AI & Compute Infrastructure — August 28, 2026

🇺🇸Federal Reserve Chair Kevin Warsh reaffirmed the 2% Personal Consumption Expenditures inflation target, citing insufficient improvement in underlying inflation. #

Warsh said the Federal Reserve’s 2% inflation objective, measured by the Personal Consumption Expenditures price index, remains a “firm and fixed” target. Although summer inflation reports were better than expected, he said they did not demonstrate a meaningful improvement in the underlying trend.

He stated that policymakers must be confident underlying inflation is moving toward the target “at a clear and sufficiently rapid pace”; otherwise, the Federal Reserve still has work to do. Warsh also said price stability should be the central bank’s predominant current focus and described labor-market conditions as consistent with full employment.

Warsh said consumer spending was healthy, labor markets were stable, and financial conditions were difficult to characterize as restrictive. Following the speech, the two-year U.S. Treasury yield finished 11.14 basis points higher at 4.3434%, while market pricing for a September 25-basis-point increase rose to 59.7%, according to the cited CME FedWatch update.

🇺🇸NVIDIA-backed Lambda raised approximately $1 billion of private short-term debt for GPUs destined for leasing to Microsoft. #

NVIDIA-backed artificial-intelligence cloud provider Lambda raised approximately $1 billion through a private short-term debt transaction. The financing was arranged by JPMorgan and marketed to private-placement investors, according to people familiar with the transaction cited in the Telegram report.

The proceeds are designated for purchases of NVIDIA graphics processing units. Those GPUs are intended to be leased to Microsoft as part of a Lambda-Microsoft compute project, distinguishing the transaction from general-purpose corporate financing.

The supplied reports describe the financing as raised rather than merely contemplated. They do not disclose the debt’s interest rate, maturity date, collateral package, amortization terms, GPU model, deployment site, or the duration and value of the associated Microsoft leasing arrangement.

• NVIDIA
• Lambda
• Microsoft

🇦🇺Blue Owl Capital led $2.4 billion of debt financing for IREN to fund Blackwell Ultra GPUs at its Canadian campus. #

A Blue Owl Capital fund led $2.4 billion of debt financing for cloud-computing provider IREN Ltd. The transaction consists of a $1.2 billion senior secured term loan and $1.2 billion of senior secured notes.

Both debt components carry a 9% interest rate and have a two-and-a-half-year term. The source describes the financing as supporting IREN’s purchase of NVIDIA Blackwell Ultra GPUs rather than funding undifferentiated data-center spending.

The compute expansion is associated with IREN’s Mackenzie data-center campus in British Columbia, Canada. The supplied report does not state the number of GPUs being purchased, the resulting megawatts of information-technology load, or a deployment-completion date.

• Blue Owl Capital
• IREN
• Blackwell Ultra GPUs

🇺🇸Volato (SOAR) disclosed a definitive merger with Alignment Engine; closing remains conditional, while preferred-share conversion requires shareholder approval. #

Volato Group, Inc. (SOAR) disclosed that it had executed a definitive merger agreement with Alignment Engine. The parties said they expected to close shortly after execution, but completion remained subject to the satisfaction or waiver of applicable closing conditions.

The merger closing itself is not conditioned on Volato stockholder approval. A later stockholder meeting will instead address the conversion of convertible preferred stock issued to Alignment Engine’s shareholders into Volato Class A common shares.

The filed press release describes Alignment Engine as developing infrastructure for energy-efficient artificial-intelligence workloads from a powered industrial campus in Ohio. It lists powered data-center infrastructure, high-performance GPU compute, advanced networking, AI training and inference infrastructure, proprietary technology, and capacity for other compute-intensive workloads as elements of Alignment Engine’s platform; these are company disclosures, not completed operating results independently verified by the filing.

• Volato (SOAR)
• Alignment Engine

🇺🇸FingerMotion (FNGR) completed the purchase of a 9.9% stake in Lyken.AI using 1,674,480 restricted shares, with no cash paid at closing. #

FingerMotion completed its acquisition of a 9.9% equity interest in Lyken AI Computing Inc., which operates as Lyken.AI, on August 17, 2026. The seller was Alset AI Ventures Inc., which retained a 90.1% controlling interest after the transaction.

Consideration consisted of 1,674,480 restricted FingerMotion common shares, and no cash was paid at closing. The transaction was presented in FingerMotion’s August 28 SEC-filed press release as the first executed step under the company’s new management team.

FingerMotion said the new team had set a direction of owning and operating enterprise AI compute capacity in North America alongside its existing mobile-data and telecommunications business in China. Its previously announced framework with BlueFlare for behind-the-meter AI and high-performance-computing infrastructure across Alberta, British Columbia, and Saskatchewan remains a memorandum of understanding, under which BlueFlare was contemplated as the primary development partner rather than disclosed as a completed deployment.

• FingerMotion (FNGR)
• Lyken.AI

SuperX AI Technology disclosed an Ezisight purchase order for 128 B300 AI servers, supporting its Australian market entry. #

Singapore-headquartered SuperX AI Technology Ltd. disclosed through a Form 6-K press-release exhibit that it had secured a purchase order from Ezisight Australia Pty Ltd. The order covers 128 B300 AI servers and was described by SuperX as marking its official entry into the Australian market.

SuperX describes its offerings as including high-performance AI servers, 800-volt direct-current infrastructure, high-density liquid-cooling systems, AI cloud services, and AI agents. Its services also cover data-center solution design, infrastructure-product integration, operations, and maintenance.

Ezisight trades as Ultimate AI Datacentre and describes itself as an Australian provider of green AI infrastructure and GPU-as-a-Service. The filed release says Ezisight focuses on modular edge-supercomputing facilities for large-model training and inference, with behind-the-meter renewable energy and water-positive cooling technology; the filing does not independently verify those promotional descriptions or state that all 128 servers have been delivered or commissioned.

• SuperX AI Technology
• Ezisight

TCL CSOT plans to enter optical communications through indium-phosphide laser chips, later evaluating optical components and high-speed modules. #

TCL CSOT (TCL华星) Chief Executive Officer Zhao Jun said at a TCL Technology investor meeting that the company plans to enter optical communications through the Shenzhen Huazhaoxingguang project. He linked the initiative to the expansion of artificial-intelligence data centers and divided the business plan into three stages.

The first stage will focus on indium-phosphide laser chips and establishing core technology and scaled manufacturing capability. TCL CSOT plans to reuse capabilities from its existing light-emitting-diode business, including factory design, chip manufacturing, facility systems, and its industrial cluster.

After laser-chip production matures, the second stage would extend into transmitter optical subassemblies and receiver optical subassemblies, commonly abbreviated as TOSA and ROSA. In the third stage, TCL CSOT would evaluate expansion into high-speed optical modules according to customer and market demand, making that module expansion an evaluation step rather than a disclosed operating deployment.

• TCL CSOT

Georgia Tech researchers reported 2.7-fold speedups using ferroelectric tuning for wafer-scale optical interconnects in mixture-of-experts training. #

Georgia Institute of Technology researchers published a technical paper examining thermal-tuning overhead in wafer-scale optical interconnects for large-language-model mixture-of-experts training. The analysis identified repeated tuning stalls during communication phases.

The researchers proposed ferroelectric-based tuning as the mitigation mechanism. Their paper reported speedups of 2.7 times in the studied setting, with the figure referring to the paper’s cross-layer analysis rather than a general performance claim for every mixture-of-experts workload.

The work specifically concerns tuning behavior in wafer-scale optical interconnects, not GPU arithmetic throughput or CPU performance. The source summary does not provide the tested model size, wafer dimensions, node count, power savings, or an independent reproduction of the result.

• Georgia Tech

Researchers demonstrated two-inch monolayer WS2 or MoS2 single-crystal growth using modified commercial sapphire surfaces. #

Researchers from Peking University, the Chinese Academy of Sciences, and other institutions reported a method for wafer-scale growth of single-crystalline two-dimensional semiconductors. The demonstrated materials were monolayer tungsten disulfide, or WS2, and molybdenum disulfide, or MoS2.

The method uses a homo-metal-element surface-modification strategy on commercial C-plane or M-plane sapphire. The cited elements include tungsten or molybdenum and aluminum, corresponding to the semiconductor material and sapphire surface system.

The reported growth scale was two inches. This is a crystal-growth result concerning monolayer semiconductor material on modified sapphire, not a disclosure of commercial chip production, transistor yield, packaged devices, or production-qualified wafer capacity.

• WS2
• MoS2

University of Michigan-Dearborn optimized liquid-cooling channels for a 2.7-kilowatt package containing two GPUs and one CPU. #

University of Michigan-Dearborn researchers published a technical paper on generative design of liquid-cooling channels for 2.5-dimensional and 3-dimensional integrated advanced packages. The work uses a physics-guided generative-design framework for cooling-channel topology optimization.

The evaluated package has a total thermal load of 2.7 kilowatts and contains two high-power graphics processing units and one central processing unit. The 2.7-kilowatt figure applies to the complete multi-chip package rather than to each individual processor.

The research focuses on liquid-cooling channel geometry within advanced packaging. The supplied abstract does not disclose commercial availability, coolant type, flow rate, inlet temperature, pressure drop, final device temperatures, or deployment in a production server.

Alignment Engine’s Ohio campus has 154 megawatts available and 480 megawatts of total capacity for planned artificial-intelligence infrastructure. #

Alignment Engine said its powered industrial campus in Ohio currently has 154 megawatts available and total capacity of 480 megawatts. The company presented the existing 154 megawatts as a foundation for deploying high-performance artificial-intelligence compute infrastructure.

The planned platform combines powered data-center infrastructure, high-performance GPU compute, advanced networking, proprietary technology, and systems supporting AI training and inference. It is also intended to support other compute-intensive workloads.

These capacity statements appeared in a press release filed with the SEC as part of Volato’s merger disclosure. The filing establishes that the company made the claims, but it does not state how much of the 154 megawatts is under customer contract, how much compute hardware has been installed, or when the campus would reach the stated 480-megawatt total.

• Alignment Engine’s Ohio campus

U.S. uranium concentrate production reached 2.1 million pounds of U3O8 in 2025, more than tripling 2024 output. #

U.S. uranium concentrate production totaled 2.1 million pounds of triuranium octoxide, or U3O8, in 2025. U3O8 is the uranium-concentrate form identified by the U.S. Energy Information Administration as the base component of nuclear-reactor fuel.

The 2025 production volume was more than three times the amount produced in 2024. It was also the highest annual U.S. U3O8 production level since 2017.

The measurement covers domestic uranium concentrate production and is expressed in pounds of U3O8. It does not represent enriched uranium, fabricated nuclear fuel, reactor output, or uranium-reserve volume.

South Africa’s Koeberg Nuclear Power Station became fully offline after Unit 1 safely shut down following a turbine trip. #

South African state-owned utility Eskom said Koeberg Nuclear Power Station Unit 1 shut down safely under established procedures after a turbine trip during the early hours of August 27. With Unit 1 no longer supplying electricity, the nuclear station was temporarily fully offline.

Eskom said Unit 1’s removal from the grid did not pose a risk to South Africa’s national electricity supply. The utility described the grid as continuing to operate safely and stably.

The event had no effect on nuclear safety and posed no risk to employees, the public, or the environment, according to Eskom’s statement. The cause of the turbine trip remained under investigation.

• Koeberg Nuclear Power Station
• Unit 1

China exported its first domestically produced F-class 50-megawatt heavy-duty gas turbine, shipping the unit to Kazakhstan. #

Dongfang Electric Group said a domestically developed F-class heavy-duty gas turbine departed Deyang, Sichuan, on August 28 for shipment to Kazakhstan. The turbine has a rated capacity of 50 megawatts.

The shipment was described as China’s first overseas export of a complete, fully domestically produced high-end heavy-duty gas-turbine unit. The distinction concerns export of the complete machine rather than an individual turbine component.

The source identifies the equipment class, output rating, origin, and destination. It does not provide the receiving power plant’s name, project commissioning date, fuel-consumption rate, thermal efficiency, contract value, or number of additional units.

Energy Transfer’s Hugh Brinson Pipeline targets September 1 operation, carrying approximately 2.2 billion cubic feet per day from the Permian Basin. #

Energy Transfer LP’s Hugh Brinson Pipeline in Texas was scheduled to begin operating on September 1. The pipeline is expected to carry approximately 2.2 billion cubic feet of natural gas per day from the Permian Basin toward East Texas.

The report said the new route could direct additional supply toward nearby Erath, Louisiana. Henry Hub, the delivery point associated with benchmark U.S. natural-gas futures, is located in that area.

Ahead of the planned start, October natural-gas futures settled at $2.888 per million British thermal units on the New York Mercantile Exchange. The contract fell 2.6 cents, or 0.9%, during the session.

• September 1

Alibaba Cloud launched its first Brazilian cloud region, using two data centers for cloud and agent-based artificial-intelligence services. #

Alibaba Cloud launched its first cloud region in Brazil. The region is composed of two new data centers, distinguishing the deployment from a single-facility edge location.

The facilities will provide cloud services to local businesses. Alibaba Cloud also plans to offer a suite of agent-based artificial-intelligence services through the Brazilian region.

The supplied report identifies the country, number of data centers, and service categories. It does not disclose the facilities’ cities, megawatt capacity, server or accelerator count, availability-zone layout, investment cost, or power sources.

• Alibaba Cloud

NVIDIA TensorRT Model Connect enables open-model deployment from checkpoint to inference using two commands. #

NVIDIA introduced TensorRT Model Connect as a way to move an open model from a checkpoint to inference using two commands. The tool is aimed at simplifying the path between obtaining model weights and running the model in an inference environment.

NVIDIA noted that deploying rapidly changing open models in native applications can otherwise require model-specific conversion and preprocessing. TensorRT Model Connect addresses that deployment workflow rather than introducing a new GPU, accelerator, server, or cloud region.

The supplied article summary does not provide benchmark results, supported-model counts, latency figures, throughput, memory requirements, or hardware-specific performance comparisons. The “two commands” claim therefore describes the documented workflow and not a two-step guarantee for every application integration.

• NVIDIA TensorRT Model Connect

BigQuery continuous queries added preview stateful processing, enabling joins, aggregations, and window functions within streaming queries. #

Google Cloud made stateful processing available in preview for BigQuery continuous queries. The feature allows continuous streaming queries to retain and use state while processing incoming data.

Supported stateful operations include joins, aggregations, and windowing functions directly within streaming queries. Google gave the example of calculating a 30-minute average over time.

The resulting real-time signals can be sent to downstream applications and AI agents. Because the feature is in preview, the source does not describe it as generally available.

Google Cloud’s Managed Service for Apache Kafka launched a generally available generator that streams synthetic data within two minutes. #

Google Cloud made a synthetic-data generator generally available for its Managed Service for Apache Kafka. The tool is designed to populate a newly created Kafka cluster with mock data without requiring the user to modify a client application or start a separate virtual machine.

Google said users can begin sending synthetic data in three clicks. The service can start streaming data to the cluster in less than two minutes.

The generator is positioned as a testing utility for clusters and new features. The announcement concerns synthetic test-data generation, not production-data ingestion performance or a new Kafka cluster-capacity tier.

Anthropic’s Claude developed and validated control logic that restored a QuEra quantum-computer laser system within seconds. #

QuEra Computing reported that an AI agent using Anthropic’s Claude developed and validated control logic for the laser system in a QuEra quantum computer. The resulting logic recovered the laser system within seconds.

QuEra contrasted that recovery time with the minutes required by a human expert. It also said the automated system held the laser steadier than a specialist’s manual tune.

The work concerned laser-system control and recovery inside a QuEra quantum computer, not the execution of a quantum algorithm by Claude. QuEra said it plans to extend the approach to other subsystems, making those additional applications plans rather than completed deployments.

• Anthropic’s Claude

SkyProduction (天工工作台) added Alibaba’s Wan 3.0, with pricing starting August 31 at RMB 0.05 per video-second. #

Kunlun Tech’s AI platform SkyProduction (天工工作台) integrated Alibaba’s Wan 3.0 video-generation model. The platform opened the model to all users free of charge for a limited three-day period from August 28 through August 30.

Paid usage is scheduled to resume on August 31, with pricing as low as RMB 0.05 per generated video-second. At that minimum rate, the source calculated that a 30-second video would cost RMB 1.50.

The quoted figure is usage pricing per output video-second, not an hourly GPU rental price or an infrastructure cost. The source described it as the lowest video-model price across the internet, but that characterization was part of the report rather than an independently documented market-wide comparison.

• SkyProduction (天工工作台)
• Alibaba’s Wan 3.0
• August 31

Industrial and Commercial Bank of China reported deploying large models across more than 600 business scenarios. #

Industrial and Commercial Bank of China (中国工商银行) Vice President Zhao Guide discussed the bank’s use of artificial intelligence during its 2026 interim-results presentation on August 28. He said the bank had cumulatively implemented more than 600 large-model business scenarios.

According to Zhao, the deployments were being used to make the bank’s support for the real economy more precise. He also said they were improving operating and management efficiency and making risk prevention and control more proactive.

The figure counts business scenarios rather than individual models, users, servers, or inference requests. The source does not disclose model suppliers, parameter counts, accelerator hardware, data-center capacity, implementation spending, or the number of scenarios currently in full production.