AI & Compute Infrastructure — August 9, 2026

🇺🇸NVIDIA agreed to invest $2 billion for roughly 20% of Lancium, a Blackstone-backed AI-campus power developer. #

The reported agreement was attributed to three people familiar with the transaction. The initial $2 billion investment would give NVIDIA an approximately 20% interest in Lancium, an infrastructure developer backed by Blackstone.

The transaction reportedly values Lancium, including the investment and company debt, at an enterprise value of approximately $10 billion. The source did not identify a public filing or state that the initial investment had already closed.

Lancium supplies power infrastructure for the Texas artificial-intelligence campus associated with OpenAI and Oracle’s Stargate project. The report said the investment is intended to help secure several gigawatts of planned power allocations for data projects serving NVIDIA chip customers.

• NVIDIA
• Lancium
• Blackstone

🇺🇸An additional $1 billion depends on Lancium securing more planned power, potentially raising NVIDIA’s ownership near 30%. #

NVIDIA’s reported commitment includes a possible additional investment of $1 billion. That second tranche is conditional on Lancium obtaining more planned power resources rather than being an unconditional, completed financing.

People familiar with the transaction said the additional investment also depends on multiple Lancium campuses satisfying specified conditions, including grid-interconnection milestones. If those conditions are met and the additional investment occurs, NVIDIA’s ownership could rise to approximately 30%.

The source described the total potential investment as up to $3 billion: $2 billion initially and another $1 billion conditionally. It did not disclose financing terms such as collateral, liquidation preferences, board rights, or a timetable for closing either tranche.

• Lancium
• NVIDIA

🇺🇸More than 150 localities enacted data-center bans during July; over 500 restrictions were active by early August. #

More than 150 U.S. towns and counties reportedly adopted temporary or permanent data-center bans in July alone. Some of the restrictions were approved through emergency meetings and accelerated votes.

By early August, the number of active data-center bans nationwide had exceeded 500. The same tracking effort counted only slightly more than 300 restrictions when it first went online at the end of June.

Residents cited higher household electricity costs, consumption of scarce water resources, and noise as concerns. Project supporters argued that data centers can create employment, attract investment, and improve infrastructure in rural and industrial areas, but the source said those arguments had not stopped the continued adoption of restrictions.

The report identified Anthropic, Google, and OpenAI as companies relying on data-center expansion to increase computing capacity. It did not quantify the capacity affected by the bans or identify how many proposed projects had been cancelled.

🇨🇳Moore Threads (摩尔线程) approved a proposed Hong Kong H-share listing, still requiring shareholder and regulatory approvals; terms remain unset. #

Moore Threads (摩尔线程) said its board had approved proposals to issue overseas-listed foreign shares, or H-shares, and seek a listing on the Hong Kong Stock Exchange’s Main Board. This was a board authorization, not a completed offering or an approved Hong Kong listing.

The proposal must still be considered by shareholders. It also requires filings, approvals and/or authorizations from the relevant governmental and regulatory authorities.

Moore Threads said it would choose an appropriate timing and issuance window within the validity period of the shareholder resolution, or another period extended by shareholders. The number of shares, offering price, proceeds, underwriters and expected timetable had not been determined.

The company expressly stated that implementation remained subject to substantial uncertainty. The disclosure therefore establishes a proposed capital-markets process rather than a committed financing.

• Moore Threads (摩尔线程)
• H-share

🇨🇳Unitree Robotics (宇树科技) scheduled its STAR Market offering at RMB 150.80 per share, targeting approximately RMB 6.099 billion in proceeds. #

Unitree Robotics (宇树科技) scheduled online and offline subscriptions for its Shanghai Stock Exchange STAR Market offering on August 10. The issue price was set at RMB 150.80 per share, with payment due on August 12.

The offering was expected to raise approximately RMB 6.099 billion in gross proceeds. The source placed Unitree’s implied market capitalization at listing at approximately RMB 60.993 billion.

The company’s securities code and offline subscription code are 688836, while the online subscription code is 787836. The online tranche was expected to include 6.471 million shares, and the reported market estimate for the online allocation success rate was 0.02%–0.03%.

The issue price corresponds to a reported offering price-to-earnings ratio of 219.23 times. Unitree’s disclosed business covers high-performance general-purpose humanoid robots, quadruped robots, robot components, and embodied-intelligence models.

• Unitree Robotics (宇树科技)
• STAR Market

Apple is testing CXMT memory chips for China-market iPhones and MacBooks; HP and Acer already use them elsewhere. #

Apple is reportedly testing memory chips from ChangXin Memory Technologies, or CXMT, across product lines including iPhones and MacBooks. The testing was described as part of efforts to address memory shortages associated with the artificial-intelligence boom.

People familiar with the matter said Apple had entered preliminary discussions with CXMT about component supply. The intended scope was limited to some Apple devices sold in China, and the report did not say that Apple had completed qualification or signed a supply contract.

HP and Acer were reported to have already started using CXMT chips in devices sold outside the United States. The source did not identify the exact memory products, capacities, process technologies, shipment volumes, or device models involved.

• Apple
• CXMT
• MacBooks
• HP
• Acer

Samsung plans to begin production at its advanced Taylor, Texas foundry by year-end and started recruiting its first interns. #

Samsung Electronics plans to begin production at its advanced semiconductor foundry in Taylor, Texas, by the end of 2026. The source did not specify the process node, customer programs, initial wafer capacity, or production-ramp schedule.

Samsung has also begun recruiting the Taylor facility’s first group of interns for the following summer. This marks the first time the Taylor factory has been included in Samsung’s annual internship program across its major global sites.

The internship is scheduled to last 11 weeks. Participants will work alongside Samsung engineers on practical projects involving products, manufacturing processes, and technology.

• Samsung

Moore Threads’ MTTS5000 intelligent-computing clusters achieved scaled sales and deployments across Beijing, Wuxi, Hangzhou, and other cities. #

Moore Threads reported that its MTTS5000 intelligent-computing clusters reached scaled sales during the first half of 2026. The clusters had been deployed in Beijing, Wuxi, Hangzhou, and other locations.

The company said the system was among the first products to pass China’s national “secure and reliable assessment” (安全可靠测评). It also reported multiple high-end training achievements involving domestic chips training domestic models, described as “guoxin xun guomo” (国芯训国模).

Moore Threads said it was expanding among internet and telecommunications-operator customers. Its MUSA software-stack ecosystem had more than 800,000 developers.

For the first half, revenue reached RMB 1.736 billion, up 147.42% year over year and above the company’s full-year 2025 revenue. Gross profit was RMB 989 million, while research and development spending was RMB 769 million; the company recorded a net loss attributable to shareholders of RMB 11.5631 million, versus RMB 271 million a year earlier.

• Moore Threads’ MTTS5000

Xinshang Micro-Assembly (Xinshang Weizhuang; 芯上微装) received repeat batch orders after its 350-nanometer AST6200 stepper passed compound-semiconductor customer validation. #

Xinshang Micro-Assembly (Xinshang Weizhuang; 芯上微装) said its internally developed AST6200 step-and-repeat projection lithography system passed process validation at a leading domestic compound-semiconductor customer. The source characterizes the system as a 350-nanometer stepper.

After validation, the customer placed repeat batch orders. The report did not disclose the customer’s identity, order value, system count, delivery schedule, or revenue-recognition timing.

The AST6200 can be used in manufacturing processes for power devices, radio-frequency front ends, optoelectronic chips, and MicroLED displays. No throughput, overlay-accuracy, yield, or utilization figures were provided.

WisdPi’s WP-UT9 external USB Type-C adapter delivers 10GBASE-T Ethernet using Realtek’s RTL8159 platform. #

ServeTheHome reviewed the WisdPi WP-UT9, an external networking adapter that connects through USB Type-C. The device provides a 10GBASE-T Ethernet interface.

The adapter is based on Realtek’s RTL8159 10-gigabit Ethernet platform. This identifies the RTL8159 as the networking platform inside the external adapter rather than a server network-interface card or an integrated motherboard controller.

The source description characterized the WP-UT9 as a stylish external adapter. It did not provide additional benchmark results, power-consumption figures, pricing, or supported operating systems in the supplied source data.

• WisdPi’s WP-UT9
• USB Type-C
• Realtek’s RTL8159

China’s electricity load reached a record 1.557 billion kilowatts; 25 provinces collectively set 78 load records this year. #

China’s national electricity load reached 1.557 billion kilowatts on August 7, setting a record for the fourth time during the summer. The new peak was approximately 4.8 million kilowatts above the previous record.

Since July 29, the national load had remained at elevated levels. Excluding weekends and holidays, each day’s maximum load exceeded the previous-year record of 1.508 billion kilowatts, producing what the source described as a plateau-like pattern in peak demand.

Six regional power grids, including North China, Northeast China, and East China, had collectively set 21 historical records during the year. Across 25 provinces, autonomous regions, and municipalities—including Beijing, Liaoning, Jiangsu, Jiangxi, Sichuan, and Gansu—provincial loads had set records a combined 78 times.

Hebei’s Yanzhao Guyuan Energy Storage Demonstration Project (燕赵沽源储能示范项目) connected 300 megawatts/600 megawatt-hours, supported by a 220-kilovolt substation. #

The Yanzhao Guyuan Energy Storage Demonstration Project (燕赵沽源储能示范项目) in Hebei reached full-capacity grid connection. The project is located in Xixinying Township, Guyuan County, Zhangjiakou.

The installation uses a 300-megawatt/600-megawatt-hour lithium iron phosphate battery energy-storage system. The power rating is 300 megawatts, while 600 megawatt-hours is the system’s stated energy-storage capacity.

The facility comprises 48 storage units, each rated at 6.25 megawatts/12.5 megawatt-hours. It also includes a 220-kilovolt step-up substation and associated electrical facilities.

POWERCHINA constructed the project. The source did not provide its total investment, owner, contracted revenue, discharge duration beyond the stated ratings, or commercial-operation date.

South Korea’s peak electricity demand reached 95.321 gigawatts, with 8.2 gigawatts of reserve; alerts begin below 5.5 gigawatts. #

South Korea’s maximum electricity demand reached 95.321 gigawatts on August 7, the highest level recorded during 2026 up to that date. The increase was attributed to a continuing heat wave.

Available reserve capacity stood at 8.2 gigawatts. An emergency electricity-supply alert is triggered if reserve capacity falls below 5.5 gigawatts.

The South Korean government expected electricity demand to continue rising during the following week. The cited factors were persistent hot weather and increased industrial production after the end of the summer holiday period.

POWERCHINA’s 96.9-megawatt Peru solar project entered energized commissioning; projected annual generation totals 250 million kilowatt-hours. #

POWERCHINA’s 96.9-megawatt Wayra solar project in Peru successfully energized its first 33-kilovolt collector circuit. The equipment operated stably, moving the project into energized commissioning rather than full commercial operation.

The project is designed with seven 33-kilovolt medium-voltage collector circuits. The remaining circuits were scheduled to undergo energization and integrated commissioning in batches.

After completion, projected average annual generation is 250 million kilowatt-hours, equivalent to approximately 685,000 kilowatt-hours per day. POWERCHINA said this could meet the annual electricity needs of approximately 230,000 Peruvian households and reduce annual carbon-dioxide emissions by about 150,000 tonnes.

The solar facility will operate alongside the existing Wayra wind farm. Together, they will form a wind-and-solar complex with total installed capacity exceeding 400 megawatts.

• POWERCHINA’s 96.9-megawatt

National Supercomputing Internet (国家超算互联网) activated China’s first fully domestic 100,000-card AI supercluster at its Zhengzhou core node. #

China’s first fully domestic 100,000-card artificial-intelligence supercluster entered operation at the Zhengzhou core node of the National Supercomputing Internet (国家超算互联网). The source described this as China’s move into a new stage of 100,000-card computing-infrastructure deployment.

The report used “cards” as the unit but did not identify the accelerator model, manufacturer, memory capacity, precision format, or interconnect technology. It therefore does not establish that all 100,000 devices are conventional graphics processing units rather than another type of AI accelerator.

The Zhengzhou node combines scientific-computing and intelligent-computing capacity. Staff said this integrated design was intended to support new categories of computing demand.

The activation was reported alongside a broader new round of capacity expansion at multiple Chinese computing nodes. No aggregate performance, power consumption, capital expenditure, utilization, or customer figures were disclosed.

• National Supercomputing Internet (国家超算互联网)

AWS is batch-closing internal development EC2 instances; reported CPU-server approval times expanded from hours to several days. #

Amazon Web Services is reportedly reviewing internal engineering teams’ use of compute resources and shutting down batches of Elastic Compute Cloud, or EC2, virtual servers previously used for software development. AWS had warned engineering teams in May to control compute consumption so the company could continue meeting customer demand.

The resource-control effort covers more than AI-specific chips. It also applies to traditional central processing unit servers used to support network and software operations.

One AWS employee said CPU-server requests that previously received approval within a few hours could now take several days. The employee described that level of internal resource queuing as unusual and said it could slow project execution.

AWS responded that market demand for computing capacity was substantial but said it could still satisfy the large majority of internal and external customer requests. The source did not quantify the number of instances being closed, the capacity recovered, or the affected regions.

• AWS
• EC2

China CAMC Engineering’s (中工国际) China IPPR data-center business remains developmental, contributes little revenue, and has limited earnings impact. #

China CAMC Engineering (中工国际; Shenzhen Stock Exchange ticker 002051) addressed the status of the data-center business operated by its subsidiary China IPPR. The response was posted on an investor-interaction platform.

The company said the business remains in a cultivation and development stage. It accounts for only a small proportion of China CAMC Engineering’s total operating revenue.

Management also said the business has a limited effect on the company’s overall operating performance. No revenue amount, order backlog, capacity under development, customer contracts, capital expenditure, or deployment schedule was disclosed.

The statement describes the business’s current financial materiality; it does not establish a major operating data-center platform or a completed pivot into AI compute infrastructure.

MiniMax open-sourced H3; it topped Hugging Face’s popularity ranking within three days and secured 100-plus Day-0 enterprise integrations. #

MiniMax formally released H3 as an open-source model. Within three days, it reached the top of the popularity ranking on Hugging Face, and more than 100 companies completed “Day-0” integrations.

The reported integrations covered the stack from chip manufacturers to inference frameworks. Stable Diffusion founder Emad Mostaque also publicly praised MiniMax following the release.

The report said MiniMax served more than one million enterprise and developer customers globally, five times the level from six months earlier. It also cited a U.S. investment institution’s estimate that as many as 80% of U.S. AI startups use Chinese open-source models in fundraising presentations.

The same article stated that Chinese open-source models represented 41% of global model downloads and that the six most-called mainstream large models were all developed by Chinese teams. These broader figures were claims reported by Economic Daily rather than metrics independently documented in the supplied source data.

• MiniMax
• H3
• Hugging Face

Alibaba’s Qwen3.8-Max (千问3.8-Max) contains 2.4 trillion total parameters and approximately 95 billion activated parameters. #

Alibaba released its new flagship Qwen3.8-Max (千问3.8-Max) model on August 3. The model has 2.4 trillion total parameters and approximately 95 billion activated parameters.

The distinction indicates that the stated total parameter count and activated parameter count are different measures; the source does not say that all 2.4 trillion parameters are active for each operation. It did not provide training-compute, accelerator-count, dataset-size, context-window, or inference-cost figures.

The report said Qwen3.8-Max improved capabilities in reasoning, coding, and agent tasks. Alibaba also began public testing of its enterprise agent product Qwen Office (千问办公) on the same day.

• Alibaba’s Qwen3.8-Max (千问3.8-Max)

Former ByteDance robotics head Kong Tao leads Xiaomi’s robotics foundation-model team; the broader division has approximately 200 employees. #

Multiple independent sources reportedly confirmed that Kong Tao, formerly the head of ByteDance’s robotics team, had joined Xiaomi. He now leads Xiaomi’s robotics foundation-model team and reportedly joined during the summer of 2025.

Kong brought several former ByteDance employees with him, according to people familiar with the matter. After joining, he held individual discussions with many Xiaomi employees to assess the robotics team’s capabilities relative to the industry.

Xiaomi’s broader robotics division has approximately 200 employees. Its research covers robot bodies, lower-level motion control described as the “cerebellum,” higher-level intelligence described as the “brain,” and robot operating systems.

Different groups inside the division also conduct preliminary research on overlapping areas, including world models. A person close to the organization said Kong’s foundation-model team works from a separate office, while the broader robotics division is maintained under a high level of internal confidentiality.

• ByteDance
• Xiaomi