🇬🇧NScale, a British AI-cloud operator, reportedly seeks up to $3 billion through a U.S. initial public offering. #
NScale was reported to be seeking as much as $3 billion through an initial public offering in the United States. The report described NScale as an AI-cloud company but did not provide a proposed exchange, ticker, valuation, share count, or pricing range.
The wording “seeks up to” describes a potential fundraising target rather than completed financing. The source did not state that NScale had publicly filed a registration statement or established a timetable for the offering.
No underwriting banks, existing shareholders, use-of-proceeds details, or offering conditions were identified in the supplied report. Accordingly, the source establishes only the reported fundraising objective and prospective U.S. IPO route.
🇺🇸August flash composite PMI reached 56, with services at 56.8 and manufacturing at 53.2. #
The preliminary U.S. S&P Global composite Purchasing Managers’ Index reached 56 in August. That exceeded both the reported forecast of 54 and the previous reading of 54.5.
Services supplied most of the upside, with the flash services PMI reaching 56.8. The figure was above the forecast of 54 and the prior reading of 54.6.
The flash manufacturing PMI was 53.2, remaining above the 50 level while falling short of the 53.9 forecast and the previous 53.9 reading. The composite result therefore strengthened even though manufacturing weakened relative to both consensus and the prior month.
🇨🇳China’s Ministry of Finance plans increased five-year funding for basic research, applied research, and national strategic technology programs. #
China’s Ministry of Finance said central-government funding during the 15th Five-Year Plan period would focus more heavily on basic research, applied basic research, and national strategic science and technology tasks. It also intends to guide local governments and private-sector participants to increase research and development spending.
During the 14th Five-Year Plan period, national general-public-budget expenditure on science and technology was 33% higher than during the preceding five-year period. Economy-wide research and development expenditure maintained average annual growth of 10%, while 2025 spending exceeded RMB3.9 trillion and reached 2.8% of gross domestic product.
The ministry also plans to improve the allocation, management, and use of fiscal science and technology funding. Its stated measures include reforming rights assignments for job-related scientific achievements, managing relevant assets separately, using government procurement to support innovation, and applying government investment funds and small-business programs to support specialized enterprises and digital transformation.
🇰🇷South Korea plans a Future Fund using additional tax revenue associated with the AI-driven semiconductor industry expansion. #
South Korea’s economic authorities announced plans to establish a Future Fund (未来基金) using additional fiscal revenue linked to the artificial-intelligence and semiconductor boom. The government characterized the additional receipts as unexpected excess fiscal revenue.
Officials said stronger global semiconductor conditions were increasing corporate-tax receipts and that domestic tax revenue was expected to rise substantially. The announcement did not specify the fund’s size, launch date, investment mandate, governance structure, or initial capitalization.
The government described the fund as a strategic investment platform intended to support long-term economic growth and increase South Korea’s potential growth rate. The source presented the proposal as a government plan; it did not report that funding had already been appropriated or deployed.
Primemas plans second-half mass production of Micron-backed Compute Express Link memory pools supporting more than 100 terabytes per server. #
Primemas reportedly plans to begin mass production of a “just a bunch of memory,” or JBOM, system with Micron Technology during the second half of 2026. The source describes collaboration with Micron; it does not identify an equity investment or other financial backing by Micron.
The system combines multiple Compute Express Link, or CXL, memory cards into a larger shared memory pool. Primemas said its CXL memory solution could support more than 100 terabytes of memory in one server.
Primemas also set a management target of generating $200 million in revenue from the CXL memory solution by the end of 2027. That figure is a forward revenue objective, not reported revenue, an order backlog, or a binding customer commitment.
FADU’s FC-6161 PCIe Gen6 controller targets 28.5 GB/s sequential reads and power consumption below 7 watts, with second-half shipments planned. #
FADU said its PCI Express Generation 6 enterprise solid-state-drive controller completed chip-out during the first half of 2026, with shipments planned to begin in the second half. The FC-6161 is specified for sequential read throughput of up to 28.5 gigabytes per second.
At peak performance, controller power consumption is stated to remain below 7 watts. The power figure applies to the controller rather than an entire solid-state drive, server, or storage rack.
FADU described the product as part of its “FADU 2.0” strategy, extending its enterprise SSD-controller portfolio from conventional central-processing-unit data centers into AI-computing environments. The company also said its roadmap extends through PCIe Gen7, with samples planned for the year after next, while arguing that inference workloads and key-value cache activity are increasing the frequency and speed of data-center storage access.
Zhaolong Interconnect (兆龙互连) mass-produces 112 Gb/s-per-channel cables, delivers 224 Gb/s samples, and validates 1.6 Tb/s assemblies. #
Zhaolong Interconnect (兆龙互连) said its 112-gigabit-per-second-per-channel high-speed cables have reached stable volume supply. Its 224-gigabit-per-second-per-channel cables have reached small-batch delivery rather than full mass production.
The company’s 1.6-terabit-per-second high-speed cable assemblies remain in customer validation. This is a pre-commercial qualification stage and was not described as volume shipment or recognized revenue.
Zhaolong also said it can develop and supply active copper cables using mainstream signal-processing chips and customize designs for customers. It noted that achievable transmission distance depends on chip performance, cable structure, and data rate, and that ongoing development is focused on signal integrity and distance improvements.
- Zhaolong Interconnect (兆龙互连)
ASUS formed a multi-dozen-person AI Lab and raised its AI-server revenue growth target from 100% to 150%. #
ASUS established an AI Lab comprising several dozen employees. The team was described as an internal base for validating and developing AI servers.
The company raised its 2026 AI-server revenue growth target from 100% to 150%. Both figures are company targets rather than reported year-to-date growth or audited full-year results.
The source did not provide the underlying revenue base, expected absolute AI-server revenue, unit shipments, accelerator configurations, customers, or deployment schedule. It also did not disclose the AI Lab’s budget or exact headcount.
Cloverleaf Infrastructure partnered with NVIDIA to deploy DSX, targeting a gigawatt-scale U.S. AI-factory build-out. #
Cloverleaf Infrastructure formed a partnership with NVIDIA to deploy the NVIDIA DSX platform at AI infrastructure sites. Cloverleaf described its intended U.S. development program as a gigawatt-scale AI-factory build-out.
Separate reports said NVIDIA was in talks to invest in Cloverleaf, with one report placing the possible investment at several hundred million dollars. Because the sources describe talks, they do not establish that an equity investment had closed or that funds had been transferred.
The supplied sources do not identify specific sites, committed power capacity, data-center completion dates, installed accelerator counts, or contracted customers. The gigawatt-scale figure is therefore a build-out target rather than reported operational capacity.
- Cloverleaf Infrastructure
- NVIDIA
- DSX
China’s internet data services used 9.9 billion kWh in July, up 40.1%; January–July usage rose 43.3%. #
China’s internet data services consumed 9.9 billion kilowatt-hours in July, an increase of 40.1% year over year. Charging and battery-swapping services used 16.4 billion kWh in the same month, up 50.3%.
Total Chinese electricity consumption reached 1.04 trillion kWh in July, rising 1.7%. Industrial consumption was 605.5 billion kWh, including 122 billion kWh for high-technology and equipment manufacturing, which increased 8.9%.
For January through July, internet data services consumed 59.3 billion kWh, up 43.3%, while charging and battery-swapping services consumed 97.4 billion kWh, up 55.8%. Total national consumption over the seven months was 6.1399 trillion kWh, an increase of 4.7%.
PJM Interconnection is considering accelerating surplus interconnections; Midcontinent Independent System Operator and Southwest Power Pool studied 15 GW and 14 GW, respectively. #
PJM Interconnection is considering an option to accelerate its surplus interconnection pathway as it seeks additional capacity. Surplus interconnection service allows new generation to use unused portions of existing grid interconnection rights, subject to the applicable study and approval process.
PJM’s existing surplus interconnection process had produced limited results, according to the report. The source did not quantify how much capacity PJM had approved through the pathway or give a deadline for adopting the proposed acceleration.
During the first half of 2026, the Midcontinent Independent System Operator studied roughly 15 gigawatts under comparable surplus-interconnection activity, while the Southwest Power Pool studied about 14 gigawatts. These figures represent capacity under study, not completed generation, energized interconnections, or guaranteed additions to the grid.
- PJM Interconnection
- Midcontinent Independent System Operator
- Southwest Power Pool
ST Telemedia Global Data Centres withdrew from a planned Ho Chi Minh City joint venture; VNG seeks its interest. #
ST Telemedia Global Data Centres withdrew from a planned data-center project in Ho Chi Minh City, Vietnam. The project had been structured as a joint venture involving Vietnamese technology company VNG.
VNG is seeking to acquire ST Telemedia Global Data Centres’ portion of the venture. The source did not state that the transfer had closed or disclose a purchase price, financing terms, or regulatory conditions.
The report said VNG aims to develop the project and other facilities instead. No operational megawatt capacity, construction budget, server capacity, customer commitments, or revised completion date was provided.
AWS Glue 6.0 combines Apache Spark 4.1, Python 3.12, and Iceberg v3 support with 30% lower pricing. #
Amazon Web Services made AWS Glue 6.0 available with a modernized runtime based on Apache Spark 4.1. The release also uses Python 3.12 and Scala 2.13.
AWS Glue 6.0 adds full support for Apache Iceberg version 3. Apache Iceberg is the table-format component referenced by the release, while Spark and Python form parts of the processing and programming runtime.
AWS stated that Glue 6.0 pricing is 30% lower than pricing for previous AWS Glue versions. The supplied source did not provide workload-specific benchmarks, regional price tables, or a comparison of total job costs after differences in execution time.
- AWS Glue 6.0
- Apache Spark 4.1
- Python 3.12
- Iceberg v3
Google Cloud extended flexible committed-use discounts to G2 NVIDIA L4 and G4 RTX Pro 6000 virtual machines. #
Google Cloud made Compute Flexible Committed Use Discounts, or Flex CUDs, available for G2 virtual machines using NVIDIA L4 graphics processing units and G4 virtual machines using NVIDIA RTX Pro 6000 GPUs. The change extends flexible spend commitments to those two GPU-accelerated VM families.
Google said customers can retain flexibility across VM families and migrate between regions under the commitment structure. A single spend commitment can also combine general-purpose compute, Google Kubernetes Engine, Cloud Run, and eligible G2 and G4 GPU VMs.
The supplied announcement did not state the exact discount percentages, commitment duration, regional availability, or minimum spending requirement. It described the commercial structure as a way to lock in savings while retaining service and regional flexibility.
Alibaba’s T-Head Zhenwu M890 chip (真武 M890) serves more than 650 external customers across over 20 industries through Alibaba Cloud. #
Alibaba management said the latest T-Head (平头哥) Zhenwu M890 (真武 M890) chip had been commercialized through Alibaba Cloud. More than 650 external customers across over 20 industries were reported to be using services based on the chip.
The cited customer sectors included internet services, autonomous driving, and financial services. The disclosure referred to external customers accessing the chip through Alibaba Cloud rather than to 650 direct physical-chip purchasers.
The information appeared in an analyst report discussing Alibaba’s quarterly results and management commentary. The source did not disclose chip shipment volumes, manufacturing process, accelerator count per deployment, pricing, performance benchmarks, or revenue attributable specifically to the M890.
Alibaba Cloud added GLM-5.3 and DeepSeek-V4-Pro application programming interfaces to its Qwen AI platform. #
Alibaba Cloud added the official versions of GLM-5.3 and DeepSeek-V4-Pro to the Qwen AI Platform (千问 AI 平台). Corresponding application programming interface services were opened so developers can call the models directly.
The expansion was described as part of Alibaba Cloud’s Model-as-a-Service, or MaaS, model-service portfolio. DeepSeek-V4-Pro, Qwen3.8-Max, and the Qwen-Audio model family were also reported as available through subscription.
Subscribers can switch among supported models in AI tools including Qoder and Codex. The source did not disclose API prices, context-window sizes, throughput limits, regional availability, or benchmark comparisons.
- Alibaba Cloud
- GLM-5.3
- DeepSeek-V4-Pro
- Qwen AI platform
NVIDIA AVO achieved 100% on ARC-AGI-3, demonstrating its long-horizon autonomous-agent architecture. #
NVIDIA reported that AVO reached a score of 100% on ARC-AGI-3. NVIDIA presented the result as a demonstration of a frontier-level, general-purpose architecture for long-horizon autonomous agents.
The accompanying description emphasized that a frontier language model is only one component of an AI agent. NVIDIA identified the surrounding agent system, often called a harness, as the mechanism that determines how the model receives and acts on information.
The result is a claim published by NVIDIA on its Developer Blog. The supplied source data did not include comparative model scores, evaluation cost, number of attempts, compute consumption, or independent replication details.
SenseTime open-sourced SenseNova U1.5 Lite, supporting native 4K generation, image editing, layout control, and complex instructions. #
SenseTime formally open-sourced SenseNova U1.5 Lite, describing it as a lightweight, natively unified multimodal model. The company said the model is now globally available for direct deployment by developers and creators.
Training data, task design, and post-training processes were reworked around real-world visual tasks. The stated capabilities include native 4K visual generation, native image editing, complex-instruction following, text and layout handling, and fine-grained visual control.
SenseTime also positioned the release around higher controllability, image quality, and cost effectiveness. The source did not provide parameter count, model weights’ license terms, inference hardware requirements, benchmark results, or per-image generation costs.
- SenseTime
- SenseNova U1.5 Lite
A Rice University-led autonomous-laboratory initiative received $19.9 million for AI, robotics, and cloud-based electronic-materials research. #
Rice University received a $19.9 million award from the U.S. National Science Foundation for a national autonomous-laboratory initiative. SUNY Polytechnic Institute is participating as a project partner.
The initiative will combine artificial intelligence, robotics, and cloud-based laboratories. Its research objective is to accelerate the discovery and manufacturing of next-generation electronic materials.
The award amount applies to the Rice-led initiative rather than solely to SUNY Polytechnic Institute. The supplied announcement did not provide the project’s duration, institutional funding allocation, computing budget, cloud provider, or specific robotic and AI systems.
