🇺🇸Nebius (NBIS) entered a $775 million facility at Secured Overnight Financing Rate plus 2.50%, collateralized by GPU infrastructure and contracted customer cash flows. #
Nebius Group N.V. disclosed that it entered its first senior secured debt facility on July 10, 2026. The facility has an aggregate principal amount of approximately $775 million and matures on October 31, 2030.
Interest is set at the Secured Overnight Financing Rate (SOFR) plus 2.50%. The collateral consists of deployed graphics processing unit (GPU) infrastructure and contracted cash flows from a customer agreement with an investment-grade counterparty.
Nebius said it intends to use the proceeds to continue expanding its AI cloud infrastructure. The company separately described the financing vehicle as backed by the deployed hardware and contracted customer cash flows.
🇺🇸Nebius sold 12.7 million shares through June 30 at a $223.60 weighted-average price, generating approximately $2.8 billion in gross proceeds. #
Nebius disclosed that it began using its at-the-market, or ATM, equity program during the quarter. Through June 30, 2026, the company sold 12.7 million Class A shares under that program.
The shares were sold at a weighted-average price of $223.60 per share. Nebius reported approximately $2.8 billion in gross proceeds, a figure stated before any applicable issuance costs or other deductions.
As of June 30, another 12.3 million shares remained available under the ATM program. Nebius also said it expected more than $9 billion in customer prepayments during 2026 and reported more than $40 billion of customer commitments; those amounts were presented by the company alongside its financing plans and forward-looking disclosures.
🇫🇷WhiteFiber (WYFI) holds a binding France-project financing term sheet, but definitive agreements remained unsigned at filing. #
WhiteFiber disclosed a binding term sheet for project-level financing associated with its planned AI compute deployment in France. The company had also secured third-party French data-center capacity for the project.
The term sheet is not a completed financing. As of the Form 10-Q filing date, definitive agreements had not been executed, and certain material terms remained under negotiation.
WhiteFiber said that, if consummated, the financing was expected to be incurred by a project-level subsidiary and would not be guaranteed by WhiteFiber, Inc. The company expected the project to be supported by the financing and customer prepayments, including 12 months of advance service fees, but explicitly said there was no assurance that financing would close on the anticipated timeline, contemplated terms, satisfactory alternative terms, or at all.
🇦🇺SharonAI Holdings filed an S-1, presenting itself as an Australian neocloud serving enterprise, government, and research customers. #
SharonAI Holdings Inc., associated in the filing with symbols SHAZ and SHAZW, filed an S-1 registration statement. It described itself as an Australian neocloud operator built for artificial intelligence and high-performance computing (HPC), including training, inference, machine-learning, and generative-AI workloads.
The company said it provides enterprise, government, and research organizations with sovereign, low-latency access to accelerated computing hardware. Its filing identified NVIDIA B200 and B300 GPUs, as well as anticipated GB300 GPUs, among the hardware addressed by its platform.
SharonAI listed NVIDIA, NEXTDC, Cisco, World Wide Technology, Lenovo, VAST Data, and Megaport as strategic partners. It also said it is an NVIDIA Cloud Partner within the NVIDIA Partner Network and identified Canva and GMI Cloud US Inc. as recent commercial wins.
The current operating subsidiary, SharonAI, Inc., was incorporated in Delaware on February 15, 2024. The prospectus also states that post-merger information was retrospectively adjusted for a 1-for-50 reverse split of outstanding common shares, including corresponding conversion-price and exercise-price data unless otherwise indicated.
🇺🇸The State Department launched a Pax Silica pilot in Panama to accelerate allied trade and logistics for AI products. #
The U.S. State Department unveiled an initiative intended to accelerate trade in AI products among allied countries. The reporting identified the program as part of the department’s Pax Silica initiative.
The initiative focuses specifically on accelerating logistics for crucial AI products. The supplied messages do not identify individual products, participating companies, funding amounts, or targeted shipment volumes.
Panama was identified as the location for the pilot program. The source described this as the starting pilot rather than a broad, completed rollout across multiple countries.
- State Department
- Pax Silica
Kioxia and Sandisk unveiled ninth-generation 2-terabit quad-level-cell 3D flash, designed for AI-infrastructure storage demand. #
Kioxia Corporation and Sandisk Corporation unveiled a ninth-generation three-dimensional flash-memory technology. Kioxia Corporation is identified as a subsidiary of Kioxia Holdings Corporation.
The new device has a density of 2 terabits and uses quad-level-cell (QLC) flash technology. The announcement characterizes it as high-performance 3D flash memory.
The companies designed the technology to support growing storage demand from AI-driven infrastructure. Their announcement framed the product as an architectural response to the rate at which storage requirements are increasing alongside AI infrastructure.
Hon Hai plans to begin shipping NVIDIA Vera Rubin platforms in the fourth quarter of 2026. #
Hon Hai was reported as planning to start shipments of NVIDIA Vera Rubin platforms. The stated timing is the fourth quarter of 2026.
The report describes shipments of platforms, rather than individual GPU deliveries, data-center capacity, or customer deployments. It does not provide a shipment volume, customer list, manufacturing location, or revenue value.
The timing is a company plan reported before the start of shipments. The source does not state that commercial shipment or customer acceptance had already occurred.
- Hon Hai
- NVIDIA Vera Rubin platforms
AMD’s Instinct MI455X accelerator uses the CDNA 5 architecture, underpinning its next-generation servers and Helios rack-scale system. #
AMD’s Instinct MI455X is an accelerator based on the company’s CDNA 5 architecture. ServeTheHome described CDNA 5 as marking the next era of AMD’s Instinct product line.
The accelerator is positioned as the backbone of AMD’s next-generation AI server offerings. This identifies the MI455X as an accelerator within server systems, rather than as a central processing unit or a complete server by itself.
The architecture also underpins AMD’s Helios rack-scale system. The cited source characterizes Helios as a large rack-scale system but does not provide deployment quantities, customer contracts, pricing, or availability figures in the supplied description.
- AMD’s Instinct MI455X accelerator
Ascend Analytics estimates over 80% of new large Texas loads seeking interconnection lack matching generation online by 2030. #
Ascend Analytics estimated that more than 80% of new large electrical loads seeking interconnection in Texas will not have matching generation online by 2030. The estimate concerns generation availability corresponding to prospective large loads, not the share of all existing Texas demand.
The assessment comes as data-center and industrial electricity demand continues to increase. Utility Dive reported the estimate in the context of supply constraints limiting peak-demand growth within the Electric Reliability Council of Texas market.
Texas had also reached a new peak-demand record. The cited report nevertheless identifies generation availability as a constraint on how quickly additional large loads can be served.
PJM, the regional grid operator, is considering ride-through standards after 3.8 gigawatts of Northern Virginia data-center load tripped offline. #
PJM Interconnection is considering reliability requirements after 3.8 gigawatts of data-center load tripped offline in Northern Virginia. The event occurred on July 22, 2026.
PJM described the incident as the latest and largest event of its kind in the grid operator’s history. The 3.8-gigawatt figure refers to electricity demand that disconnected, not lost generation capacity.
The operator is considering “ride-through” standards intended to govern how loads respond during grid disturbances. The broader standards discussion identified in the article covers both data centers and cryptocurrency-related loads.
Virginia regulators ordered Dominion to directly assign certain transmission costs to data centers and may examine upstream costs. #
The Virginia State Corporation Commission ordered Dominion Energy to directly assign certain transmission costs to data centers. Direct assignment associates the covered costs with the relevant data-center customers rather than describing them solely as general system costs.
The order applies to certain transmission expenses; the supplied source does not state that every transmission cost will be directly assigned. It also does not provide a total dollar amount or a per-customer allocation.
The commission said an upcoming docket may consider whether the same policy “could or should” apply to more upstream transmission costs. That broader treatment remains a matter for examination rather than a completed order in the cited report.
The U.S. Energy Information Administration forecasts marketed natural-gas production averaging 122.5 billion cubic feet daily in 2026, above 2025’s 118.5 billion cubic feet daily record. #
The U.S. Energy Information Administration forecasts that marketed U.S. natural-gas production will average 122.5 billion cubic feet per day in 2026. The estimate appears in the agency’s August 2026 Short-Term Energy Outlook.
The forecast exceeds the previous annual record of 118.5 billion cubic feet per day, set in 2025. The difference between the two annual averages is 4.0 billion cubic feet per day.
The measure is marketed natural-gas production and is expressed as an average daily rate across the year. Because the 2026 figure is an EIA forecast, it is not a completed full-year production result.
- U.S. Energy Information Administration
Con Edison plans 28 new substations by 2035 to support electrification-driven growth in New York. #
Con Edison plans to add 28 substations by 2035. The plan concerns new utility substations rather than data centers, generating plants, or individual server facilities.
The New York-based utility tied its growth plans to electrification. It said this approach aligns with New York city and state climate policies.
The source characterizes the substations as planned infrastructure, not completed or already energized assets. It does not provide individual substation capacities, locations, construction schedules, or project costs in the supplied description.
WhiteFiber scheduled 216 NVIDIA GB200 GPUs to begin deployment and revenue generation in July under a 12-month service order. #
WhiteFiber entered a service order in April 2026 to provide services using 216 NVIDIA GB200 GPUs. The order’s initial term is 12 months beginning on the service-commencement date.
After the initial term, the service order automatically renews for successive one-month periods unless either party terminates it. This renewal mechanism does not convert the disclosed initial commitment into a longer fixed term.
WhiteFiber’s quarterly filing said deployment and revenue generation were scheduled to begin in July 2026. It associated the order with total revenue of up to $6.8 million, making that figure a maximum disclosed contract amount rather than revenue already recognized.
The filing uses scheduled deployment language and does not state in the cited passage that all 216 GPUs had been accepted or were generating revenue as of the filing date. WhiteFiber also disclosed other service orders and customer terminations during the broader reporting period.
WhiteFiber’s five-year Paris-region contract exceeds $160 million; service is expected in September, subject to equipment delivery and acceptance. #
WhiteFiber entered a five-year agreement in May 2026 to provide AI compute infrastructure to an investment-grade technology customer in the Paris region. The deployment is to use advanced NVIDIA GPU systems.
The company reported a total contract value exceeding $160 million. That amount is the stated contract value over the agreement’s term, not revenue already recognized or cash already received.
Service had previously been expected to commence in July 2026 but was moved to an expected September 2026 start. The revised commencement remains subject to final equipment delivery and acceptance milestones.
WhiteFiber said it secured third-party data-center capacity in France to support the deployment. It also expects customer prepayments, including 12 months of advance service fees, to support the project, while the related project-level financing remained at the binding-term-sheet stage without definitive agreements at filing.
SharonAI secured up to 87 megawatts through NEXTDC, hosting most GPU infrastructure with that data-center provider. #
SharonAI disclosed that it had secured up to 87 megawatts of data-center capacity through NEXTDC. The figure measures potential power-linked data-center capacity and is not a count of GPUs, servers, racks, or currently utilized megawatts.
The company relies on NEXTDC to host the majority of its GPU infrastructure. SharonAI described NEXTDC as its primary data-center provider.
The filing identifies this concentration as a material dependency. SharonAI said construction delays, power-supply issues, financial difficulties, or failure by NEXTDC to deliver contracted capacity on schedule could prevent GPU deployment, impede customer-contract fulfillment, or delay anticipated revenue.
SharonAI further stated that its revenue projections and growth strategy materially depend on NEXTDC delivering capacity on time and as contracted. These statements appear in the S-1’s discussion of business dependencies and risks, rather than as confirmation that all 87 megawatts are already operational.
Alibaba released open weights for Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter model; NVIDIA documented serving it on GB300 NVL72 systems. #
Alibaba released open weights for Qwen3.8-2.4T-A95B, also identified by NVIDIA as Qwen3.8-Max. NVIDIA described it as Alibaba’s largest open-weight model.
The model contains 2.4 trillion parameters. NVIDIA’s account characterized it as bringing near-frontier capabilities to an open-weight release.
NVIDIA documented how to serve the model on GB300 NVL72 systems. The blog’s title also identifies configurable reasoning as part of the serving configuration.
The source distinguishes the model release from the hardware used to serve it: Alibaba released the model weights, while NVIDIA provided the GB300 NVL72 deployment documentation. The supplied description does not specify a commercial customer deployment, service price, or cluster count.
- Alibaba
- Qwen3.8-2.4T-A95B
- NVIDIA
- GB300 NVL72 systems
Researchers from NVIDIA and Duke University published SCALE, a self-supervised method targeting place-and-route design-rule violations at sub-2-nanometer nodes. #
Researchers from NVIDIA and Duke University published a technical paper titled “SCALE: Self-Supervised Constraint-Aware Layout GEneration for Local P&R DRV Fixing at Advanced Nodes.” P&R refers to place-and-route, while DRV means design-rule violation.
The work focuses on local place-and-route design-rule-violation fixing as semiconductor manufacturing advances toward sub-2-nanometer process nodes. The researchers identify complex rule interactions, dense multi-layer routing geometries, and foundry-specific constraints as limitations affecting this task.
SCALE is described as a self-supervised and constraint-aware layout-generation method. Its target is local correction of design-rule violations in advanced-node physical layouts, rather than model training, inference serving, or manufacturing equipment operation.
The publication is a research paper rather than a disclosed commercial electronic-design-automation product or production deployment. The supplied source does not provide commercial availability, customer adoption, or manufacturing-yield results.
- NVIDIA
- Duke University
- SCALE
