Domestically developed AI large models are gaining pricing power—will the industry chain undergo a r

钛媒体APP

joined discussion ·

【Digital Intelligence Weekly】ChangXin Technology Becomes the Largest by Market Cap on A-Share Market; ByteDance Restructures Feishu, Huoshan, and Doubao Divisions; Kimi K3 Goes Open-Source and Instantly Gains Popularity

【Digital Intelligence Weekly】will consolidate this week's most significant trends, major policies, and industry research reports in enterprise services, cloud computing, and big data.

DeepSeek-V4-Flash official API enters public beta

Unlike the previous tiered logic of 'Pro strong, Flash weak,' this update sends a noteworthy signal: the official Flash version now performs comparably to—or even surpasses—the V4-Pro preview version from three months ago across multiple Agent benchmark tests.

ByteDance and MiniMax released new models on the same day

On July 31, MiniMax launched its next-generation multimodal generative model, MiniMax H3, and announced it will be open-sourced soon. H3 is MiniMax’s first open-source multimodal generative model. The video generation pricing for MiniMax H3 is RMB 0.8 per second (at 2K resolution), only one-third the price of comparable flagship video models in the industry.

On the same day, ByteDance announced the release of its new video creation model, Seedance 2.5. According to the company, Seedance 2.5 builds upon the unified multimodal audio-video joint generation architecture of Seedance 2.0, with key advancements in long-form narrative capabilities, multimodal reference understanding, and editing functionality.

ByteDance restructures Lark, Volcano Engine, and Doubao organizational structure, integrating its ToB business

ByteDance issued an internal email announcing the integration of the Lark product team and the Doubao product team into a new Doubao product team, led by Zhao Qi, head of Doubao. Xie Xin, head of Lark, will now report to Zhao Qi. This marks ByteDance’s largest adjustment to its ToB business since implementing its business segmentation strategy in 2021. On the GTM (Go-to-Market) side, the Lark GTM team will merge with the Volcano Engine team to form a new ToB GTM organization called the 'Creativity Service Platform,' which will oversee marketing, sales, and customer service for ByteDance’s MaaS and SaaS cloud offerings. Tan Dai, head of Volcano Engine, will lead this new organization, with Lin Chan, head of Lark sales, and Shi Zhijun, head of Lark strategy and marketing, reporting to him. Following the reorganization, Lark’s existing products and services will remain unchanged, and Lark will deepen product collaboration with Doubao in productivity scenarios. It is reported that Doubao Enterprise Edition, co-developed with deep involvement from the Lark product team, has already entered closed beta testing with select Lark customers.

Moonshot AI’s Kimi completes Series F funding round exceeding USD 3.5 billion

Moonshot AI's Kimi has completed its Series F funding round, raising over USD 3.5 billion and increasing its post-money valuation to USD 35 billion. The round closed early as it exceeded its target by more than threefold. Its next round—Series G (Pre-IPO)—originally scheduled to begin in August, has already commenced, with a pre-money valuation now at USD 50 billion.

360 launches its next-generation enterprise AI agent workspace, NanoWork

Zhou Hongyi, founder of 360 Group, officially launched NanoWork, a next-generation enterprise AI agent workspace designed for business owners, entrepreneurs, solo companies, and employees. Zhou announced that each of the first batch of users will receive a trial allowance of 100 million tokens. Through hands-on training and guided support, the platform aims to help 1,000 small businesses successfully implement AI and improve operational efficiency. Zhou emphasized, 'For enterprises adopting AI, the boss must use it first.'

Kimi K3 goes open-source and instantly goes viral; National Supercomputing Internet urgently launches Token Plan

Following the open-sourcing of the Kimi K3 model, the National Supercomputing Internet AI community has officially launched related model files and API access services. Enterprises and developers can now quickly download the Kimi K3 model for deployment and development, or invoke it directly via a one-click API call. On its launch day, API request volume far exceeded the National Supercomputing Internet’s expectations. To meet diverse user needs, the platform has urgently introduced a monthly subscription-based Token Plan alongside its existing pay-as-you-go model, featuring granular credit-based billing and compatibility with mainstream terminals and AI programming tools.

Kimi K3 Officially Open-Sourced: The World's First 3-Trillion-Parameter Open-Source Model Goes Live

Moonshot AI has officially open-sourced the full model weights of Kimi K3, marking the world's first deployed open-source large language model with 3 trillion parameters. According to third-party platform Artificial Analysis, K3 maintains top-tier performance while its cost per task on the BrowseComp benchmark is only half that of GPT-5.6 Sol and nearly an order of magnitude cheaper than Claude Fable 5. Moonshot AI stated that anyone can download and deploy the Kimi K3 model—freely usable for internal R&D or embedded into end-user products.

Unisound: Upgrades to Speech Foundation Models U2-ASR and U2-TTS

Unisound announced in a Hong Kong Stock Exchange filing that it has recently completed comprehensive upgrades to its U2-ASR and U2-TTS capabilities. In this upgrade, U2-ASR added support for 13 new international languages, covering key overseas markets in Europe, Southeast Asia, the Middle East, and Latin America; U2-TTS added voice synthesis capabilities for eight additional Southeast Asian languages.

Tencent WorkBuddy is now officially available on the HarmonyOS PC app store

According to Tencent Cloud, Tencent WorkBuddy has officially launched on the HarmonyOS PC app store, becoming the first desktop office AI agent on the HarmonyOS platform.

MiniMax H3 Officially Launched

InfyTech has officially launched MiniMax H3, a general-purpose multimodal generative model capable of unified understanding of multimodal contexts composed of text, images, video, and audio. It can generate native stereo audio-video outputs, supporting up to 15 seconds at 2K resolution.

Jensen Huang launches X account with open letter on first day; 25 tech companies sign joint appeal supporting open models

On July 25 local time, NVIDIA CEO Jensen Huang officially launched his personal account on X (formerly Twitter) and posted his first message—a joint open letter signed by 25 tech companies including NVIDIA, Microsoft, and Meta, urging the U.S. government to support the development of open models. The letter argues that America’s leadership in AI should not be measured solely by whether it possesses the most advanced single model, but rather by whether it has built a robust, open ecosystem capable of permeating all industries. It states that open models enhance security defenses, accelerate innovation diffusion, foster market competition, and ensure AI benefits are widely shared rather than concentrated among a few corporations. The letter acknowledges that open models carry risks of modification or misuse but emphasizes that the solution lies in strengthening safety research—not restricting openness.

NVIDIA’s Market Cap Plummets Again, Overtaken Once More by Apple

NVIDIA’s stock price dropped sharply, causing its market capitalization to be overtaken again by Apple.

Hugging Face demands $100 million in computing power compensation from OpenAI

The open-source AI community Hugging Face was infiltrated by an OpenAI agent. The co-founder and CEO of Hugging Face has made two demands to OpenAI: first, to fully disclose the agent's activity logs, and second, to allocate $100 million worth of computing power to support the open-source community. On the 21st of this month, OpenAI acknowledged that one of its agents had 'gone out of control' during internal testing, breached its isolated environment, and autonomously infiltrated the Hugging Face community. The incident was ultimately resolved only after seeking assistance from Chinese open-source models.

AI cloud service provider Nscale advances IPO plans, targeting a potential valuation of $25 billion

According to reports, AI cloud service provider Nscale is advancing its initial public offering (IPO) plans and has already begun meetings with potential investors. The company plans to host an investor day event in New York next week to present its IPO strategy to analysts and investment institutions. Nscale could go public as early as this fall, though the IPO timeline may be delayed depending on market conditions. Investment banks previously estimated the company’s target valuation at approximately $25 billion, up from the $14.6 billion valuation it received during its March funding round.

Meta discloses nearly $700 billion in future spending commitments related to AI data centers and cloud computing

Meta announced that it has committed nearly $700 billion in future expenditures through long-term and short-term agreements for artificial intelligence data centers, cloud computing, and other areas. The company has signed irrevocable contracts amounting to $349.3 billion, primarily covering third-party cloud agreements, servers, and network infrastructure. Meta stated this figure is a conservative estimate, as for agreements with variable terms, 'we do not estimate total obligations beyond the minimum purchase quantities.' Additionally, the company has $347 billion in lease commitments that have not yet commenced and therefore are not reflected on its balance sheet. Of this, $68 billion was added in July alone, with related payments scheduled to begin in 2027 and 2028. These expenditures are in addition to existing leases and primarily involve data centers, colocation facilities, and 'specific network infrastructure.'

Anthropic admits its large model 'ran amok,' infiltrating three companies

On July 30 local time, Anthropic stated that due to a misconfiguration in its testing system, its Claude series AI models accidentally connected to the internet during cybersecurity tests and infiltrated the systems of three organizations. Anthropic reviewed more than 140,000 test instances to confirm whether Claude had accessed the internet from its test environment, which was supposed to be isolated from external networks. These tests included so-called 'capture-the-flag' evaluations, which task models with breaching other systems to obtain specified information—a common method for assessing AI models’ offensive cybersecurity capabilities.

Anthropic releases new model Opus 5

On July 24, Anthropic launched Opus 5. Claude Opus 5 maintains the same pricing as its predecessor, Opus 4.8, but delivers significantly improved performance, particularly excelling at software engineering tasks with real-world value. Over the past two months, Claude Opus 4.8, Fable 5, Mythos 5, and Sonnet 5 have all been successively released.

Anthropic's Texas data center is in talks to secure $15 billion in financing, with Google providing financial guarantees and chips.

According to reports citing informed sources, a data center developer partnering with Anthropic is entering advanced negotiations to raise $15 billion to build a large-scale data center campus and an accompanying power plant in Texas. Google will provide financial guarantees and chips for the project. Under the terms currently under discussion, a banking consortium led by Morgan Stanley will extend a $15 billion loan to Nexus Data Centers to construct the campus in Hubbard, Texas. The site will include an on-site natural gas-fired power plant capable of generating up to 1.6 gigawatts. The deal could be announced as early as today. Sources indicate that to support Nexus’s financing, Google has guaranteed Anthropic’s multi-billion-dollar lease payments and electricity obligations to protect against potential default by the startup. One source noted that Google’s guarantee is limited, covering only the minimum amount required by banks to finalize the financing. The guarantee encompasses four data center lease agreements signed by Anthropic, along with associated power purchase agreements, with electricity supplied by the on-site power plant serving the campus.

OpenAI reduces pricing for select GPT-5.6 models.

OpenAI announced optimized pricing and performance for its GPT-5.6 series models to enhance cost-effectiveness for enterprise AI applications. Specifically, the price of the GPT-5.6 Luna model has been reduced by 80%, and the GPT-5.6 Terra model by 20%. These price reductions will also apply to billing calculations for paid subscriptions used with Codex and ChatGPT Work. Additionally, OpenAI has introduced a 'Fast' mode in its API, replacing the previous Priority Processing service. In Fast mode, GPT-5.6 Sol can achieve processing speeds up to 2.5 times faster than standard mode, at twice the price of standard processing, with no change in model capabilities.

OpenAI acknowledges that the AI model breach incident involved multiple platforms.

An updated investigation report from OpenAI reveals that during the period when its AI model breached Hugging Face’s systems, it also leveraged publicly available online information to access accounts on multiple public service platforms.

Musk says Grok 4.6 will launch around August 7, followed by Grok 4.7 several weeks later.

Musk stated that xAI’s next-generation large model, Grok 4.6, is expected to launch around August 7, featuring a parameter count of 1.5 trillion and significant improvements in supervised fine-tuning and reinforcement learning. He added that Grok 4.7 will be released several weeks after Grok 4.6, with an estimated parameter count of 2.1 trillion. Compared to Grok 4.6, Grok 4.7 will offer further enhancements across the board but with slightly slower inference speed and higher token efficiency.

Morgan Stanley: In the inference era, generative AI could deliver incremental returns on capital of 25–50%.

A new analysis by Morgan Stanley estimates that in the AI inference era, incremental returns on capital from generative AI investments could reach 25% to 50%, significantly exceeding market expectations and addressing concerns about the difficulty of recouping massive AI-related capital expenditures. The report argues that scaled players such as Amazon, Google, Microsoft, and Meta will continue to benefit as inference demand surges.

AI drives a surge in software vulnerability discoveries, with Microsoft and Google reporting record-high security disclosures

According to data from the U.S. National Vulnerability Database, over 45,000 software vulnerabilities have been recorded since the start of 2026—nearing the total for all of 2025—and the number of vulnerabilities discovered this year could reach double last year's figure. Oracle, Microsoft, and Google have all disclosed record numbers of vulnerabilities in their recent security updates. Specifically, Oracle’s July security update patched 1,449 vulnerabilities, far exceeding the 309 reported during the same period last year; Microsoft disclosed 642 security flaws in July, nearly five times the amount from the same month last year; and Google Chrome’s recent update fixed 433 vulnerabilities, 401 of which were identified internally by Google. Security agencies note that AI tools are enhancing enterprise security teams’ ability to detect vulnerabilities, accelerating discovery rates to levels previously unattainable.

Trump administration’s voluntary AI regulatory framework nears implementation

According to reports citing informed sources, the Trump administration is finalizing a voluntary AI regulatory framework that would require artificial intelligence companies to submit their most advanced frontier models to the government for review before public release. Sources indicate that the White House Office of the National Cyber Director distributed draft versions of the framework to OpenAI, Anthropic, and Google about two weeks ago, and the three companies jointly submitted feedback. President Trump signed an AI-related executive order in early June, mandating completion of this regulatory framework by August 1. The framework aims to establish standardized procedures to govern next-generation, high-compute AI models capable of rapidly identifying cybersecurity vulnerabilities.

Anthropic clarifies its position: it has never advocated banning open-weight AI models

Anthropic responded to recent statements, clarifying it has never advocated for a complete ban on open-weight AI models. CEO Dario Amodei stated the company does not agree with the view that 'open-weight models inherently facilitate stronger AI safety mechanisms.'

Oracle integrates Google Gemini

On July 30, Oracle and Google Cloud announced an expanded partnership under which Oracle will integrate Google’s Gemini large language model across its full suite of enterprise applications. The two companies also plan to make Gemini available on Oracle’s AI Agent aggregation development platform, giving Oracle customers broader model choices and multimodal capabilities when building native agents—for example, using the ultra-fast, lightweight Gemini 3.1 or the more complex logic- and task-capable Gemini 3.5 Turbo.

Zuckerberg says the U.S. shouldn’t ban Chinese AI models just to stay ahead

Meta CEO Zuckerberg stated that the United States should not ban Chinese AI models solely to maintain an advantage in the AI race. His remarks come as some senior U.S. officials have publicly indicated plans to investigate alleged 'distillation' of cutting-edge American AI models by Chinese firms and potentially sanction Chinese companies on grounds such as 'stealing U.S. intellectual property.' Zuckerberg openly opposes such bans, arguing that U.S. companies should instead systematically identify and address bottlenecks and barriers to better compete with China in AI.

U.S. restricts imports of humanoid robots citing national security concerns

On July 28, the U.S. Federal Communications Commission (FCC) updated its Covered List, adding two new categories of equipment: one is 'advanced robotic devices' manufactured outside the United States—specifically humanoid robots, quadrupedal robots, and other mobile robots (hereinafter referred to as 'covered robots'); the other is grid-connected inverters produced outside the United States. Under these new rules, newly introduced models of covered robots will largely be prohibited from export to the U.S. from regions such as China. Previously FCC-registered legacy models of covered robots may continue to be sold in the U.S.; however, new models of advanced robots seeking entry into the U.S. market must apply to either the U.S. Department of Defense or the U.S. Department of Homeland Security. Only after these agencies determine that the products pose no 'unacceptable risk' will they receive 'conditional approval.'

CiYuan Infinite has completed its Angel++ round of financing, bringing its total raised capital to several hundred million RMB.

Beijing CiYuan Infinite Technology Co., Ltd., an enterprise-grade AI Agent infrastructure provider, announced the completion of its Angel++ round of financing. This round was led by Linxin Investment with participation from Huakong Fund. This marks the second financing round CiYuan Infinite has closed within a month, bringing its cumulative funding to several hundred million RMB. The proceeds will primarily accelerate the development of China's first fully autonomous, high-security, self-evolving enterprise-grade AI Agent infrastructure platform, deepen joint research initiatives with Tsinghua University and Beihang University, and further build next-generation infrastructure tailored for the Agent application paradigm.

ChangXin Technology Becomes the Largest Company by Market Cap on the A-Share Market

On July 27, ChangXin Technology made its debut on the STAR Market of the Shanghai Stock Exchange. By the close of trading that day, its share price had surged 465.82% to RMB 49.00 per share, with a trading volume of RMB 141.2 billion—surpassing East Money’s single-day record of RMB 90 billion set on October 9, 2024. At market close, ChangXin Technology’s market capitalization reached RMB 3.28 trillion, making it the new leader by market value on the A-share market.

Unitree Robotics to List on the STAR Market; Preliminary Bookbuilding on August 5, Subscription on August 10

On July 30, 2026, Unitree Robotics disclosed its 'Announcement on Arrangements for Initial Public Offering and Listing on the STAR Market and Preliminary Bookbuilding.' According to the announcement, Unitree plans to issue 40,446,434 new shares, representing approximately 10.00% of its post-IPO total share capital. All shares offered are newly issued, with no secondary offering of existing shares. Following the offering, the company’s total share capital will amount to 404,464,340 shares. The IPO aims to raise approximately RMB 4.202 billion, with China International Capital Corporation (CICC) acting as sponsor and lead underwriter.

132K Views