China Securities Co., Ltd.: Domestic models accelerate iteration, computing power remains robust.

China Securities Co., Ltd. has released a research report stating that domestic models are accelerating iterations and the demand for computing power remains robust. Domestic models continue to evolve, with GLM-5.2 and Kimi K2.7 enhancing 1M context, long-range agents, Agentic Coding, and real-world engineering delivery capabilities, pushing domestic models from general Q&A towards developer tools and enterprise-level workflows. Kimi is strengthening international operational capabilities, DeepSeek's financing enhances expectations for the industrialization of leading models, and WeChat's AI gray testing indicates that the AI entry point is shifting from independent apps to a super application ecosystem, which is expected to increase model usage frequency and reasoning token consumption. Internationally, Anthropic's revenue, number of large clients, and financing scale continue to be revised upward, leaving the demand for enterprise workflows and Agentic Coding in a phase of expansion. On the side of computing power, AI cloud vendors are experiencing growth in backlog, GPU leasing prices are rising, and supply constraints in CPU, DRAM/NAND are spreading, with favorable conditions for computing power, storage, and leasing services expected to continue. Domestic models are continuously iterating, with long-range agents and engineering capabilities becoming the main focus this week. Z.AI launched GLM-5.2 on June 16, supporting 1M lossless context and enhancing scenarios such as complex systems engineering, deep debugging, and project-level context support; Kimi K2.7 Code was released and open-sourced on June 12, with improvements of 10.4%, 11.4%, and 76.2% in Program-Bench, MCP Mark Verified, and SWE Marathon respectively compared to K2.6, and reasoning token usage decreased by 30%. The competition among domestic models is shifting from general Q&A to long context, Agentic Coding, and real-world engineering delivery, with improved model availability expected to drive developer usage and enterprise trials. Domestic models are accelerating commercialization, and leading vendors are enhancing their global and content operational capabilities. According to reports from Game Value Theory and 36Kr on June 18, MiHoYo's former international president Jin Wenyi officially joined the "Dark Side of the Moon" about two weeks ago to lead the Kimi related business; Sina News cited Game Insights reporting that Jin left MiHoYo at the end of May and joined the "Dark Side of the Moon" in early June, where she will be responsible for Kimis international expansion and globalization. Jin joined MiHoYo in 2017 and built overseas distribution, localization operations, and overseas community systems while participating in the branding of HoYoverse. Kimi possesses long-context, multi-modal, and agent capabilities, augmented by experience in international gaming and content commercialization, with room for expansion in directions such as overseas subscriptions, AI content, AI tools, gaming industrialization, and enterprise services. Capitalization of domestic models is accelerating, with DeepSeek's financing strengthening expectations for the industrialization of leading models. According to a June 19 report from Sina Technology citing Fast Technology, corporate records show that on June 15, Hangzhou DeepSeek announced news related to its Series A financing of approximately 51 billion yuan, with a company valuation nearing 400 billion yuan; investors include Liang Wenfeng, Tencent, Contemporary Amperex Technology, NTES, JD.com, IDG Capital, and others. The financing capacity of leading domestic model companies is increasing, which is expected to further enhance R&D investment, computing power reserves, ecological cooperation, and the advancement of Agent commercialization. WeChat AI has entered gray testing, indicating that the domestic AI entry point is shifting from independent apps to a super application ecosystem. According to a June 20 report from IT Home, WeChat's native AI assistant "Xiao Wei" has begun to expand its gray testing range. Tencent customer service states that it supports operations of WeChat's native functions via text or voice, including adjusting settings, sending messages, making calls, ordering takeout, and generating images; on June 8, WeChat provided developers with the capability to access the WeChat AI ecosystem, allowing AI to call, access, and operate mini-programs, with both automatic and development modes available. With high-frequency ecosystem entry points such as social networking, payment, mini-programs, public accounts, and video accounts, the gray testing of Xiao Wei signifies the evolution of domestic AI applications from standalone chat entry points to task execution and transaction loops, which is expected to boost model usage frequency, reasoning token consumption, and the AI demand within the mini-programs ecosystem. The demand for overseas models continues to grow, with Anthropic's revenue and large customer metrics being revised upward. Anthropic disclosed on April 6 that its run-rate revenue had surpassed $30 billion, a significant increase from approximately $9 billion by the end of 2025, with the number of enterprise clients with annual spending exceeding $1 million increasing from over 500 in February to over 1,000; by mid-May, market tracking metrics by SemiAnalysis indicated that Anthropic's ARR had risen to about $44 billion; on May 28, Anthropic completed a $65 billion Series H financing with a post-funding valuation of $965 billion and disclosed in an official announcement that its run-rate revenue had surpassed $47 billion earlier in May. The simultaneous increases in revenue, number of large clients, and financing scale indicate that demand for Claude in enterprise workflows, developer tools, and Agentic Coding scenarios remains in a phase of expansion. Top-tier scientific and hardware talents are flowing into Anthropic, continuing the competitive extension of model vendors toward specialized scenarios and computational efficiency. TechCrunch reported on June 20 that John Jumper, the core leader of AlphaFold and 2024 Nobel Prize winner in Chemistry, is leaving Google DeepMind to join Anthropic; on June 9, DCD reported that Clive Chan, an early core member of OpenAI's in-house chip team, has joined Anthropic and previously participated in OpenAI's AI chip project in collaboration with Broadcom. The simultaneous influx of talent from the research and chip sides reflects Anthropic's ongoing intensification in areas such as life sciences, complex reasoning, chip adaptation, and inference efficiency. Leading model companies are continually increasing their investment intensity in computing resources and underlying infrastructure. China Securities Co., Ltd. believes that the acceleration of model commercialization will further strengthen the demand for computing power on the inference side. Internationally, Anthropic's revenue and number of large clients are rapidly rising; Claude Code, enterprise workflows, and Agentic Coding remain important directions for demand expansion. Domestically, Kimi is enhancing international operations, DeepSeek is advancing financing, and WeChat AI is entering gray testing, as model capabilities are accelerating into developer tools, super applications, content ecosystems, and transaction scenarios. Compared to the one-time large-scale investments on the training side, demand on the inference side more relies on user scale, calling frequency, task chain length, and multi-modal ratios. As 1M context, long-range Coding, parallel agents, WeChat mini-program calls, and AI content generation gradually come to fruition, the per-user computing power consumption and volume of concurrent calls are expected to continue increasing, benefiting GPU, CPU, storage, and computing power leasing services. The long-term supply of computing power continues to grow, and AI cloud vendors' orders are validating the certainty of downstream demand. NVIDIA's FY2027 Q1 data center revenue reached $75.2 billion, a 92% year-over-year increase, and the company continues to emphasize strong demand for Blackwell; CoreWeave's Q1 2026 revenue backlog reached $99.4 billion, a 284% year-over-year growth, and it signed an additional commitment of $21 billion with Meta in the first quarter. The rapid growth of orders and backlog from leading AI cloud vendors indicates that model vendors, cloud vendors, and corporate clients are still locking in computing power resources early. Storage prices continue to rise, and AI servers are occupying DRAM and NAND capacity. TrendForce predicts that in Q2 2026, conventional DRAM contract prices will rise by 58%-63% quarter-on-quarter, while NAND Flash contract prices will rise by 70%-75%; this judgment is based on the continued redirection of DRAM capacity toward servers and AI-related applications, with more NAND capacity allocated to enterprise SSDs. AI servers, HBM, high-capacity RDIMM, and enterprise SSDs are jointly squeezing capacity, shifting the storage chain from cyclical recovery to structural tightness, with ongoing price elasticity expected. Supply bottlenecks are spreading to CPUs, with server CPU prices rising alongside delivery times. TrendForce cited supply chain information on April 22 indicating that server CPU prices have risen by 10%-20% since March; Intel adjusted server CPU prices on April 1, with market expectations of an 8%-10% increase in the second half of the year; AMD server CPUs are expected to see cumulative price increases of around 16%-17% in the second and third quarters of this year. Additionally, the average delivery time for CPUs has been extended from 1-2 weeks to 8-12 weeks. AI inference and Agent workflows require extensive task scheduling, memory management, data handling, and multi-model coordination, which is expected to further enhance the value and supply constraints on CPUs. High-end GPUs continue to be in short supply, with B200 leasing prices and delivery cycles reflecting a worsening supply-demand imbalance. Reports indicate that Baseten CEO Tuhin Srivastava revealed at an Altimeter event that its cloud vendor had pre-notified in May 2026 that the leasing rate for B200 GPUs would increase from $2.63 to $5.10 per hour upon renewal in October, an increase of approximately 94%; delivery schedules for purchasing 1,000 GPUs have generally been extended to the second quarter of 2027, with waiting periods of 12-15 months. The report also cited the Bloomberg Index, stating that by the end of March 2026, B200 leasing prices had risen to $5.47 per hour, a quarter-on-quarter increase of 23.5%, with a cumulative rise of 48% in the first half of 2026; H100 leasing prices were $2.64 per hour in March, increasing by 8.6%, marking the fourth consecutive month of rise. China Securities Co., Ltd. believes that the current round of rising computing power prices is not solely due to the shortage of GPUs but is a result of the resonance in the industrial chain formed by the recovery in model demand, rapid growth of ARR, domestic model financing and internationalization progress, locked-in long-term computing power orders, rising GPU leasing prices, CPU price increases, and shortages of DRAM/NAND/enterprise SSDs. In the short term, there are supply constraints in high-end GPUs, CPUs, storage, cabinets, and electricity; in the medium term, Agent and multi-modal inference will continue to raise per-user computing power consumption.