🌎 Macro, Policy & Capital
🇦🇺AI-infrastructure developer Firmus signed a multiyear OpenAI partnership for Malaysian compute; its total contracted capacity now exceeds 900 megawatts. #
Firmus announced a multiyear strategic partnership under which OpenAI will purchase dedicated AI compute from two Malaysian AI-factory campuses and become an anchor customer. The supplied announcement report is dated September 8, although it is included here to preserve the September 7 digest’s coverage.
The reported more than 900 megawatts represents Firmus’s total contracted capacity following the partnership, not capacity attributed exclusively to OpenAI or Malaysia. It is a power-denominated infrastructure figure, rather than a count of graphics processing units (GPUs) or a measure of computational throughput.
Firmus said its portfolio comprises seven AI factories across Australia, Singapore, Indonesia, and Malaysia: two operating and five under construction. The company plans to bring the five construction-stage facilities into service over the following 24 months; this is a deployment plan, not an operating-capacity result.
The sources describe a concluded partnership but do not provide its contract value, exact duration, financing arrangements, or OpenAI’s share of contracted capacity. They also do not disclose a financing round or valuation associated with the announcement.
🇨🇳Memory supplier Longsys (江波龙; HKEX: 09976) priced H-shares at HK$236, with HK$6.8032 billion net proceeds; trading starts September 8. #
Longsys (江波龙; Hong Kong Stock Exchange: 09976) disclosed its offering allotment results on September 7, according to the Telegram reports of its exchange announcement. The final offer price was HK$236 per H-share, with reported net proceeds of HK$6.8032 billion.
The Hong Kong public offering attracted subscriptions equal to 40.32 times the offered amount, while the international offering attracted 3.88 times subscription. H-share trading was scheduled to begin on September 8, 2026; the supplied material reports pricing and allotment, not first-day trading performance.
A separate company announcement concerned its subsidiary Xizang Yuanshi (西藏远识), whose commitment to Suzhou Puhua Chenguang Venture Capital Partnership (苏州璞华宸光创业投资合伙企业) remained RMB80 million. The partnership’s total subscribed commitments increased from RMB371 million to RMB1.057 billion, reducing the subsidiary’s commitment share from 21.56% to 7.57% without changing its committed amount or contribution method.
That percentage reduction concerns the investment partnership, not dilution of Longsys shareholders through the H-share offering. The supplied offering summaries do not disclose the share count, shareholder dilution, use-of-proceeds allocation, or settlement conditions; the separate partnership announcement said the new limited partners had no related-party relationships or interest arrangements with the company and specified controlling parties.
- Longsys (江波龙; HKEX: 09976)
- HK$236
- HK$6.8032 billion
- September 8
🇨🇳Printed-circuit-board manufacturer Kinwong Electronic (景旺电子) passed its Hong Kong listing hearing and submitted post-hearing listing documents. #
Kinwong Electronic (景旺电子), a printed-circuit-board (PCB) manufacturer, passed its Hong Kong listing hearing according to a Telegram report citing exchange documents. Another message reported that it had submitted its post-hearing listing documents.
The named joint sponsors are CITIC Securities (中信证券), BofA Securities (美银证券), and Guolian Securities (国联证券). The supplied updates identify the listing milestone and advisers but provide no offering price, fundraising amount, or trading date.
This is a listing-process development, not evidence of a completed offering or receipt of proceeds. The source summaries also contain no share-issuance size, dilution calculation, customer concentration, or allocation of proceeds to manufacturing expansion.
- Kinwong Electronic (景旺电子)
🇪🇺The EU announced a €200 million Greenland investment plan for 2026–2027, covering connectivity, hydropower, critical raw materials, and other sectors. #
European Commission President Ursula von der Leyen announced the €200 million investment plan while visiting Nuuk on September 7. She also signed a joint declaration with Danish Prime Minister Mette Frederiksen and Greenland’s Prime Minister Jens-Frederik Nielsen concerning EU–Greenland relations.
The infrastructure components include improvements to satellite and fiber-optic cable connectivity, alongside support for hydropower and critical raw materials. The Commission’s reported timetable covers investment support during 2026 and 2027, rather than completed infrastructure expenditure.
The package also covers education, fisheries, housing improvements, sustainable tourism, small businesses, and youth exchanges. Greenland’s prime minister separately described the declaration as encompassing infrastructure, connectivity, maritime management, and civil preparedness.
The €200 million is the overall package, not a dedicated allocation to communications or energy alone. The supplied reports do not provide sector-by-sector budgets, generating-capacity figures, cable specifications, or a data-center project commitment.
🇨🇳China’s 2030 telecommunications plan targets 9,800 exaflops of AI computing capacity and 320 million gigabit-or-faster broadband subscribers. #
China’s Ministry of Industry and Information Technology released the 15th Five-Year Plan for Information and Communications Industry Development (信息通信行业发展“十五五”规划). The plan sets 2030 objectives across overall industry development, innovation, green development, infrastructure, and application adoption.
Its computing target is 9,800 exaflops, or exa-floating-point operations per second (EFLOPS), of AI computing capacity. The broadband target is 320 million gigabit-or-faster subscriptions; these are national planning targets, not reported current deployments.
Other stated 2030 objectives include RMB4.1 trillion in information and communications industry revenue, 50 fifth-generation mobile network base stations per 10,000 people, including 5G-Advanced, and 95% 5G user penetration, also including 5G-Advanced. The plan describes building a next-generation communications network with comprehensive coverage and leading performance by 2030.
A separate supplied news summary reported national AI computing capacity of 2.45 million petaflops at the end of July. Neither summary specifies numerical precision or a detailed measurement methodology, so no like-for-like growth calculation is presented here.
🖥️ Chips, GPUs & Systems
CXMT (长鑫存储) commercialized 16GB LPDDR6 memory packages in Xiaomi (小米) 18 Fold, with maximum transfer rates of 12,800 megabits/second. #
CXMT (长鑫存储) said its internally developed sixth-generation low-power double-data-rate memory, LPDDR6, had entered mass production and commercial use in Xiaomi (小米) 18 Fold. Its announcement described this as the first commercial deployment of LPDDR6 in a flagship smartphone worldwide; that distinction is a company claim reported through Telegram.
The memory has a 16-gigabit individual-die capacity and a 16-gigabyte packaged capacity, which are different scopes and units. CXMT said the product follows the JEDEC LPDDR6 standard and uses a 1,295-ball fine-pitch ball-grid-array (FBGA) package.
The stated maximum transfer rate is 12,800 megabits per second, while the reported rate when paired with the smartphone’s system-on-chip (SoC) is 10,667 megabits per second. The maximum memory rating therefore should not be presented as the handset’s reported operating rate.
CXMT also claimed 60% higher bandwidth relative to the cited 10,667-megabit-per-second LPDDR5X comparison and 20% lower power consumption than the preceding LPDDR5X generation. Both percentages carry asterisks in the supplied text, but their explanatory footnotes and test conditions are absent, so they remain qualified manufacturer claims.
- CXMT (长鑫存储)
- Xiaomi (小米) 18 Fold
Huawei (华为) launched Mate XT2 with Kirin 9050 Pro (麒麟9050 Pro) chips across the lineup, commercializing its logic-folding technology. #
Huawei (华为) introduced its Mate XT2 tri-fold smartphone at a September 7 launch event in Guangzhou. The report states that every configuration uses the Kirin 9050 Pro (麒麟9050 Pro) processor.
The launch report describes this as Huawei’s first introduction of a new Kirin chip at a flagship launch event in six years, following the Mate 40 global launch. A separate supplied daily summary lists the Mate XT2’s starting price as RMB19,999.
An Economic Daily (经济日报) commentary identifies Kirin 9050 Pro as the first high-performance chip using logic-folding technology (逻辑折叠技术) and describes improvements in performance and energy efficiency. Here, logic folding refers to the chip technology described by that source, distinct from the handset’s physically folding display.
The supplied material provides no quantitative processor benchmarks, fabrication-node specification, or architectural explanation sufficient to evaluate those improvements independently. The commentary itself cautions that progress in one chip does not establish comprehensive leadership across China’s semiconductor industry.
Xiaomi (小米) commercialized its Xring O3 (玄戒O3) smartphone processor; cumulative investment in chip development reached RMB 21 billion. #
Xiaomi (小米) said Xring O3 (玄戒O3) had entered mass-production commercialization with the launch of two Xring-equipped devices. At its September 7 autumn flagship event, founder, chairman, and chief executive Lei Jun (雷军) stated that cumulative chip-development investment had reached RMB21 billion, against a previously outlined RMB50 billion, ten-year research-and-development plan.
The Xiaomi 18 Fold combines the company’s Xring O3 processor, HyperOS 4 (澎湃OS 4), and MiMo on-device model. The launch report describes it as the first device to bring together Xiaomi’s internally developed chip, AI, and operating-system technologies, following more than 20 months of project development.
The handset starts at RMB10,999 and includes a 6,000-milliampere-hour Jinshajiang battery (金沙江电池), a stated single-side thickness of 5.02 millimeters, a weight of 219 grams, and support for up to six split-screen panes. A silicon-nitride ceramic special edition offers 16GB memory and 1TB storage for RMB15,999, with production limited to 1,500 units.
Lei also said Xring O100 (玄戒O100) and Xring D100 (玄戒D100) were planned for commercialization in the first half of 2027. Those are management’s future product plans, distinct from O3’s reported commercial status; the sources describe Xiaomi’s broader work across smartphone SoCs, AI accelerators, and intelligent-driving chips without providing detailed specifications for the two forthcoming products.
Broadcom (博通) generated $16.7 billion in fiscal third-quarter AI semiconductor revenue, increasing 221% year over year. #
A supplied Telegram summary of a China Securities (中信建投证券) research report states that Broadcom (博通) generated $29.59 billion in total revenue in fiscal 2026’s third quarter, up 86% year over year. Within that total, AI semiconductor revenue reached $16.7 billion, increasing 221% year over year.
The report attributes the AI semiconductor growth primarily to custom AI accelerators and high-speed networking demand. The revenue category therefore should not be treated as GPU revenue or accelerator revenue alone.
The same summary cites a fiscal fourth-quarter AI semiconductor revenue expectation of $21.7 billion, representing 236% year-over-year growth. That figure is an outlook rather than a realized result, and the supplied material contains neither Broadcom’s earnings release nor a securities filing to establish its full wording and assumptions.
⚡ Power, Grid & Data Centers
Zhiyuan Holdings (至源控股; HKEX: 00990) secured Malaysian approval for its AI data-center project, conditional on water and power arrangements. #
Zhiyuan Holdings (至源控股; Hong Kong Stock Exchange: 00990) announced that its AI data-center project received approval from Malaysia’s Ministry of Investment, Trade and Industry (MITI) on September 7. The supplied source is a Telegram report of the company’s Hong Kong announcement, rather than the approval document itself.
The approval carries water- and electricity-related conditions, which the company said its team was working to implement. It therefore represents conditional project approval, not confirmation that all utility arrangements are secured.
Management said it would accelerate preparations and seek to begin construction and production as soon as possible. The report does not disclose a firm construction date, commissioning date, megawatt capacity, hardware order, customer contract, or financing package, so these intentions should not be treated as deployment results.
Shanghai Electric (上海电气) won Malaysia’s Samalaju Unit 3 project, which will construct a 500-megawatt-class gas–steam combined-cycle generating unit. #
Shanghai Electric (上海电气) reported winning the Samalaju combined-cycle gas-turbine Unit 3 project in Sarawak, Malaysia. The project is located in Samalaju Industrial Park.
The stated construction scope is one 500-megawatt-class gas–steam combined-cycle generating unit. This is an electricity-generation capacity classification, not a data-center load commitment or a measure of AI computing capacity.
The source reports a project award, not commissioning or completed electricity delivery. It supplies no contract value, construction schedule, named electricity purchaser, or connection to the separate Malaysian AI-campus projects covered elsewhere in the digest.
- Shanghai Electric (上海电气)
- Samalaju Unit 3
Zaporizhzhia Nuclear Power Plant restored external electricity through its 330-kilovolt Ferroalloy-1 line, returning plant systems to normal operating mode. #
Zaporizhzhia Nuclear Power Plant announced on September 7 that external electricity supply had been restored through the 330-kilovolt Ferroalloy-1 (铁合金-1) line. The supplied reports attribute the update to the plant’s official communication.
Following reconnection, the plant’s systems switched back to normal operating mode. The 330-kilovolt figure identifies the line’s voltage, not its delivered power or the plant’s generation capacity.
The more detailed report says repairs were completed after a new ceasefire arrangement took effect on September 5. The update concerns restoration of incoming electricity and plant-system operation; it does not report reactor restarts or resumed electricity exports.
- Zaporizhzhia Nuclear Power Plant
☁️ Cloud & Compute Infrastructure
Firmus’s two Malaysian AI campuses will supply dedicated compute to OpenAI, with planned NVIDIA Vera Rubin NVL72 systems and liquid cooling. #
OpenAI will procure dedicated AI computing capacity from two Firmus campuses in Malaysia, according to Firmus’s reported announcement. OpenAI is described as an anchor customer under the multiyear partnership, rather than as the owner of the facilities.
The planned infrastructure includes NVIDIA Vera Rubin NVL72 systems, the NVIDIA DSX AI Factory platform, and Firmus’s internally developed HyperCube liquid-cooling infrastructure. These are named deployment components; the source does not give system counts, GPU counts, rack counts, or per-campus power allocations.
Firmus reported two operating facilities and five under construction across its seven-facility, four-country portfolio, with the construction-stage facilities planned to enter service over the following 24 months. The report does not identify which individual facilities are already operating or provide a separate Malaysian commissioning schedule.
OpenAI said the Malaysian data centers would help meet growing product demand in Asia-Pacific and globally. That is the customer’s stated purpose for the expansion, not evidence of compute already delivered or a quantified regional demand forecast.
- Firmus
- OpenAI
- NVIDIA Vera Rubin NVL72
Claude Fable 5.1 became available on AWS last week, supporting coding, scientific research, and long-running enterprise workflows. #
The September 7 Amazon Web Services (AWS) Weekly Roundup states that Claude Fable 5.1 became available on AWS the previous week. This is a model-availability update reported directly by AWS.
The excerpt attributes its capability description to Anthropic, which says the model supports ambitious tasks in coding, scientific research, and enterprise workflows. These are vendor-described capabilities, not independently measured results in the supplied material.
Anthropic describes the model as built for long-running, high-stakes work lasting hours and spanning multiple applications. The provided excerpt contains no pricing, region list, context-window size, throughput figures, hardware allocation, or benchmark results.
Shenhao Technology (申昊科技) has not finalized customers for its compute-rental business; the business has generated neither orders nor revenue. #
Shenhao Technology (申昊科技) president Gu Xiongfei (顾雄飞) discussed the business at the company’s 2026 interim-results presentation on September 7. He said discussions with downstream compute-rental customers remained ongoing and that final customers had not been determined.
The compute-rental business had generated no orders and no revenue contribution, according to his statement. This explicitly separates customer-development activity from contracted business and recognized revenue.
Gu also described the company’s overall order book as ample and its new-order trend as favorable, but those company-wide comments do not establish orders for compute rental. The source provides no GPU inventory, deployment capacity, rental pricing, capital expenditure, or utilization data for the business.
- Shenhao Technology (申昊科技)
🤖 AI Models, Software & Research
China Mobile Cloud (移动云) launched AutoLab for Quantum, enabling natural-language quantum-circuit construction, noisy simulation, quantum-neural-network training, and experiment reproduction. #
China Mobile Cloud (移动云) launched AutoLab for Quantum on September 7 for universities and research teams. The service is presented as a one-stop research agent through which users describe experimental objectives in natural language rather than undertaking extensive coding and environment configuration.
The announced functions include quantum-circuit construction and visualization, noisy simulation, Grover search, and quantum-neural-network training. It also supports experiments using the variational quantum eigensolver (VQE) and quantum approximate optimization algorithm (QAOA).
Additional capabilities include pulse-level simulation and optimization, literature retrieval, and experiment reproduction. The source does not specify quantum hardware access, processor or simulator capacity, supported qubit counts, pricing, or validation results, so the launch should not be described as a new quantum-computing hardware deployment.
GPT-6 Astra completed all 48 levels of an “I’m not a robot” CAPTCHA game without errors in a released demonstration. #
A Telegram report citing foreign media says OpenAI researchers released a video showing GPT-6 Astra completing all 48 levels of an “I’m not a robot” (我不是机器人) CAPTCHA game without errors. The report emphasizes completion of the entire sequence rather than an isolated stage.
CAPTCHA stands for Completely Automated Public Turing test to tell Computers and Humans Apart. The source describes these tests as presenting challenges intended to be straightforward for people but difficult for machines.
The reported result is limited to the released game demonstration. The supplied source contains no repeated-trial results, independent evaluation, timing data, or evidence establishing performance against production CAPTCHA services, so it does not support a universal CAPTCHA-bypass claim.
