Chinese AI infrastructure funding surge: Zhiping completes ~50 billion yuan Series B; Keling AI secures ~3 billion USD; Anthropic launches Claude Science.
Model manufacturers worldwide have raised prices in recent months, primarily due to rising computational costs across the globe. Whether overseas or domestic chips, supply cannot keep pace with token demand growth. This assessment comes from Huang Zhenxin, Kimi's B2B business lead at Moonshot AI, during a recent media briefing. Explosive growth in downstream enterprise AI application needs is meeting upstream computing supply constrained by chip production capacity. Price increases across the API market are inevitable under this dual supply-and-demand squeeze, he argues. Addressing industry concerns about bubbles, he acknowledged their existence while noting that "Anthropic has already achieved quarterly profitability, with real, tangible returns." Bubbles exist, but fundamentals are sound. When productivity genuinely transforms, companies that can account for costs will continue investing—a sign of industry health.
**Meta CEO: AI Agent Development Has Not Accelerated "As We Expected" in Recent Months**
Meta Platforms CEO Mark Zuckerberg stated that AI agent development over the past four months has not accelerated "as we expected." The company's 2026 organizational restructuring "could have been cleaner." Investments made in the restructuring "have not yet yielded results."
Aburekemu (commonly known as Amu), vice president and chief strategy officer for China at Lenovo Group, recently outlined the "Three Laws of Token Economics." First is the "Law of Inertia"—the inevitable continuous decline in per-unit token costs, driven by three factors: chip, energy, and foundational model innovations; integrated compute-electric optimization and tuning; and runtime token scheduling. Second is the "Law of Acceleration"—the accelerating release of per-unit value with application depth. Amu emphasized the concept of "effective tokens," arguing that AI's value depends not just on models but on three additional factors: carbon-silicon fusion density, harness engineering depth, and AI governance maturity. Third is the "Singularity Law," which reveals the dynamic inflection point where total token costs and total token value reverse. As enterprise usage increases, Amu noted, total token costs will rise. Simultaneously, the value delivered by AI agents also rises and is released. When the curves of total token cost and total token value intersect, that is when AI value will explode.
Jiyi Intelligence recently launched in Changchun the embodied inspection robot "QC Home" for industrial quality control, the commercial service robot "Xiaomù," and its self-developed Z-1 embodied large model. The company took just eight months from founding to first product launch. "QC Home" has solved automotive wiring harness inspection, delivering three times human efficiency with over 30% total cost reduction. "Xiaomù" contains a database of over 100,000 pharmaceutical SKUs and has partnered with Jilin Pharmacy. The Z-1 embodied model ranks first in the RoboCasa open benchmark with an average score of 80.0%. Jiyi's parent company is Jieyi Technology, a precision automotive parts manufacturer, targeting tens of thousands of units in annual capacity by 2030.
OceanBase released a lake-warehouse unified AI database for the AI era, integrating data lake open storage, database transactional and analytical processing, and multimodal data handling into one strongly consistent data foundation, enabling agents to acquire complete business context in a single pass. The product system simultaneously covers the Lakebase underlying engine, DataStudio governance layer, and DataPilot business interface. It has been validated in scenarios including Ant Fu and Lingguang, with total cost of ownership reduced 30–50% compared to traditional multi-system approaches. CEO Yang Bing stated that AI database standards remain unfixed, and domestic vendors can shift from "followers" to "co-definers."
**Unbounded Power Releases First "Long-Sequence Bidirectional Physical Causal Chain" Latent Space World Model**
Unbounded Power formally released the world's first "long-sequence bidirectional physical causal chain" latent space world model—the MWA™ embodied general brain. The model operates through sequential chunk-level inverse dynamics modeling, outputting continuous multi-step latent action chunks to provide robots with a new latent-space sequential modeling approach for long-duration, coherent, and precise task execution across diverse complex scenarios. The model topped RoboCasa GR1 TableTop, the authoritative embodied intelligence leaderboard launched by top institutions including Stanford University, surpassing industry mainstream models including Nvidia's GR00T-N1.6, Daxiao Robot's ACE-EGO-0, Xiaopeng's DIAL, and Aode's ABot-M0.
The National Supercomputing Interconnect has launched the SCNet 10,000-Card Co-Creator Incentive Plan. Based on the scaleX 10,000-card supercluster, it offers global AI developers, startups, and research institutions up to billion-level token allowances and thousand-card-hour domestic computing resources. Centered on AI and AI4S core themes, it solicits best practices in domestic computing, paper readings, training livestreams, developer competitions, and AI application incubation co-creation projects, forming an AI innovation ecosystem of computing sharing, content co-creation, and brand co-building. This incentive plan runs continuously. By December 31, 2026, global users can submit project descriptions, computing needs, and co-creation directions through the supercomputing platform's dedicated registration channel. The platform will complete qualification review and project matching within five business days, and approved computing resources will be issued to user accounts.
Reports indicate that China Mobile, following the establishment of its Digital Intelligence Business Department and Computing Power Office, has also established a Token Office aimed at connecting the full cycle of "creating, delivering, and applying tokens" to break down silos where "each secondary department manages its own area." Alibaba confirmed the establishment of the Token Office internally, though other claims circulating online—such as "the Token Office holds higher rank and is directly led by group core leadership"—remain unconfirmed.
Multiple sources report that ByteDance plans to complete the design of its next-generation self-developed CPU no later than early next year, targeting mass production and large-scale deployment in the second half of 2027 to support its expanding AI infrastructure. One source noted that early versions of the CPU have been in internal use since late last year. However, another source suggested that due to urgent demand, the new-generation CPU's tape-out timeline may be accelerated. For ByteDance, this CPU is part of its broader strategy to build its own chip portfolio and integrate more self-developed hardware into its AI infrastructure. Two sources revealed that to accelerate R&D and secure wafer capacity, ByteDance is collaborating with Qualcomm, the American chipmaker actively expanding AI data center business. Working with Qualcomm, which uses a fabless model outsourcing manufacturing to Taiwan Semiconductor Manufacturing Company and other foundries, aims to help ByteDance address highly strained advanced chip supply chains.
Tencent Cloud announced that due to product business adjustments, after careful assessment, Cloud Mate will cease service on August 31, 2027 at 23:59:59. After service termination, the Cloud Mate console and related functions will no longer be accessible, with data retention, export, and recovery arrangements handled per the announcement and Tencent Cloud policies.
Meituan officially released its next-generation trillion-parameter large model LongCat-2.0 and open-sourced it. Trained end-to-end on a 50,000-card domestic computing cluster with 1.6 trillion total parameters and approximately 48 billion average activation across a dynamic range of 33B–56B, LongCat-2.0 features pre-training data exceeding 30 trillion tokens, covering Chinese, English, multilingual, and code data, with native support for 1 million token context length.
Om AI formally released the world's first edge-side streaming multimodal model series VLX for the physical world. The VLX series consists of three collaborating models forming a complete capability system for real-time physical intelligence. VLX-Flow handles continuous perception through incremental encoding and cached inference mechanisms; VLX-Seek provides precise localization, delivering reliable spatial perception for edge devices; VLX-Go handles action execution, converting visual understanding directly into robot-executable short-term waypoints and motion trajectories, enabling devices to independently complete following, obstacle avoidance, and navigation.
Huawei open-sourced the 92-billion-parameter openPangu-2.0-Flash model. The open-source openPangu brand aims to provide best-practice references for leveraging Kunpeng through native Ascend training and inference technologies.
Tencent Cloud announced that effective August 1, 2026, it will eliminate free quotas for "Cloud-Native Build-GPU" and "Cloud-Native Development-GPU." GPU resources remain available through pay-as-you-go usage at unchanged rates and pricing. For organizations with bound Tencent Cloud budgets: no further GPU free quota issuance; all usage is deducted from budgets at standard pricing. For organizations without bound budgets: GPU-related features require binding a budget first; proceed to organization settings—usage management—to bind your budget.
**Alibaba Cloud: Meoo Night Plan Now Live, Qwen 3.7-Max Discounted to Minimum 20% Off**
Alibaba Cloud announced that Meoo Night Plan is now live: nightly from 22:00 to 08:00 the next day (Beijing time), users selecting designated models on meoo.com's web interface will automatically receive off-peak pricing. Qwen 3.7-Max nightly discount reaches minimum 20% off; Qwen 3.7-Plus reaches minimum 40% off. Users need not change settings; logging into meoo.com during off-peak hours automatically applies discounts. Discounts affect price only; model service quality and response speed remain unchanged. Requirements: available only to Pro/Max paid subscribers; Free-tier users and those having purchased only credits are unsupported; coverage limited to meoo.com web interface conversations; Meoo CLI and AI service calls are excluded; requests outside off-peak windows and other models apply standard rates.
Zhang Shuyu, product manager at Tencent AI Assistant QClaw, posted on social media that she officially ended her role at Tencent on June 29. She joined Tencent in September 2025, serving for approximately ten months, and has not disclosed her next position. Zhang stated that she has already completed certain staged goals she set earlier and hopes to "solve one or more problems troubling humanity" next.
Sources within Alibaba revealed that due to recent disclosures of security vulnerabilities in Claude Code involving backend injection, Alibaba comprehensively assessed the tool and added it to its high-risk software list. Effective July 10, Alibaba will comprehensively prohibit internal employees from using Claude Code in office environments and recommends Qoder as an alternative.
According to Douyin platform announcements, due to product feature adjustments, the agent feature will be discontinued on July 15, 2026. After discontinuation, users can still view agent information and historical conversation data and save them for a period of time. Starting October 15, 2026, Douyin will process agent-related data per its privacy policy, after which it will no longer be viewable or recoverable within Douyin. For important content, advance backup via screenshots or export-text sharing is recommended. The agent feature allows users to create interactive AI characters on Douyin. Users can establish custom AI agents through basic settings, meeting work, life, and entertainment needs.
According to the latest announcement on the Chinese Academy of Sciences' ChinaXiv preprint platform, Huawei semiconductors head He Tingbo released on July 3 the V2 version of "Time-Shrink Theory for Multi-Level Electronic Systems" (also known in industry as the "Tao Law"). Compared to the V1 version released May 25, the new paper supplements the original theoretical framework with extensive engineering implementation details, measured quantization data, and product evolution roadmaps, further perfecting the post-Moore scaling theory system centered on the time constant τ. In engineering implementation, the V2 version deeply explains the core technology LogicFolding's gear ratio concept: when hybrid bonding pitch approaches top metal routing dimensions, 3D design space shifts from traditional "macro-block discrete optimization" to "unit-level continuous optimization," enabling globally optimal vertical logic division and breaking traditional 3D stacking's functional-block-only segmentation limits. V2 additionally provides mass-production measured data tables, explicitly giving Kirin 2026 and baseline Kirin 9030 Pro voltage, frequency, normalized power, area, and power density parameters.
The South Korean government stated it expects to construct four chip factories in the southwest, investing approximately 800 trillion Korean won; anticipated chip-sector investments over the next 15 years will reach at least 300 trillion Korean won, including next-generation memory, edge AI, and defense. Investment in the chip packaging cluster in the Chungcheong region is expected to reach 81 trillion Korean won.
On Sunday the 28th (local time), Musk posted on his X social media platform that his latest large language model Grok 4.5 has formally entered beta testing within SpaceX and Tesla before broader rollout. Musk further stated that SpaceX plans to release a "completely from-scratch trained" brand-new AI model monthly for the remainder of the year. These remarks suggest future versions may not be simple improvements to existing systems but wholly new foundational models built through new training.
Reports indicate Google has begun restricting Meta's access to its Gemini AI large model because Meta's computing demand exceeds Google's current capacity. Constrained supply directly disrupted schedules for multiple Meta AI projects, forcing R&D work to be postponed. Reports state Google's action directly exposes that computing supply remains the core bottleneck in AI industry development. Despite Google's continued increases in AI infrastructure investment, it cannot ensure sufficient computing capacity to meet surging market demand.
Semiconductor research institutions report that Jia Yangqing, AI veteran, former Alibaba Cloud VP, and Lepton AI founder and CEO, has left Nvidia, likely because the co-developed DGX Lepton AI supercomputing cloud service failed to meet Huang Ren-shun's expectations. In March last year, Nvidia acquired Lepton AI, publicly disclosing only a figure of hundreds of millions of dollars; latest reports claim the actual acquisition price was 700 million dollars. Lepton, founded in 2023, was established by Jia Yangqing, known as the "father of Caffe," after leaving Alibaba. Its primary business involves leasing Nvidia GPU servers and developing software to help startups build and manage applications in the cloud. Reports indicate Nvidia initially planned to open-source Lepton's core software before 2026, though it has not yet done so.
Reports indicate Samsung formally announced an investment plan totaling 265.5 trillion Korean won (approximately 1.168 trillion RMB), investing 203 trillion Korean won in semiconductor industry clusters in Yongin and Pyeongtaek cities. Previously, South Korea stated that Samsung Group and SK Hynix would announce major investment plans totaling 200 trillion Korean won (approximately 880 billion RMB) over the next decade, focused on semiconductors, AI computing data centers, and physical AI.
Reports indicate SK Group plans 110 trillion Korean won in medium-to-long-term investment to build a Yongin semiconductor supply chain, phasing in 100 trillion Korean won in 15GW AI data center capacity.
Reports indicate Millennium Management is establishing an AI laboratory to expand R&D and application of cutting-edge technologies. A memo indicates the new laboratory will begin operations within weeks. The laboratory will focus on accelerating early AI product acquisition and evaluation while collaborating with AI companies on projects and attracting top AI talent.
Google released its new self-developed AI image and video generation model Nano Banana 2 Lite, claiming faster speed and lower usage costs versus the previous generation. The model generates an image in four seconds with notably lower latency, suited for work requiring rapid iterative modification or bulk image generation within short timeframes. Generating one 1K image costs approximately $0.0336 (about 0.23 RMB); by contrast, standard Nano Banana 2's 1K image is approximately $0.067 (double Lite), and Nano Banana Pro's 1K/2K image approximately $0.134 (quadruple Lite).
Anthropic announced the launch of "Claude Science," an AI work platform for automating a range of biology and chemistry research tasks, including protein structure prediction. Anthropic stated the software integrates multiple tools scientists commonly use, including over 60 scientific databases, aiming to make it easier for scientists to automate multi-step tasks, ask questions in natural language, and obtain answers without querying multiple information sources individually. Anthropic said Claude Science will first launch as beta to enterprise paid users.
On June 30 local time, Intel held a groundbreaking ceremony for a new facility at its Bowers Campus in Santa Clara, California. Intel CEO Chen Lihe, Intel Foundry Services VP Naga Chandrasekaran, and Intel Mask Division VP Frank Abboud among other executives attended. The new facility, leveraging Intel's over 40 years of mask operations experience, will expand manufacturing capacity to support global future semiconductor innovation and production capacity.
Siemens announced a €300 million investment in Germany, significantly expanding production capacity for critical technologies supporting global energy transition and AI data centers. Investment includes constructing a new supplier facility in Offenbach, Germany, and expanding two existing factories in Frankfurt. Construction begins July this year, with the supplier facility expected operational spring 2027. The investment will create 700 new jobs by the end of 2030.
**Anthropic Includes Kimi K2.7 with Opus 4.8 and GPT-5.5 in Joint Safety Testing**
Silicon Valley AI giant Anthropic today issued an announcement removing export controls on its Fable 5 and Mythos 5 advanced models, and in simultaneously updated latest safety technical notes, includes China's Kimi K2.7 alongside Claude Opus 4.8 and GPT-5.5 in core safety capability assessments. The report indicates that in safety testing, Kimi K2.7, GPT-5.5, and Opus 4.8 successfully identified identical core vulnerabilities, while on demonstration tasks involving single vulnerability exploitation, Kimi K2.7 provided results consistent with Fable 5.
South Korean chip giant SK Hynix announced plans to invest 1 trillion Korean won in Cheongju, North Chungcheong Province, for massive expansion to ease chip shortages driven by the AI boom. Specific plans show SK Hynix will launch construction of the new foundry "M17" next year, planning 800 billion Korean won investment (approximately $5.146 billion) by 2029 for NAND flash chip production; simultaneously, the company plans to allocate an additional 200 billion Korean won by end of 2027, simultaneously constructing an advanced chip packaging factory in Cheongju.
According to reports, an internal memo indicates Tesla last month informed all employees that starting July 6 it is capping employee AI usage expenses at $200 weekly. This action sends a clear signal: even enterprises aggressively pushing AI technology implementation and attempting to reshape operations and product systems now must strictly control such costs. Previously, Tesla had been actively promoting various AI tools internally, encouraging employees to integrate them into daily workflows.
Reports indicate that three people directly familiar with the project revealed Anthropic has begun developing its own AI chips and negotiated with Samsung Electronics as a potential manufacturing partner, emulating competitor OpenAI's attempt to gain greater control over the expensive computing systems behind its models. Should the Claude maker continue the chip effort, it will be relatively new compared to other companies in developing proprietary AI server chips.
South Korea's Ministry of Science and Technology Information and Communications deputy director stated on the 3rd that the South Korean government is strategically reassessing construction of its autonomous AI foundational model. The government will advance large-scale projects centered on AI infrastructure, physical AI, and autonomous model capability building. The deputy director further noted that to rank among the world's top three AI powers, South Korea raised its AI budget this year to 9.9 trillion Korean won, triple last year. The South Korean government is planning to expand AI data center construction, expecting 8.4 gigawatts of data center capacity by 2029 and an additional 10 gigawatts by 2035. The government also set a target to lead globally in physical AI by 2030. Simultaneously, it will promote the "AI for All" initiative enabling all citizens to access AI services.
On July 2 local time, Amazon used United Launch Alliance rocket launches to place 29 satellites for its low-orbit satellite internet project into designated orbit. Chris Weber, Amazon Leo business and product VP, posted on X social media that with over 390 satellites deployed, Amazon has "completed enough launches" to launch preliminary internet service this year, with future missions further expanding coverage and scale.
Microsoft's AI programming assistant GitHub Copilot today announced formally integrating the Kimi K2.7 Code model, marking GitHub Copilot's first integration of an open-source model since launch.
Archon Robotics, a general-purpose full-body embodied brain company (also known as Yuance Robotics), completed a hundred-million-yuan seed round, with investors including top dollar funds True Fund, Gaorong Ventures, IDG Capital, Five Sources Capital, Gobi Ventures partnering with University of Hong Kong, QiJi Ventures, and Shanghai Innovative Institute for Technology, with Light Source Capital serving as exclusive financial advisor. Funding will primarily support full-body humanoid foundation model R&D, multi-modal full-body action data collection, talent team expansion, and multi-location R&D centers and industry partnership ecosystems, accelerating open-source humanoid base model deployment within the year.
Flexible embodied AI company Qingyu Robotics (Shenzhen) completed Pre-A round financing, with Shunwei Capital and Five Sources Capital as joint investors and Gaogulu Capital as exclusive financial advisor. The company, founded March 2025, enters through biomimetic flexible robot platforms, extending upward into embodied data flywheel and cross-embodiment foundation models, building universal embodied intelligence systems for multiple robot morphologies.
Moonshot AI issued a formal statement stating that recent market fraudulent information falsely citing the company has emerged regarding financing and equity transactions, with certain institutions and individuals suspected of fraud. Moonshot stated all financing activities are directly handled by the company, which has not hired or authorized any third-party institutions as financial advisors or financing consultants for any financing transaction, including new equity or secondary stock transactions. Secondary stock transfers must receive internal approval; unapproved secondary stock transactions are invalid, and the company will refuse any form of confirmation registration, reserving all legal remedies to pursue transferees' breach of contract liability.
General-purpose intelligent robot enterprise ZhiPingFang completed approximately 5 billion yuan in funding, with valuations surpassing 20 billion. This round investors span national-team capital, Greater Bay Area industrial capital, insurance companies, leading brokerages, and multiple Tesla supply-chain enterprises. Currently, ZhiPingFang has built semi-automated production lines with annual capacity exceeding 2,000 units, achieving routine batch delivery, with the domestic first tens-of-thousands-unit productivity embodied humanoid production line launch planned for the second half of 2026.
**Embodied "Superpower" ZiVariable Completes Four Rounds in Two Months, Post-Investment Valuation Exceeds 20 Billion Yuan**
Embodied "superpower" ZiVariable, following announcement of a B-round completion led by Xiaomi strategic investment in late April, consecutively completed B+, B++, and C-round financing, completing four consecutive heavy financing rounds with all closings within two-plus months. Investment roster includes China Mobile, Sequoia China, IDG Capital, Source Code Capital, Dachen Venture Capital, CITIC Capital, and over 30 top institutions, internet giant capital, industrial capital, national and local funds, and more.
Chinese quantum computing enterprise SpinQ announced completion of 1 billion yuan D-round financing, with new investors including CITIC Capital, Shanghai Semiconductor Industry Investment, China Aviation Honghua, Shanghai Electronics Science, China National Capital, Sichuan Xingchuan, Jinpu Investment, Hengxu Capital, Shang Qi Capital, Hongtai Fund, Luhua Investment, Xingzheng Investment, Shenzhen News, QingKe Holdings, Bojia Capital, and others, with prior stockholders including Yuncontrol Donghai and Jingkai Capital, Qingdao Hanrui, Hainan Fengkai Xiang, and others continuing to add capital, with senior industry investors Lu Dalong and Lankun also participating. Raised funds will be invested entirely in fault-tolerant universal quantum computing full-chain R&D...