AI & Compute Infrastructure — September 13, 2026

🇨🇳Zhipu AI (智谱; HKEX: 02513) completed approximately $5 billion in financing—$2 billion in shares and $3 billion in convertible bonds—for models and compute infrastructure. #

Zhipu AI (智谱), identified in the reports as Hong Kong-listed stock 02513, announced the financing’s completion on September 13. The approximately $5 billion total comprises approximately $2 billion in share-placement proceeds and approximately $3 billion from convertible-bond issuance, rather than a prospective financing facility.

The reported uses include next-generation GLM models, a fully self-training system, and associated compute infrastructure. Another announcement summary also identifies business expansion as a use of proceeds, but neither provides an allocation among these categories.

The infrastructure component is a stated use of funding, not a report of installed equipment or commissioned capacity. The supplied summaries do not identify GPU purchases, server quantities, data-center power capacity, deployment dates, or infrastructure counterparties.

🇨🇳Zhipu’s zero-coupon bonds priced at 100.5% of principal, with an initial conversion price of HK$892.50; placement shares represented 4.50% of enlarged capital. #

The share placement was priced at HK$714 per share, approximately 9.96% below the closing price preceding the announcement. The placement shares represented approximately 4.50% of enlarged issued share capital; that percentage describes the placement, not the potential additional dilution from bond conversion.

The convertible bonds carry a zero coupon and were issued at 100.5% of principal. The reported terms specify redemption at principal at maturity, distinguishing the issue price from the maturity redemption amount.

The initial conversion price is HK$892.50 per share, a 25% premium to the placement price and approximately 12.55% above the preceding closing price. The supplied announcement summaries do not give the maturity date, collateral provisions, conversion adjustments, or total potential conversion-related dilution.

• HK$892.50

🇨🇳Agricultural Bank of China (中国农业银行) signed an RMB8 billion indicative credit agreement with Jiechuang Intelligent (杰创智能); formal financing agreements remain pending. #

Agricultural Bank of China’s (中国农业银行) Guangzhou branch signed the strategic-cooperation indicative credit agreement with Jiechuang Intelligent (杰创智能) on September 12. The signing occurred during the branch’s ninth Qiushi Lecture Hall (求是讲堂) event of 2026.

The agreement specifies an indicative credit amount of RMB8 billion. The reported cooperation focuses on artificial intelligence and compute infrastructure, with the stated objective of supporting AI deployment at scale.

The report explicitly says Jiechuang must complete applicable compliance procedures before separately signing formal credit and related agreements. Consequently, the RMB8 billion is not reported as cash received or a drawn loan, and the source provides no interest rate, maturity, collateral, or drawdown conditions.

• RMB8 billion
• Jiechuang Intelligent (杰创智能)

🇺🇸Anthropic reportedly selected Nasdaq for its potential IPO, targeting October; the listing has not completed. #

A Telegram report citing Business Insider and unnamed sources says Anthropic selected Nasdaq for a potential initial public offering (IPO), with October as its target. The supplied material contains no offering prospectus, final pricing, share count, or completed listing announcement.

Separate reports describe financial information shared with investors during IPO preparations. One says second-quarter revenue exceeded $11.5 billion and that Anthropic achieved its first profitable quarter; a Financial Times summary more specifically describes expected current-quarter adjusted operating profit, excluding stock-based compensation and certain special or one-time costs.

Unnamed sources also reported a gross margin above 80% before revenue-sharing payments to distribution partners such as Amazon and model-training costs. Those exclusions are essential to the metric’s scope: the figure is not presented as a margin after those costs.

Anthropic reportedly identified improved compute efficiency as a central contributor to margin improvement. The same investor-material summary says the company cautioned that future quarterly profitability was not assured because of continued spending on frontier models and compute infrastructure; the current-quarter profit statement remains an expectation, not a completed result.

• Anthropic
• Nasdaq

🇨🇳AI-chip developer Enflame (燧原科技) listed on the Shanghai Stock Exchange’s STAR Market on September 11. #

A Telegram report attributed to China Securities Journal states that Enflame (燧原科技) listed on the Shanghai Stock Exchange’s STAR Market on September 11. The source identifies Enflame as an AI-chip enterprise and describes the listing as completed, rather than an application or planned offering.

The report places the listing alongside the China Securities Regulatory Commission’s (中国证监会) stated capital-market priorities for the 15th Five-Year Plan (十五五). Those priorities include adjusting listing standards and improving the fundraising, investment, management, and exit cycle for private-equity and venture-capital funds.

The supplied report does not provide Enflame’s ticker, offering price, proceeds, valuation, or post-offering ownership. It also supplies no associated chip-shipment, customer-contract, or compute-deployment figures, so those cannot be inferred from the listing.

• Enflame (燧原科技)
• September 11

Global server-market vendor revenue reached $166.3 billion in Q2 2026, increasing 52.0% year-over-year and 35.7% quarter-over-quarter. #

A Telegram summary of International Data Corporation’s (IDC) latest report puts worldwide server-vendor revenue at $166.3 billion in the second quarter of 2026. Revenue increased 52.0% from the year-earlier quarter and 35.7% from the preceding quarter.

The result exceeded the previous quarterly record of $125.3 billion in the fourth quarter of 2025. The reported measure covers global server-market vendor revenue, not GPU revenue, cloud-service revenue, or data-center construction spending.

IDC attributed the expansion principally to AI-infrastructure investment, alongside renewed shipment growth and rising average selling prices for both AI and conventional servers. The supplied summary does not quantify shipment growth or separate the contributions of pricing, AI servers, and conventional servers.

• Q2 2026

Infineon presented its case for using RISC-V in automotive processors for next-generation vehicles at Hot Chips 2026. #

ServeTheHome’s September 13 article summary reports that Infineon presented a case for using RISC-V in automotive processors at Hot Chips 2026. The stated application is next-generation vehicles, rather than data-center servers or cloud AI accelerators.

The summary refers to using RISC-V across various automotive processors. It does not identify particular processor products, automotive customers, or production programs.

The supplied source contains only the article title and description, not the presentation’s detailed technical content. It therefore supports coverage of the presentation, but not claims about core counts, performance, power consumption, safety certification, or commercial availability.

• Infineon
• RISC-V
• Hot Chips 2026

Copper-clad-laminate supplier Goldenmax International (金安国纪) denied NVIDIA and Huawei supply-chain certification claims; high-frequency products remain in research and customer sampling. #

Goldenmax International (金安国纪) said online claims that its products had obtained certification for inclusion in NVIDIA’s and Huawei’s (华为) supply chains were false. According to the announcement summaries, it had neither contacted the two companies nor undertaken any form of business cooperation with them.

The company also said it had not identified applications of its copper-clad-laminate products in AI servers or compute infrastructure. Its high-frequency, high-speed laminates remained in laboratory research and customer sampling, with no revenue generated from those products.

The clarification followed September 10–11 trading in which cumulative closing-price gain deviation reached 20%, triggering an abnormal-trading disclosure. Goldenmax additionally warned that its trailing price-to-earnings ratio exceeded the industry average.

A separate financing matter remained in progress: the company was advancing its 2025 plan to issue A-shares to specified investors. Regulatory registration had been approved, but implementation remained uncertain; the report does not describe that issuance as completed.

• Goldenmax International (金安国纪)
• NVIDIA
• Huawei

Guangdong Goworld (超声电子) denied supplying NVIDIA; printed circuit boards for 800G/1.6T optical modules remain under development without corresponding revenue. #

Guangdong Goworld (超声电子) said the market rumor that its high-frequency boards had passed NVIDIA certification was untrue. The company stated that it currently supplied no products to NVIDIA.

Its printed circuit boards (PCBs) supporting 800G and 1.6T optical modules remained under development. The announcement summaries explicitly say these products had not reached scaled supply and had generated no corresponding revenue; the speed designations concern the optical modules, not GPU or server compute performance.

The disclosure followed September 10–11 trading when cumulative closing-price gain deviation exceeded 20%. The company warned that the short-term price increase and turnover substantially above normal levels indicated risks of speculative trading and overheating.

Goworld reported normal operations, no material change in fundamentals, and no undisclosed material matters. It also said its controlling shareholder had not bought or sold company shares during the abnormal-trading period.

• Guangdong Goworld (超声电子)
• NVIDIA

Samsung Electronics and SK hynix reportedly rejected Korea Electric Power Corporation’s KRW25 trillion, five-year electricity-prepayment proposal, intended primarily for grid infrastructure. #

A Telegram report citing Yonhap and unnamed industry officials says Samsung Electronics and SK hynix rejected Korea Electric Power Corporation’s (KEPCO) proposal. The requested KRW25 trillion represented combined prepayment for five years of electricity bills.

The proposed split was KRW20 trillion from Samsung Electronics and KRW5 trillion from SK hynix. KEPCO reportedly said the prepaid funds would primarily support core electricity-grid infrastructure.

The chipmakers reportedly cited uncertainty in the industry outlook and the burden that large advance payments would place on their medium- and long-term operating strategies. They had communicated their rejection to KEPCO, according to the report; it describes neither a completed prepayment nor a funded grid-construction commitment.

• Samsung Electronics
• SK hynix
• Korea Electric Power Corporation
• KRW25 trillion

Baker Hughes secured Venture Global contracts covering 13 gas-compression systems and four liquefaction units containing eight modules. #

The supplied report says Baker Hughes secured work to provide gas-compression systems for a Louisiana pipeline project and modular liquefaction equipment for Venture Global’s Plaquemines facility. These are pipeline and liquefied natural gas (LNG) equipment awards, not data-center power-generation contracts.

The pipeline scope includes 13 gas-compression systems. The Plaquemines scope comprises four liquefaction units containing eight liquefaction modules in total, preserving the distinction between units and modules.

The liquefaction equipment is described as supporting additional LNG production capacity at Plaquemines. The source does not disclose contract value, delivery dates, incremental LNG output, or any allocation of that output to power plants or compute facilities.

• Baker Hughes
• Venture Global

The U.S. Environmental Protection Agency reportedly plans to repeal coal- and gas-power-plant carbon-emissions limits; final action remains pending. #

Reports say the U.S. Environmental Protection Agency (EPA) was preparing to repeal carbon-pollution standards for fossil-fuel power plants. Unnamed sources said an announcement could occur as early as Monday during the Group of Twenty energy ministers’ meeting in Houston; the supplied material does not contain a final rule.

The reported action concerns Biden-era requirements for existing and future fossil-fuel power plants to substantially reduce greenhouse-gas emissions over coming decades. The source identifies carbon capture and storage among the technologies contemplated by those requirements.

According to the detailed report, the EPA first proposed withdrawing the power-plant climate regulations in the preceding summer and was preparing to finalize part of that proposal. Officials also planned a separate rule to withdraw the finding that power-plant greenhouse-gas emissions independently constitute a threat under the Clean Air Act.

The report places this action after the agency’s reported withdrawal earlier in the year of analogous vehicle standards and the 2009 greenhouse-gas endangerment finding. It provides no quantified change in generating capacity, electricity prices, or data-center power availability.

• U.S. Environmental Protection Agency

Thermal-material developer Wuhan Jiyu Technology (武汉基域科技) launched Supercrystalline Film (超晶膜) materials and thermal-control components for applications including space computing. #

Wuhan Jiyu Technology (武汉基域科技) introduced its new thermal-control materials and components during a debut-product event at the 2026 Computing China Exhibition (2026 算力中国展览会). Founder and general manager Chen Pengfei (陈鹏飞) presented the product, named Chaojingmo (超晶膜), translated as “Supercrystalline Film” in the digest.

The announcement reports a thermal-conductivity figure of 2,200, with the source printing the unit as “W/m-1·k-1.” Because that notation is irregular, it should be checked against a technical specification before use in quantitative comparisons; the supplied report includes no test method or measurement conditions.

The product is presented for high-power devices in aerospace, energy storage, and 5G communications, with space computing identified in the report’s headline. A low-cost advantage is also claimed, but no price, comparison baseline, customer order, or deployed-system result is provided.

• Wuhan Jiyu Technology (武汉基域科技)
• Supercrystalline Film (超晶膜)

China Mobile Cloud (中国移动云) unveiled a hybrid inference system combining Lynxi (灵汐科技) neuromorphic chips with Iluvatar CoreX (天数智芯) GPUs. #

China Mobile Cloud (中国移动云) introduced the heterogeneous hybrid system for large-model inference at the 2026 Computing China Exhibition (2026 算力中国展览会). The presenter was Du Yujian (杜宇健), technical director of the company’s neuromorphic and optical laboratory.

The architecture combines Lynxi (灵汐科技) neuromorphic chips with Iluvatar CoreX (天数智芯) graphics processing units (GPUs). The report describes a mixed-chip inference system, rather than a new GPU product or an exclusively neuromorphic implementation.

The presentation positioned the approach against an NVIDIA-and-Groq solution, rendered in the source as “Vera Robin+Groq.” That source wording does not establish a precise comparison configuration, and the report supplies no chip counts, server counts, memory specifications, or commercial deployment schedule.

• China Mobile Cloud (中国移动云)
• Lynxi (灵汐科技)
• Iluvatar CoreX (天数智芯)

The hybrid system claims over twice comparable GPU-cluster performance and over 40% lower operating costs for DeepSeek V4 Flash inference. #

The reported performance and cost claims specifically concern DeepSeek V4 Flash inference. China Mobile Cloud’s presentation described a performance increase of more than 100% relative to comparable GPU clusters, alongside an operating-cost reduction exceeding 40%.

These are system-level comparisons for the named inference workload, not figures for an individual neuromorphic chip or GPU. The source does not specify whether “performance” measures throughput, latency, concurrency, or another benchmark metric.

The supplied message also omits the baseline cluster configuration, numerical precision, batch size, power consumption, and operating-cost accounting methodology. The figures should therefore remain attributed presentation claims, rather than independently established results applicable to other models or infrastructure configurations.

• DeepSeek V4 Flash

Health 160 (健康160) completed delivery and customer acceptance of an RMB12.2 million solution integrating health monitoring, compute infrastructure, and commissioning. #

A Telegram summary of Health 160’s (健康160) Hong Kong exchange announcement says the group completed delivery through a mainland subsidiary. The customer issued acceptance documentation for the RMB12.2 million integrated solution, making this a reported delivery-and-acceptance milestone rather than an intended market entry.

The customer is described as a healthcare-sector business focused on research and development of smart wearable health devices. The delivered package comprised an intelligent health-monitoring system, compute infrastructure, and integration and commissioning services.

The RMB12.2 million applies to the complete solution, not solely to the compute hardware. The report does not identify the customer, equipment specifications, GPU quantities, payment collection, or accounting recognition of revenue, and it does not establish a GPU-as-a-Service operating business.

• RMB12.2 million

Space-compute developer Dongfang Tiansuan (东方天算) signed Pearl Constellation (明珠星座) cooperation agreements with 27 industry partners, covering hardware, energy, communications, and software. #

Dongfang Tiansuan (东方天算) signed the agreements during the September 13 space-computing industry ecosystem forum at the Pujiang Innovation Forum in Shanghai. The Mingzhu Constellation (明珠星座), translated as “Pearl Constellation” in the digest, was formally launched as a program at the event.

Named partners included the Chinese Academy of Sciences’ Innovation Academy for Microsatellites (中科院微小卫星创新研究院), Shanghai Institute of Space Power-Sources (上海空间电源研究所), and Shanghai Institute of Satellite Engineering (上海卫星工程研究所). Commercial participants included Alibaba Cloud (阿里云), iFLYTEEK (科大讯飞), Biren Technology (壁仞科技), and Iluvatar CoreX (天数智芯).

The cooperation spans compute chips, space energy, inter-satellite communications, satellite platforms, computing payloads, and software services. The source reports signed industry-cooperation agreements but gives no monetary commitments, equipment purchase quantities, or binding capacity reservations.

The program’s stated long-term objective is a gigawatt-scale space-based computing constellation, with in-orbit validation of domestic computing payloads, optical-computing payloads, and new energy systems preceding initial satellite deployment and network expansion. Its objective of progressively establishing a globally covering space–ground compute infrastructure and service network by 2030 remains a program plan, not operational capacity; the gigawatt designation is not a computational-throughput figure.

• Dongfang Tiansuan (东方天算)
• Pearl Constellation (明珠星座)

Kailong High Technology (凯龙高科) deployed embodied robots commercially at Wuxi Institute of Commerce (无锡商业职业技术学院), within a mobile-service outlet. #

The supplied report says embodied-intelligence robots developed by Kailong High Technology (凯龙高科), identified by stock code 300912, were formally deployed in a Wuxi college-campus setting. It specifically describes commercial service deployment inside a mobile-service outlet at Wuxi Institute of Commerce (无锡商业职业技术学院).

The reported milestone is the deployment of the company’s robots in that service environment. The source does not describe a broader campus rollout or establish deployments at additional institutions.

No robot count, contract value, service-task breakdown, autonomy measure, or customer-use statistics are supplied. The report therefore supports the existence and location of the commercial deployment, but not conclusions about its revenue contribution or operating performance.

• Kailong High Technology (凯龙高科)

Microsoft plans to publish a code of conduct for its internally developed MAI models on September 14, inviting public comment. #

Microsoft chairman and CEO Satya Nadella said the company would publish a code of conduct for its first-party MAI models “tomorrow,” in remarks reported on September 13. The announced step is publication for public comment on September 14, not a completed consultation or a new model release.

Nadella framed Microsoft’s approach around human benefit and control, saying AI that does not help humanity and remain under human control is not worth developing. He also supported research and deliberate development pacing to make alignment a design objective, together with mechanisms such as embedded evaluators.

For enterprises, Nadella emphasized control over proprietary and tacit knowledge, continuous-learning loops, and the models and weights into which organizational knowledge is incorporated. He said organizations should be able to build these systems without dependence on a single model supplier.

Nadella also advocated broad access and choice throughout the AI technology stack, an ecosystem supporting both open- and closed-source models, and participation across countries, sectors, and academia. The supplied reports describe these principles but do not contain the forthcoming MAI code’s actual provisions.

• Microsoft
• MAI models
• September 14