🇺🇸GPU-cloud operator WhiteFiber (WYFI) disclosed over $550 million in recent cloud contract value, including signed agreements not yet generating revenue. #
WhiteFiber’s September investor presentation, disclosed through SEC EDGAR on September 14, reports more than $550 million in recent cloud total contract value (TCV). The presentation identifies wins from May through August 2026 across Canada, Iceland, U.S. owned-fleet deployments, and Paris.
The company explicitly states that the figure includes customers with signed definitive agreements that are not yet generating revenue. It therefore represents contract value, not revenue already recognized or cash already collected; the excerpt does not provide a revenue-conversion schedule.
The presentation documents WhiteFiber’s development from a GPU-cloud business launched in January 2024 into an integrated cloud and data-center platform. It lists the acquisition of Enovum in October 2024, the WhiteFiber brand launch in February 2025, and an initial public offering in August 2025.
WhiteFiber separately reports 7,500+ NVIDIA GPUs contracted and 20+ customers, with the customer count also including signed customers not yet generating revenue. Its approximately 1.5-gigawatt data-center pipeline consists of acquisition opportunities under management evaluation—not operating or committed capacity—and the SEC disclosure establishes that WhiteFiber reported these figures, rather than independently validating them.
🇺🇸Anthropic signed a $13.7 billion, six-year compute-procurement agreement with Rum Group, formerly Rumble. #
Financial_Express reports that Anthropic signed a $13.7 billion agreement covering six years of compute procurement. The message identifies the counterparty as Rum Group, formerly Rumble, preserving the distinction between the company’s current and former names.
The same message states that, including this transaction, Anthropic’s compute-procurement agreements over the preceding year totaled at least $517 billion. That is a reported aggregate agreement value, not an amount described as already spent or recognized as supplier revenue.
The source describes a signed procurement agreement, rather than a financing round or preliminary memorandum. However, the supplied material contains no underlying contract or regulatory filing, and does not specify hardware, power capacity, deployment dates, payment milestones, termination conditions, collateral, or equity issuance.
🇺🇸Water-from-air technology company AirJoule acquired BitSink, a manufacturer of cooling, electrical-distribution, and racking infrastructure for AI data centers. #
AirJoule Technologies Corporation announced on September 14 that it had acquired BitSink. The supplied announcement describes an acquisition already made, rather than an intention to negotiate or a transaction awaiting a stated closing date.
AirJoule’s existing business is water-from-air platform technology, while BitSink is a U.S.-based designer and manufacturer of data-center infrastructure. BitSink’s stated product scope covers cooling, electrical distribution, and racking for AI and high-performance computing (HPC) data centers.
The transaction therefore adds physical data-center infrastructure capabilities to AirJoule’s stated technology base; it is not described as an acquisition of a GPU-cloud operator. The supplied excerpt gives no purchase price, financing structure, share consideration, dilution, customer contracts, manufacturing capacity, or integration timetable.
🇯🇵Memory manufacturer Kioxia (铠侠) reportedly considers raising at least $10 billion through American depositary receipts; discussions remain preliminary. #
Kioxia Holdings (铠侠控股), the Tokyo-based memory-chip manufacturer, is reportedly considering raising at least $10 billion through American depositary receipts (ADRs). Financial_Express attributes the information to people familiar with the discussions, rather than a company announcement or an offering filing.
The report names Bank of America, Goldman Sachs, and JPMorgan among banks involved in discussions about a possible offering in 2027. Both the offering size and participating bank group remain subject to change.
The same report says Kioxia previously repurchased billions of dollars of shares in Japan and is considering the U.S. listing to obtain additional liquidity. No exact repurchase amount is supplied, and the report does not establish a finalized issuance structure.
This remains a preliminary capital-markets proposal, not a completed fundraising or confirmed listing. The supplied source provides no ADR ratio, offering price, share count, dilution estimate, exchange selection, or detailed allocation of proceeds.
🇺🇸The 10-year Treasury yield reached 5.0121% intraday, before retreating to 4.9731% in late New York trading. #
The U.S. 10-year benchmark Treasury yield reached 5.0121% at 22:26 Beijing time on September 14, according to the supplied market report. It subsequently retreated to 4.9731% in late New York trading, when it was still 0.62 basis points higher on the day.
For historical comparison, the source cites the 5.0187% intraday high on October 23, 2023. The September 14 peak remained below that level; other supplied messages describe the session as the first move above 5% since 2023.
The two-year Treasury yield also rose, reaching 4.6451% in late New York trading, up 1.97 basis points for the session. It had increased for six consecutive trading days, accumulating a 28.11-basis-point rise, and traded within a daily range of 4.5984%–4.6769%.
Fujitsu scheduled global sales of its FUJITSU-MONAKA CPU for November 2026, featuring 2nm technology and 3D stacking. #
Fujitsu announced that global sales of its next-generation FUJITSU-MONAKA central processing unit (CPU) would begin in November 2026. This is a scheduled commercial launch, not a statement that global sales had already started on September 14.
The announcement identifies the processor as designed and developed in Japan and describes 2-nanometer technology and three-dimensional stacking. Its headline also identifies a server launch, while the description refers to a server integrated, developed, and manufactured in Japan.
Fujitsu positions the CPU and server around AI inference and describes their performance as world-class. The supplied excerpt contains no benchmark results, core counts, power measurements, prices, or shipment volumes, so that performance characterization remains an attributed company claim.
- Fujitsu
- FUJITSU-MONAKA CPU
- November 2026
Custom-chip designer Alchip will showcase 3nm production, 2nm design, and multi-die packaging capabilities at the September 15–17 AI Infra Summit. #
Alchip Technologies announced plans to showcase its process-technology and advanced-packaging achievements at AI Infra Summit 2026. The event runs September 15–17 at the Santa Clara Convention Center, with Alchip exhibiting at Booth 1705.
The announcement distinguishes 3nm production capabilities from 2nm design capabilities. It does not describe 2nm production as already underway, and those two stages should not be treated as equivalent manufacturing milestones.
The presentation also covers application-specific integrated circuit (ASIC) design, three-dimensional integrated circuit (3DIC) capabilities, and multi-die advanced packaging. The supplied excerpt names no customer designs, production volumes, yields, package specifications, or new commercial contracts.
Kyocera (京瓷) and Tohoku University (东北大学) developed laser annealing that integrates optical isolators onto silicon-photonics chips without damaging surrounding components. #
Kyocera (京瓷) and scientists at Japan’s Tohoku University Research Institute of Electrical Communication (东北大学电气通信研究所) developed a new laser-annealing technique. The university referenced here is Japan’s Tohoku University.
The reported technique heats only the region requiring treatment, allowing optical isolators to be integrated directly onto silicon-photonics chips without damaging neighboring components. This is a chip-integration research result, rather than an announcement of a complete optical networking system.
Financial_Express, citing Science and Technology Daily (科技日报), reports that the associated paper appeared in IEEE Access, a journal of the Institute of Electrical and Electronics Engineers. The supplied message contains no processing temperatures, optical-performance measurements, manufacturing yields, or commercialization schedule.
- Kyocera (京瓷)
- Tohoku University (东北大学)
Canadian telecom operator Bell (贝尔) announced plans to expand its Saskatchewan AI data-center project to 1.2 gigawatts. #
Bell’s (贝尔) chief executive announced plans to expand the company’s Saskatchewan AI data-center project to 1.2 gigawatts. A separate message quoting Saskatchewan Premier Moe describes the expansion as four times the previous project scale.
The capacity figure concerns a planned data-center project, not a reported GPU count or currently operating compute fleet. The supplied messages do not identify whether the 1.2 gigawatts represents gross site power or information-technology load, and provide no energization or customer-deployment schedule.
The capital-investment messages contain a material currency conflict: one states C$50 billion (500 亿加元), while another states US$50 billion (500 亿美元). Both describe required project investment rather than money already raised or spent, but the currency cannot be resolved from the supplied material.
Constellation Energy agreed to buy a 609-megawatt Rhode Island gas plant from Shell; the acquisition has not been described as closed. #
Utility Dive reports that Constellation Energy entered a deal to buy a 609-megawatt natural-gas power plant in Rhode Island from Shell. The supplied source describes an acquisition agreement, not a completed transfer of ownership.
The asset adds to Constellation’s portfolio in ISO New England, the regional independent system operator’s market. The 609 megawatts is the plant’s reported generation capacity; it is not a dedicated AI-data-center power allocation or a compute-capacity figure.
In a separate transaction, Shell acquired a 169-megawatt Pennsylvania gas portfolio to expand its position in the PJM Interconnection electricity market. The excerpt supplies no price, closing timetable, financing terms, or dedicated data-center customer for Constellation’s Rhode Island purchase.
- Constellation Energy
- Shell
Oracle (甲骨文) clarified that New Mexico’s court stay concerns adjacent microgrid air permitting, not its Project Jupiter (木星计划) data-center campus. #
Oracle (甲骨文) said the New Mexico Supreme Court’s stay of an administrative air-quality permitting process concerns a neighboring microgrid project. According to Oracle, the order does not cover the Project Jupiter (木星计划) data-center campus.
The company describes the data center and microgrid as separate facilities on adjacent but distinct parcels. It also says the stay concerns the microgrid’s air-permitting process, rather than county-level building permits.
Oracle further argues that the data-center buildings themselves do not emit air pollutants and therefore do not qualify as a pollution source under New Mexico law. These are Oracle’s reported descriptions and legal position; the supplied material does not include the court order itself, while a separate message says Oracle will continue working with the state on the project.
- Oracle (甲骨文)
- Project Jupiter (木星计划)
Google is exploring a potential data center in Lea County, New Mexico; discussions continue, with project details unresolved. #
Google said it is exploring a potential data center in Lea County, New Mexico. Both financialjuice and Financial_Express carried reports of the company’s consideration of the site.
The more detailed message states that discussions remain ongoing and many details have not been finalized. This is an exploratory project update, not a confirmed construction start, completed financing, or operating-capacity announcement.
The supplied reports give no investment amount, power requirement, hardware specification, customer commitment, or opening date. They identify this as Google’s Lea County proposal, distinct from the separately reported Oracle Project Jupiter permitting matter.
WhiteFiber began initial billing at its NC-1 data-center campus in August 2026, alongside active customer deployment. #
WhiteFiber’s SEC-disclosed presentation identifies August 2026 as the start of active customer deployment and initial billing at NC-1. The company acquired the campus in May 2025, making the billing update an operating milestone distinct from its development plans.
The presentation also records a December 2025 Nscale anchor contract for 40 megawatts of information-technology (IT) load, with approximately $865 million in total contract value. It does not state that all 40 megawatts were deployed or billing in August, nor does it quantify NC-1’s initial billing amount.
WhiteFiber describes NC-1 as having planned total power capacity of at least 99.0 gross megawatts by May 2029, with additional expansion capacity potentially available. This future gross-power figure is different in both scope and timing from the Nscale contract’s IT-load figure.
Across its platform, WhiteFiber reports approximately 11 gross megawatts online in 2025 and estimates approximately 70 gross megawatts online by year-end 2026. The presentation also cites buildout costs of approximately $8 million–$10 million per gross megawatt; the year-end capacity and cost statements are company estimates, not independently verified completed deployment or expenditure.
Google Cloud Dataflow made batch-job Pause/Resume generally available, allowing failed long-running jobs to resume instead of restarting. #
Google Cloud announced the general availability of Pause/Resume for Dataflow batch jobs. Dataflow supports both batch and streaming pipelines, but this feature’s stated scope is batch processing.
Google explains that some customer batch jobs run for several days and that, before this feature, failed jobs required a complete retry because users could not access the work already processed. Pause/Resume allows customers to resume failed long-running jobs rather than restart them from scratch.
The feature also permits deliberate pauses and resumptions according to business requirements. Google identifies reallocating accelerated resources—such as GPUs and tensor processing units (TPUs)—from lower-priority batch jobs to higher-priority feature-engineering or inference workloads as a supported use case, without providing a quantified savings benchmark.
- Google Cloud Dataflow
- Pause/Resume
Dataflow added G4 virtual-machine support with NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs, offering 96GB memory and 1.6 TB/s bandwidth. #
Google Cloud announced Dataflow support for G4 virtual machines (VMs) powered by NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs. The announcement adds this option to Dataflow’s existing support for multiple GPU and TPU types used in AI inference.
Google specifies 96GB of virtual GPU (vGPU) memory and 1.6 terabytes per second of bandwidth for the described GPU offering. These are GPU memory and bandwidth specifications, not total VM system memory, network throughput, or aggregate cluster capacity.
The blog says customers can run inference inside Dataflow jobs using models described as “up to 70B+” parameters. It also identifies continued support for native capabilities including RunInference, right fitting, and GPU-enabled autoscaling, while claiming performance gains over NVIDIA L4 without supplying a numerical comparison in the excerpt.
- Dataflow
- G4 virtual-machine
NVIDIA expanded CUDA-Q with CUDA-Q Logical, an open orchestration layer for designing and testing fault-tolerant quantum-computing applications. #
NVIDIA announced an expansion of its open-source CUDA-Q platform through the addition of CUDA-Q Logical. The supplied announcement describes the addition as an open, programmable orchestration layer for the quantum-computing industry.
Its stated purpose is to support the design and testing of fault-tolerant quantum-computing applications. The announcement names drug discovery, financial modeling, and materials development as application areas, rather than reporting completed commercial results in those fields.
This is a software-platform announcement, not the deployment of a new quantum processor or a demonstrated fault-tolerant machine. The supplied excerpt provides no logical-qubit count, error-rate measurement, application benchmark, or hardware-capacity figure.
- NVIDIA
- CUDA-Q
- CUDA-Q Logical
Perplexity Portable Computer, its local multistep-task agent, became available on Windows RTX PCs, using local models accelerated by NVIDIA GPUs. #
NVIDIA announced that Perplexity Portable Computer is available on Windows RTX PCs. Portable Computer is the local version of the Perplexity Computer agent, rather than a newly announced physical computer.
The agent plans and carries out multistep tasks, using local models accelerated by NVIDIA GPUs. The supplied NVIDIA description specifically identifies data analysis and bringing together information among its functions.
NVIDIA frames local execution as a way to perform work on a PC while keeping sensitive information on the device. The supplied excerpt does not provide supported model names, minimum GPU-memory requirements, pricing, performance benchmarks, or a guarantee that every possible task operates without cloud access.
- Perplexity Portable Computer
- Windows RTX PCs
- NVIDIA
Waymo launched paid robotaxi service in Las Vegas, its 15th U.S. service city, initially deploying dozens of vehicles. #
Alphabet’s Waymo announced the launch of paid robotaxi service in Las Vegas on September 14, making it the company’s 15th U.S. service city. Access is being opened in phases, with dozens of vehicles initially deployed and a stated plan to expand to hundreds.
The initial Las Vegas service does not include highway access. Airport trips and highway service are described as additions for later phases, rather than capabilities available at launch.
Moove manages the Las Vegas operating fleet and also manages Waymo vehicle operations in Phoenix and Miami. The same report says Waymo has more than 4,000 robotaxis across the United States and provides more than 500,000 paid rides per week; those figures are nationwide, not Las Vegas-specific.
Waymo had previously stated a goal of operating in 20 cities globally and providing one million paid rides per week in 2026. These remain company targets, distinct from the reported operating figures, while the message also notes preparations for testing in London and Tokyo.
Apple began releasing Siri AI in English beta; initial availability excludes EU users on iOS, iPadOS, and watchOS. #
Apple began rolling out Siri AI in English beta on September 14, supported by a new generation of Apple Intelligence and updates beginning with iOS 27. The reported language-expansion schedule adds French, Japanese, Korean, Portuguese, and Spanish in October, which is a rollout plan rather than current availability.
Apple describes Siri AI as offering personal-context understanding, screen awareness, broader knowledge, and additional system-level application actions. Reported capabilities include answering questions and retrieving relevant information from a user’s messages, email, and photos; these are Apple’s stated capabilities, not independent test results.
Initial availability excludes European Union users on iOS, iPadOS, and watchOS. A separate financialjuice message attributes to Apple the statement that Siri AI and Apple Intelligence are unavailable in China because of regulatory requirements, while another supplied message reports Siri AI’s addition to macOS 27.
The device-compatibility report distinguishes operating-system support from full AI-feature eligibility: iOS 27 supports iPhone 11 and later, plus second-generation iPhone SE and newer models, while full Siri AI capabilities initially target Apple Intelligence-compatible devices, described as including iPhone 15 Pro and newer models. The software release also includes new child-protection tools in iOS 27, iPadOS 27, and macOS 27, separately from the Siri AI rollout.
