AI & Compute Infrastructure — August 30, 2026

🇨🇳Enflame Technology plans a RMB 6 billion IPO, offering 43.04 million shares for fifth- and sixth-generation AI-chip development. #

Enflame Technology (燧原科技) is scheduled for subscription on September 2 on the Shanghai Stock Exchange’s STAR Market. Its disclosed stock code is 688801, with online subscription code 787801.

The company proposes to issue 43.035173 million shares, equivalent to 10% of its enlarged share capital. Post-offering share capital would be approximately 430 million shares, while the IPO’s proposed fundraising amount is RMB 6 billion.

The proceeds are designated principally for research, development and industrialization of fifth- and sixth-generation AI-chip product families, as well as an advanced AI hardware-software co-innovation project. These are proposed uses of IPO proceeds rather than completed investments or deployments.

Enflame reported revenue of RMB 301 million in 2023, RMB 722 million in 2024 and RMB 990 million in 2025, representing a three-year compound annual growth rate of 81.32%. The source described Enflame, Moore Threads, MuXi and Biren Technology collectively as four leading domestic graphics processing unit companies.

• Enflame Technology

🇨🇳Pujiang China (01417.HK) approved an AI-infrastructure expansion while retaining property management, leasing, catering, and sanitation operations. #

Pujiang China (浦江中国; 01417.HK) disclosed in a voluntary announcement that its board had decided to expand into an AI-infrastructure business. The expansion is being pursued alongside, rather than as a replacement for, the company’s existing operations.

Those continuing businesses include property-management services, commercial-building leasing, catering and urban sanitation. The announcement therefore describes a corporate expansion into a new business area, not a disposal or termination of the legacy operations.

The company said it would gradually construct an “AI Token Economy” (AI 词元经济) business system centered on AI-token production and distribution. “Token” in this context refers to the units processed by AI models, not necessarily a cryptocurrency or blockchain token.

Pujiang China has established an intermediate holding company, TokenVerse Holdings Limited. Its proposed mainland Chinese operations will be conducted through Shanghai Ciyuan Wanxiang Information Technology Co., Ltd. (上海词元万相信息科技有限公司); the source did not disclose financing size, installed compute capacity, hardware orders or a deployment timetable.

• Pujiang China (01417.HK)

🇨🇳Guangdong launched Token Loan (词元贷) with three banks; Bank of China completed the first credit approval. #

Guangdong’s first specialized financial product for the token economy, Token Loan (词元贷), was formally launched in Guangzhou’s Haizhu District. Bank of China, China CITIC Bank and Bank of Guangzhou are participating in the initiative.

Bank of China completed the first credit approval under the product. The source characterizes this as a credit approval, without disclosing the approved borrower, loan amount, interest rate, maturity, collateral structure or whether funds had already been drawn.

The product is aimed at AI-industry-chain companies described as having light assets, limited financial statements and fragmented cash flows. Their core assets may consist of algorithms, models, data and code rather than factories, real estate or heavy machinery commonly used as traditional loan collateral.

Token Loan is intended to address the timing mismatch between upfront compute purchases and the receipt of operating revenue. The source states that the facility is designed to ease cash-flow shortages during rapid expansion, but it does not provide aggregate facility size or bank-by-bank commitments.

• Token Loan (词元贷)

🇺🇸 🇮🇷U.S. forces struck two Iranian launchers on Larak Island; Asian WTI crude opened nearly 2% higher. #

A U.S. official said American forces struck two Iranian launchers on Larak Island near the Strait of Hormuz. The official said Islamic Revolutionary Guard Corps personnel had been observed preparing to use the launchers to send rockets carrying sea mines into the strait.

Separate reporting described the action as the first U.S. strike on an Iranian target in more than a month. Iran’s Islamic Revolutionary Guard Corps subsequently said the attack caused military and civilian casualties and stated that it would respond.

West Texas Intermediate crude opened nearly 2% higher at $84.29 per barrel, while Brent opened 1.7% higher at $89.43 per barrel. Another early-Asian-session report described international crude prices as rising by more than 2% following the Larak Island strike.

The Strait of Hormuz was described as a route through which one-fifth of global oil and liquefied-natural-gas volumes had passed. The U.S. military also said it had completed mine-clearing work along the international shipping route through the strait.

• WTI crude

🇨🇳STAR Market companies reported first-half revenue of RMB 1.01 trillion and net profit of RMB 144.887 billion. #

Shanghai Stock Exchange data showed that STAR Market companies generated combined first-half revenue of RMB 1.01 trillion. Revenue increased 38.6% year over year.

Combined net profit reached RMB 144.887 billion, an increase of 437.6% from the corresponding period. The exchange said first-half net profit exceeded the STAR Market’s profit for the entirety of the previous year.

Within the market, 36 companies categorized in the growth tier reported a 29.1% year-over-year increase in revenue. Their aggregate net losses narrowed by 62.3%.

These figures cover STAR Market companies collectively and should not be read as results for the broader Shanghai Stock Exchange or the entire Chinese listed-company market. The source did not provide a sector-level breakdown of how much of the revenue or profit came specifically from semiconductors, AI or compute infrastructure.

• STAR Market

NVIDIA Jetson Orin Nano 2 delivers twice its predecessor’s performance using a new Ampere-based Orin system-on-chip. #

NVIDIA announced the Jetson Orin Nano 2 as a new entry-level edge-AI board. The product is positioned within the Jetson line for AI processing at the edge rather than as a data-center GPU or complete server.

ServeTheHome reported that the upgraded board is twice as fast as its predecessor. This is a performance comparison with the previous Jetson Orin Nano generation, not a statement about data-center accelerators or cluster-scale throughput.

The board uses a new Orin system-on-chip, or SoC, based on NVIDIA’s Ampere architecture. A system-on-chip combines multiple computing functions on one piece of silicon; the source specifically characterizes this as new Orin silicon rather than only a board-level refresh.

The supplied article description does not state the board’s price, memory capacity, power envelope, performance in tera operations per second, availability date or individual benchmark results. Those figures therefore cannot be established from the provided source.

• Ampere-based Orin system-on-chip

University of Oxford researchers reported that High-Bandwidth Flash (HBF) provides 16-fold more capacity per stack than High-Bandwidth Memory at comparable bandwidth. #

University of Oxford researchers published a technical paper titled “Hardware-Managed Heterogeneous High-Bandwidth Memory and Flash in LLM Inference Systems.” The work examines a hybrid memory architecture for large-language-model inference.

The paper describes High-Bandwidth Flash, or HBF, as a denser alternative to High-Bandwidth Memory, or HBM. It reports 16 times more capacity per stack at comparable bandwidth.

HBM is high-speed stacked memory used close to processors and AI accelerators, while HBF in this research refers to stacked flash designed to provide high bandwidth with substantially greater capacity. The comparison concerns capacity per stack and bandwidth, not processor compute performance.

The paper investigates hardware management of heterogeneous HBM and HBF resources for inference systems. The supplied abstract excerpt says replacing HBM with HBF can address capacity constraints, but the source data does not provide full latency, energy, cost or end-to-end inference benchmark results.

• High-Bandwidth Flash (HBF)
• High-Bandwidth Memory

TCL Technology targets a 12-layer glass-substrate sample this half, versus 20 layers for mainstream high-performance packaging. #

TCL Technology (TCL科技; 000100) said during an institutional-investor survey that mainstream glass-based advanced packaging for high-compute and high-performance chips requires approximately 20 layers. The company’s planned sample for the second half of 2026 is expected to have approximately 12 layers.

TCL said it is relying on external resources to establish the technical route and prepare a complete sample. It attributed the lower layer count to technical limitations associated with those external resources and acknowledged that the planned sample remains short of the 20-layer mainstream level it cited.

The company stated that it would initiate construction of a pilot line during 2026 if the sample performs as expected and the technical route is substantially established. That makes the pilot-line construction conditional on sample and process validation rather than an already completed or unconditional deployment.

The disclosure concerns a glass-substrate packaging sample, not a finished AI accelerator, server or production-scale packaging line. No pilot-line investment amount, production capacity, customer commitment or mass-production date was disclosed.

• TCL Technology

MuXi reported first-half revenue of RMB 1.324 billion, up 44.67%, as GPU shipments increased significantly. #

MuXi (沐曦股份) reported first-half revenue of RMB 1,323.5937 million, equivalent to approximately RMB 1.324 billion. Revenue increased 44.67% year over year.

The company attributed the increase principally to significant growth in shipments of its graphics processing unit products. It said its products and services had received broad recognition and continued purchasing from downstream customers.

First-half net profit attributable to shareholders was RMB 612.4495 million. MuXi described this as a return to profitability from a loss in the corresponding period.

The disclosure reports shipment growth but does not provide the number of GPUs shipped, model-level sales, average selling prices, customer concentration, installed cluster capacity or a split between hardware and service revenue. It therefore establishes financial growth and the stated shipment driver, but not unit volumes.

• MuXi

Purdue University’s distributed-GPU simulator covers Ampere, Hopper, and Blackwell, achieving 99% Pearson correlation against physical H100 silicon. #

Purdue University researchers published a technical paper titled “Architecting the Next Generation of Asynchronous, Distributed GPUs for the AI Era.” The paper presents a cycle-level simulation framework for distributed GPU systems and AI workloads.

The simulator covers NVIDIA’s Ampere, Hopper and Blackwell GPU generations. These are GPU microarchitecture families, rather than separate server, rack or cluster products.

The researchers reported validation against physical silicon, including an NVIDIA H100 GPU based on the Hopper architecture. The framework achieved a 99% Pearson correlation coefficient in that validation.

Pearson correlation measures how closely two sets of results vary together; it is not itself an absolute-error percentage or a direct performance-speedup figure. The provided source excerpt does not disclose the simulator’s runtime overhead, complete benchmark suite or absolute prediction error.

• Purdue University
• Ampere
• Hopper
• Blackwell
• H100

SpaceX is developing a Bastrop foundry for turbine blades and vanes, potentially shortening gas-turbine deployment by 18 months. #

SpaceX is preparing a foundry in Bastrop, Texas, dedicated to manufacturing blades and vanes for large gas turbines. Blades and vanes are distinct turbine components, not complete turbines or generating plants.

The original report came from The Information, after which Elon Musk confirmed the plan on X, according to the supplied Telegram report. Musk said in-house casting of these components could bring gas turbines online as much as 18 months earlier.

The stated 18-month figure is Musk’s estimate of the maximum schedule reduction, not a reported result from an already completed turbine deployment. The source does not identify a turbine supplier, generation capacity, project count, foundry investment amount or production start date.

The project specifically addresses the casting of turbine blades and vanes, which the report characterized as technically difficult and capacity-constrained components in the global turbine supply chain. It should not be interpreted as confirmation that SpaceX is manufacturing complete gas-turbine systems.

• SpaceX

SpaceX and Tesla are each building annual solar-manufacturing capacity targeting 100 gigawatts. #

Elon Musk said SpaceX and Tesla are each building solar-manufacturing capacity at an annualized target of 100 gigawatts. The statement describes separate capacity targets for the two companies rather than a combined 100-gigawatt figure.

Musk said both companies were proceeding as rapidly as possible. The source does not specify whether the planned capacity covers solar cells, modules, integrated systems or another manufacturing stage.

The report does not provide factory locations, capital expenditure, equipment orders, current installed capacity, construction milestones or target completion dates. Consequently, the 100-gigawatt figures remain stated annual manufacturing targets rather than verified operating output.

In the same statement, Musk said natural gas would remain necessary for several years to supplement and guide solar generation. No corresponding gas-generation capacity target was provided.

• SpaceX
• Tesla

Guizhou’s internet-data-service electricity consumption approached two billion kilowatt-hours during January–July, increasing 101.15% year over year. #

Guizhou Province’s internet-data-service sector consumed close to two billion kilowatt-hours of electricity from January through July 2026. Consumption increased 101.15% from the corresponding seven-month period in 2025.

The metric is electricity consumed by internet data services, not a direct measure of computing performance, GPU count, data-center floor area or contracted power capacity. It captures energy use over the period rather than instantaneous electrical demand in watts.

The update was presented in the context of compute-power and electricity coordination, known as “compute-power coordination” (算电协同). The supplied source does not break consumption down by facility, operator, workload type or energy source.

No peak-load figure, renewable-energy share, power-usage effectiveness measurement or province-wide data-center capacity was disclosed. The reported change therefore establishes the scale and growth of electricity consumption but not the efficiency of the underlying facilities.

European Union gas storage stood at 63%, versus approximately 80% during comparable periods in recent years. #

European Union gas storage was reported at 63% during the final week of August. That compares with an average of approximately 80% during comparable periods in recent years.

The report attributed accelerated inventory depletion to cold weather at the end of the previous winter and increased gas-fired electricity generation during Europe’s summer heat waves. It also said disrupted Gulf energy exports following the U.S.-Israeli war with Iran had made the EU’s pre-winter 80% storage objective difficult to achieve.

Experts cited by the report said storage could enter the winter heating season at its lowest level since 2013 if current conditions persisted. This was an expert projection, not a confirmed future storage level.

The percentages refer to the fill level of gas-storage facilities, not annual European gas demand or gas-fired generation capacity. The source did not provide stored volumes in cubic meters, country-level inventories or individual terminal data.

Beijing University of Posts and Telecommunications’ space compute cloud completed hundreds of calls for more than 100 users and achieved large-model inference efficiency of 10 tokens per joule using domestic hardware and software. #

Beijing University of Posts and Telecommunications (北京邮电大学) said the space compute cloud it leads had entered regular operation, providing in-orbit computing services externally. It was described as the first space compute cloud to offer external in-orbit experimental services.

Since entering service, the platform has completed hundreds of space-compute calls and served more than 100 users. It has supported in-orbit experiments involving big-data processing, sixth-generation communications, distributed storage and mixed-criticality task deployment.

Using a domestic-native hardware and software ecosystem, the cloud achieved large-model inference energy efficiency of 10 tokens per joule. This measures generated or processed model tokens relative to energy consumption and is not a GPU throughput figure in floating-point operations per second.

The service reported space-compute coverage exceeding 10%. It also said it had established continuing service capabilities for real-time data processing, low-latency intelligent decision-making and iterative development of satellite-network capabilities, although the source did not define the denominator used for the coverage figure.

Barclays estimates AI model companies spend $35–$40 per $100 of revenue on inference across AWS, Azure, and Google Cloud. #

A Barclays research report estimated that AI model companies direct approximately $35–$40 of every $100 of revenue to inference-compute costs paid to Amazon Web Services, Microsoft Azure and Google Cloud Platform. The calculation concerns paid inference rather than model-training expenditure.

Barclays estimated that the cloud providers receive approximately $10–$20 of operating profit from that spending. The report associated this with cloud-provider operating margins of 35%–45%.

The report also said adjusted gross margins for AI laboratories’ paid-inference businesses had risen from low-double-digit levels in 2025 to 50%–65% or higher in 2026. It estimated a year-over-year improvement of 30–50 percentage points.

Barclays analysts said actual margins could be higher than the report’s estimates. They also expected margins to decline gradually as competition among frontier models intensified and compute supply expanded; both the cost and margin figures remain analyst estimates rather than company-reported segment results.

• Barclays
• AWS
• Azure
• Google Cloud

RuiPath 2.0 uses seven billion parameters; its two-million-parameter Edge version runs directly on ordinary hospital PCs. #

Ruijin Hospital (瑞金医院) and Huawei Cloud jointly released the RuiPath 2.0 pathology model at the 2026 Medical Artificial Intelligence Innovation Forum in Shanghai. The primary RuiPath 2.0 model has seven billion parameters.

The developers said the model provides precise, interpretable labeling across dimensions including histological subtype, depth of invasion and mitotic figures. These are pathology-analysis functions rather than general-purpose conversational-model capabilities.

They also introduced RuiPath 2.0 Edge, a lightweight model containing two million parameters. The Edge version can be deployed directly on an ordinary hospital personal computer, according to the release.

The source does not provide diagnostic-accuracy percentages, inference latency, hardware requirements for the seven-billion-parameter model, regulatory-approval status or a count of hospital deployments. The announcement therefore establishes the model sizes, stated pathology functions and local deployment capability, but not clinical outcomes.

• RuiPath 2.0

Serbia’s Chinese-Serbian robot factory began production; its completed project targets up to 20,000 humanoids and robot dogs annually. #

Serbia’s first robot factory formally began production on August 29 in the western city of Šabac. The facility was jointly built by Chinese and Serbian parties and was described as Europe’s first mass-production base for humanoid robots.

The first phase involved an investment of €20 million and occupies approximately 3,000 square meters. Its activities include assembly of humanoid robots and robot dogs, calibration of movement and vision systems, and performance testing.

Humanoid robots produced at the factory are initially intended for industrial production, with later expansion planned into education and public services. The planned products are intended for European and global markets.

A subsequent robot industrial park is planned in Serbia with total investment of approximately €200 million. Once the overall project is completed, planned annual capacity is up to 20,000 humanoid robots and robot dogs; this is a project target, not the factory’s reported current production volume.

Hanshow Technology’s smart shopping carts entered batch delivery; its inspection robots remain in overseas supermarket proof-of-concept trials. #

Hanshow Technology (汉朔科技; 301275) said during an institutional-investor survey that its smart shopping carts had entered the batch-delivery stage. The disclosure distinguishes these commercial deliveries from other solutions that remain in testing.

The company is developing AI capabilities for smart shelf management, real-time in-store execution, AI shopping guidance, self-checkout, intelligent inspection and cleaning. It described the portfolio as an end-to-end system extending from sensing to execution.

Hanshow said these solutions are moving gradually from pilots toward scaled implementation. However, its inspection robots and related products remain in proof-of-concept, or POC, trials at overseas supermarkets.

A proof of concept demonstrates and evaluates a solution in a limited setting and is not equivalent to a binding volume order or batch deployment. The source did not disclose delivery quantities for the carts, trial customer names, revenue contribution or a timetable for converting inspection-robot trials into commercial orders.