPoint presentations, spreadsheets and browser operations. China Telecom's TeleAgent ranked in the top three. It scored 3.49 on regular tasks and 3.36 on complex tasks, received a full 5 points for task performance among nine capabilities, and had the highest cost-efficiency score in the assessment. TeleAgent's desktop version entered internal trial in April and V1.0 launched in July; its user base is now close to 1.2 million.

TeleAgent uses context management, autoDream long-term memory, a self-developed Go core of about 30,000 lines, compatibility work for Windows PowerShell, Kylin and UOS, and a ModelRouter that assigns tasks to lightweight, balanced or flagship models. QbitAI said the router can reduce inference costs by about 40 percent, keep simple task response times at 3 to 5 seconds and routing latency below 10 milliseconds. TeleAgent also uses PD disaggregation, KV Cache and load-aware scheduling. Its context window exceeds 400K. On security, Skill sources, viruses and vulnerabilities are checked; short-term and long-term memory are managed separately; file reads and writes are restricted to designated working directories; high-risk commands are monitored; and code and file operations are isolated where possible. It was among the first intelligent agent products to pass a China Academy of Information and Communications Technology security evaluation, passed 98.2 percent of China Telecom Research Institute internal security tests and 99.7 percent of 336 test cases across 28 scenarios with external security vendors. China Telecom's internal Skill plaza has listed 56,000 skills, added nearly 2 million times, with 20,000 employees contributing. IDC's April survey showed 48.5 percent of enterprises were evaluating, piloting or using general agents, and 96.9 percent of enterprise agent applications involved office automation.

Leiphone also reported exclusively that a Google Gemini pre-training expert recently joined a major Chinese internet company. The researcher is at Google level L7-L8, close to director level, has long been based in the UK, will work on pre-training and frequently travels to China. A person familiar with the matter described it as possibly the first time Chinese internet companies have found the right person in this round of competition for Gemini talent. Earlier hires from Gemini by several major companies did not all touch the core; some worked only on edge parts of pre- or post-training or in management roles that did not cover core pre-training. Since late last year, Chinese foundation-model teams have shifted attention toward pre-training as scaling of compute and data has run into the ceiling of the scaling law. Leiphone reported that choices in pre-training technical routes, data distribution strategies and first- and second-order optimizers can waste time and compute if wrong, and much of the experience is not published in papers. Silicon Valley AI pre-training researchers told Leiphone that Google deliberately keeps core pre-training talent in its DeepMind UK office, away from Silicon Valley's high exposure. Several people with DeepMind or Gemini backgrounds have joined Chinese companies, including Tian Yonglong at Tencent in July, Wu Yonghui and Qiao Siyuan at ByteDance, and Zhou Hao at Alibaba, but Gemini employees told Leiphone that many of those scientists came from Silicon Valley offices rather than the UK and did not truly master Gemini's core pre-training methods. The recent move from Silicon Valley to London has made the target more specific.