AI & Compute Infrastructure — August 23, 2026

🇨🇳Alibaba priced its HK$80 billion placement at HK$112.70, with completion expected August 26; proceeds will fund full-stack AI infrastructure. #

Alibaba Group Holding (09988.HK; BABA.N) priced 710 million new ordinary shares at HK$112.70 each, for a placement totaling approximately HK$80 billion, or about US$10.2 billion. The placement had been priced but was not yet completed; completion was expected on August 26, 2026.

The shares were offered to non-U.S. persons outside the United States, and Alibaba described the transaction as its first new-share placement since its 2019 Hong Kong listing. China International Capital Corporation (CICC), HSBC, Morgan Stanley and UBS participated in the placement.

Alibaba said 100% of the net proceeds would be invested in its full-stack artificial-intelligence capabilities, including expanding and upgrading AI infrastructure. The disclosed transaction creates 710 million new shares; the supplied sources did not provide a post-placement ownership percentage or other dilution calculation.

Demand reportedly included sovereign wealth funds from the Middle East, Europe and Asia. Reports that the order book was oversubscribed describe investor demand, while the placement’s legal completion remained scheduled for August 26.

• Alibaba
• HK$80 billion
• HK$112.70
• August 26

🇺🇸Hugging Face is exploring a sale after receiving interest implying a valuation of at least $13 billion; no transaction has been reached. #

Hugging Face, an AI model-hosting and open-source community platform, has received merger-and-acquisition interest associated with a potential value of at least US$13 billion. The reports describe inbound interest and a sale exploration, not an agreed acquisition.

The company is reportedly working with a bank to assess interest from potential bidders. No prospective buyer was identified in the supplied reports, and no definitive agreement, purchase price or transaction structure was disclosed.

The transaction therefore remained exploratory as of the report. The US$13 billion figure represents the value implied by reported acquisition interest rather than the consideration under a signed or completed sale.

• Hugging Face

🇯🇵Japan’s Green Transformation Acceleration Agency invested ¥20 billion across fusion developer Kyoto Fusioneering and recycler JEPLAN. #

Japan’s Green Transformation Acceleration Agency invested a combined ¥20 billion, reported as approximately US$125 million, in two startups. The recipients were fusion-technology developer Kyoto Fusioneering and plastics-recycling company JEPLAN.

The source did not disclose how the ¥20 billion was divided between the two companies. It also said detailed investment terms were expected to be announced soon, so no ownership percentages, valuation terms or individual funding amounts were available.

The investment was described as public-private capital intended to encourage additional private-sector participation and accelerate commercialization of advanced technologies. It forms part of a broader Japanese plan to direct ¥370 trillion in public and private funds to 17 strategic fields by fiscal 2040.

• Green Transformation Acceleration Agency
• Kyoto Fusioneering
• JEPLAN

🇭🇰Hong Kong reported AI-related IPO fundraising of nearly HK$100 billion between December and May, representing approximately 55% of total proceeds. #

Hong Kong Financial Secretary Paul Chan said AI-related initial public offerings raised nearly HK$100 billion from December through May. That represented approximately 55% of all IPO proceeds raised in Hong Kong during the same period.

Chan also said multiple AI-related companies had recently been added to Hang Seng Indexes, reflecting their growing weight in Hong Kong’s economy and financial market. Separately, Hong Kong’s exports had recorded high-double-digit growth for several consecutive quarters amid global demand for AI-related products.

The Hong Kong government allocated HK$50 million in its budget for an “AI for All” training program. Through Cyberport, Hong Kong Science and Technology Parks and the Hong Kong Productivity Council, the program targets more than 200 activities and approximately 50,000 participant attendances across 2026 and 2027.

Within the government, an AI efficiency task group had initiated 30 efficiency projects spanning 13 departments. A cited research estimate said that matching large-company AI adoption rates among small and medium-sized enterprises by 2035 could unlock as much as HK$65 billion in economic benefits.

• HK$100 billion

Samsung described fabricating High Bandwidth Memory (HBM) base dies on a logic node to free package area for compute. #

At Hot Chips 2026, Samsung discussed evolving the base die used in High Bandwidth Memory (HBM). The company described fabricating the HBM base die on a logic process node rather than treating it solely as a conventional memory component.

Samsung’s presentation linked that change to opportunities for freeing more package area for compute. The supplied reports did not provide a production date, named customer, process-node dimension or quantified package-area saving.

This was a packaging and memory-architecture presentation rather than the announcement of a shipping accelerator, server or complete computing system. It addressed how the HBM base die could evolve within packages serving machine-learning workloads that require increasing memory capacity and bandwidth.

• Samsung
• High Bandwidth Memory (HBM)

d-Matrix presented its Raptor generative-inference accelerator, stacking three-dimensional DRAM and logic instead of relying on HBM. #

d-Matrix presented its Raptor accelerator at Hot Chips 2026. The device is designed for generative-AI inference, distinguishing its target workload from general-purpose central processing units or training-focused accelerator systems.

Raptor uses a three-dimensional architecture that stacks Dynamic Random-Access Memory (DRAM) and logic. ServeTheHome characterized the design as moving away from reliance on High Bandwidth Memory (HBM).

The supplied source did not provide performance, power, memory-capacity, pricing or shipment-volume figures. It also did not identify a server configuration, cloud deployment or named customer associated with Raptor.

• HBM

Montage Technology secured PCI-SIG certification for its PCIe 6.x and CXL 3.x retimer, joining the organization’s supplier list. #

Montage Technology (澜起科技) announced that its PCI Express (PCIe) 6.x and Compute Express Link (CXL) 3.x retimer had passed a series of certifications from the PCI Special Interest Group (PCI-SIG). The product was consequently added to PCI-SIG’s official supplier list.

The certification applies to a retimer chip, an interconnect component, rather than to a GPU, accelerator, server or complete rack. The disclosed compatibility scope was PCIe 6.x and CXL 3.x.

Montage said inclusion on the supplier list would substantially simplify procurement for large original-equipment manufacturers and hyperscale data-center operators. The source did not disclose customer orders, shipment volumes, pricing or revenue associated with the certification.

• Montage Technology
• PCI-SIG

Innolight said customers provided 2027 orders for 800G and 1.6T optical modules; near-packaged optics shipments could begin in 2027. #

Innolight (中际旭创; 300308.SZ) said major customers had provided guidance and orders for 2027. It reported that demand for 800-gigabit-per-second and 1.6-terabit-per-second optical modules in 2027 continued to grow rapidly compared with 2026.

Historically, optical-module orders were commonly placed for approximately three months, according to the company. Innolight said many customers had now provided 2027 orders, while some customers were discussing longer-term agreements whose details were not disclosed.

For new products, Innolight was developing or sampling 2.4T modules, near-packaged optics (NPO) and XPO products for scale-out and scale-up interconnect scenarios. It said these products could enter mass production in the second half of 2027, with larger-scale shipments possible in 2028; separately, it said some customers had supplied explicit NPO demand guidance and orders, supporting potential shipments in 2027.

The company was expanding capacity and preparing supplies of optical chips, electrical chips and printed circuit boards. It said a small portion of its factory expansion would be delivered in the second half of 2026, most would be delivered in 2027, and much of the associated infrastructure spending would begin contributing to revenue in 2027.

• Innolight

Zhongwu High-tech approved ¥190 million for annual capacity of 140 million precision micro-tools serving high-speed AI printed circuit boards. #

China Tungsten High-Tech (中钨高新; 000657.SZ), referred to in the digest as Zhongwu High-tech, approved an expansion by its controlled subsidiary Shenzhen Jinzhou Precision Technology. The project targets annual capacity of 140 million high-precision micro-tools for high-speed, high-frequency AI printed circuit boards (PCBs).

The estimated investment is ¥190 million. Funding consists of ¥178.5 million from Jinzhou’s own funds and ¥11.5 million in bank borrowing, and the investment was described as neither a related-party transaction nor a major asset restructuring.

The project will use leased factory space in the eastern section of the company’s Shenzhen base. Construction is scheduled for one year, with the project expected to reach designed capacity in its second year.

The company’s project calculation showed an after-tax financial internal rate of return of 14.55%. That figure is a project estimate rather than a realized operating return, and the source did not disclose customer commitments or contracted tool volumes.

• Zhongwu High-tech

Guangdong approved a demonstration power station using China’s domestically developed 300-megawatt-class F-class heavy-duty gas turbine. #

The Guangdong Provincial Development and Reform Commission formally approved a first-of-a-kind demonstration power-station project using a domestically developed 300-megawatt-class F-class heavy-duty gas turbine. The approval was issued as Guangdong Development and Reform Approval Document No. 6 of 2026.

The turbine was independently developed by China United Heavy-Duty Gas Turbine Technology, associated with State Power Investment Corporation. The source described it as China’s domestically developed gas turbine with the highest power rating and technical class.

The approval concerns a demonstration power station and its turbine equipment, not 300 MW of data-center capacity. The supplied report did not state the station’s construction cost, completion date, site-level generation configuration or commercial-operation schedule.

Hungary’s sole nuclear plant targets full output Wednesday after drought-related restrictions shut three reactors; two restarted over the weekend. #

Hungary was gradually restoring generation at the Paks nuclear power plant, the country’s only nuclear station. The plant targeted a return to full output by Wednesday.

Drought conditions had forced three reactors to shut during the preceding month. Two of those reactors restarted over the weekend.

At its lowest point, the station’s output had fallen to approximately one-tenth of total capacity. Following the two restarts, output recovered to more than half of total capacity, according to the report.

The disruption was linked to an unusual heat wave and drought affecting Europe. The source did not provide the plant’s megawatt output, the identity of the reactor still offline or a precise hour for full restoration.

More than 28,000 Indiana customers remained without electricity nearly two weeks after severe storms and flooding began. #

Severe storms and flooding had affected Indiana since August 11. By the morning of August 23, local electricity providers reported that more than 28,000 customers remained without power.

The broader event had caused outages for hundreds of thousands of customers cumulatively. At least seven people had died, according to the report.

Gary was identified as one of the more severely affected cities. The source did not provide restoration schedules, utility-by-utility outage counts or the amount of damaged transmission and distribution infrastructure.

Iran announced a Fars Province gas discovery exceeding 7.5 trillion cubic feet, including approximately 5.7 trillion cubic feet recoverable. #

Iranian Oil Minister Mohsen Paknejad announced a gas discovery in southern Fars Province containing more than 7.5 trillion cubic feet, equivalent to approximately 212.4 billion cubic meters. The announcement was made on August 23.

Paknejad said a recovery rate exceeding 72% would yield approximately 5.7 trillion cubic feet, or about 161.4 billion cubic meters, of recoverable gas. These figures were presented by the Iranian government and were not accompanied in the supplied sources by an independent reserve assessment.

The minister described the discovery as “sweet gas,” a characteristic he said would reduce development and extraction costs. He also said the discovery included a large volume of condensate, but no condensate quantity was provided.

The sources did not disclose a field-development budget, production start date, planned daily output or pipeline connection. The announcement therefore establishes a stated discovery and recoverable-resource estimate, not current production capacity.

China reported 2,185 exa floating-point operations per second (EFLOPS) of intelligent-computing capacity at June-end. #

China’s intelligent-computing capacity reached 2,185 exa floating-point operations per second (EFLOPS) at the end of June 2026. The figure was published in the Green Computing Development Research Report 2026 (《绿色算力发展研究报告（2026年）》).

The report was released by Ao Li, vice president of the China Academy of Information and Communications Technology, at the 2026 Green Computing and Artificial Intelligence Conference. The disclosed figure covers national intelligent-computing capacity rather than the capacity of one data center, operator or accelerator cluster.

The conference also reported progress in higher-efficiency computing equipment, greener computing facilities and coordination between computing infrastructure and electricity supply. Computing networks, token-based services and overseas computing services were identified as emerging business models.

The supplied source did not provide a breakdown by accelerator type, operator, province, utilization rate or precision format. It therefore does not establish how much of the 2,185 EFLOPS was active, commercially available or allocated to particular workloads.

China Mobile recorded ¥52.9 billion in compute-service revenue; AI data-center revenue rose 486.1%, while intelligent-compute revenue reached ¥5.3 billion. #

China Mobile reported first-half compute-service revenue of ¥52.9 billion, up 14% year over year. This category is broader than GPU-cloud revenue and includes multiple computing-service activities.

Revenue from artificial-intelligence data centers (AIDCs) rose 486.1% year over year. Intelligent-computing service revenue reached ¥5.3 billion, representing growth of 130.1%.

China Mobile said intelligent-computing services contributed 46.1% of the incremental revenue generated by its compute-service business. The source did not disclose the absolute AIDC revenue figure, installed accelerator count, data-center power capacity or customer concentration.

The figures were reported in an assessment of first-half results from China’s three major telecommunications operators. The report said computing and intelligent services were increasing as a share of their principal businesses, although new operations had not yet fully offset pressure on traditional services.

• China Mobile

China Telecom intelligent-compute revenue rose 95% with 94% utilization; China Unicom compute revenue reached ¥41.9 billion, up 13%. #

China Telecom reported that first-half intelligent-computing revenue increased 95% year over year. It also reported a 94% utilization rate for intelligent-computing capacity.

China Unicom reported total compute revenue of ¥41.9 billion, an increase of 13%. Within that business, intelligent-computing-center revenue rose 11%, while intelligent-computing revenue increased 9%.

These figures distinguish broad compute revenue from the narrower intelligent-computing categories. The source did not report corresponding absolute intelligent-computing revenue for China Telecom or provide accelerator counts, megawatt capacity or precision-adjusted computing performance for either operator.

The data appeared alongside results for China Mobile in a review of the three state-owned telecommunications operators. The review identified AI data centers, intelligent-computing services and token-based operations as areas of business transition.

• China Telecom
• China Unicom

Baidu reported AI-cloud infrastructure revenue growth of 50%, while its GPU-cloud revenue increased 283%. #

Baidu reported that revenue from AI-cloud infrastructure increased 50% year over year. Revenue from its graphics processing unit (GPU) cloud business rose 283%.

The company also said AI business revenue had reached 50% of its general-business revenue. The source did not provide absolute revenue for AI-cloud infrastructure or GPU cloud, installed GPU capacity, utilization or customer concentration.

Baidu separately stated that it had no plan for an additional share placement. It said existing funds and operating cash flow were sufficient to support its planned business development.

At the end of the second quarter, Baidu reported ¥283.1 billion in cash and investments. Operating cash flow had remained positive for four consecutive quarters, according to the company representative cited in the report.

• Baidu

Pasqal researchers developed an AI agent that generates quantum-computing code from English instructions and executes it on physical quantum computers. #

Researchers at French quantum-computing startup Pasqal developed an AI agent that accepts instructions written in English. The agent generates corresponding quantum-computing code from those natural-language requests.

The system can then run the generated code automatically on physical quantum computers. This distinguishes the work from code generation that is evaluated only in simulation or left for a user to execute manually.

The approach was described as extending “vibe coding” from conventional software development into quantum computing. Researchers said it could lower the technical barrier to quantum-computing experiments, but the supplied sources provided no benchmark, success-rate, hardware-count or comparative performance data.

No commercial-release date, price or supported-hardware list was disclosed. The report attributes the development to Pasqal researchers and cites reporting by the journal Nature.

• Pasqal

Noitom released HiPHI, containing 617.5 hours of motion-capture data, including 245.7 hours of human-object interactions. #

Noitom (诺亦腾) released HiPHI at the 2026 World Robot Conference. HiPHI is a high-precision optical motion-capture dataset intended for humanoid-robot learning, digital-human development and computer-graphics research.

The dataset contains 617.5 hours of data, including left-right mirrored samples. Its composition is 371.8 hours of full-body human-motion data and 245.7 hours of human-object interaction data.

For the human-object interaction portion, each object’s movement trajectory and three-dimensional mesh were recorded in synchronization with the motion-capture performer. This supplies aligned human and object information rather than human skeletal motion alone.

HiPHI was made publicly available on Hugging Face. The source did not disclose the dataset’s storage size, licensing conditions, number of performers, number of objects or capture-frame rate.

• Noitom
• HiPHI

United Imaging and Tianjin University launched uMR Divine Vision (uMR 神观), spanning magnetic-resonance imaging, decoding, neural modulation, evaluation, and algorithm iteration. #

United Imaging Healthcare (联影医疗) and Tianjin University jointly released uMR Divine Vision (uMR 神观) at China’s first national magnetic-resonance brain-computer-interface conference in Tianjin. The event’s theme was “Integration, Decoding and Modulation—Opening a New Era of Brain-Computer Interaction.”

The organizations described uMR Divine Vision as a full-stack magnetic-resonance brain-computer-interface solution. It connects imaging, simulation, neural modulation, evaluation and algorithm iteration in a closed loop.

The technical system includes high-spatial-and-temporal-resolution magnetic-resonance imaging, magnetic-resonance hardware adaptation and optimization, a toolkit for magnetically compatible brain-computer-interface equipment, and magnetic-resonance-guided neural-modulation methods. A related formulation described the workflow as acquisition, decoding, modulation and evaluation.

United Imaging and Tianjin University also initiated a magnetic-resonance brain-computer-interface innovation alliance. The supplied reports describe the system as the world’s first solution of its kind, but provide no clinical-trial results, regulatory approval, pricing or commercial deployment figures.

• United Imaging
• Tianjin University