🇺🇸NVIDIA revised its proposed guarantee for OpenAI’s Ohio campus from $250 billion to below $120 billion; the proposal initially covers roughly 5 gigawatts, while OpenAI continues unsigned negotiations for a binding 10-gigawatt lease. #
People familiar with the discussions said NVIDIA and OpenAI were nearing a financing agreement for a large Ohio data-center campus. The reported transaction had not been signed, although an agreement could reportedly be reached as early as the weekend.
Under the revised structure, NVIDIA would initially guarantee only half of the multibillion-dollar expansion rather than the entire project. The reported change would reduce NVIDIA’s financial guarantee from $250 billion to less than $120 billion and was described as a response to investor concerns about the chipmaker’s exposure.
The current framework assigns the guarantee to a first phase totaling approximately 5 gigawatts of power capacity, with financing for subsequent phases to be determined later. Separately, OpenAI was still negotiating a binding lease covering 10 gigawatts and had not yet signed it.
- NVIDIA
- OpenAI’s Ohio campus
- OpenAI
🇺🇸SpaceX completed its $60 billion acquisition of Cursor, issuing approximately 389 million Class A shares as consideration. #
SpaceX disclosed that its merger with AI coding-tool company Cursor became effective on August 14, two months after the acquisition agreement was announced. Cursor’s implied equity value in the completed transaction was $60 billion.
Cursor shareholders received approximately 389 million shares of SpaceX Class A common stock. The principal consideration therefore consisted of SpaceX shares, and the source said the transaction did not involve a public offering.
Cursor’s vested restricted stock units were converted into approximately 1.75 million SpaceX Class A shares. Its unvested restricted stock units and stock options were converted into approximately 29.13 million SpaceX restricted stock units and 44.37 million SpaceX stock options, respectively.
SpaceXAI said Cursor would join its team and support work on AI tools beginning with software engineering and expanding into knowledge work. Cursor’s AI assistant was launched in 2023 to help programmers write and debug code.
🇺🇸Micron Ventures launched the $250 million Paradigm Fund, its third and largest vehicle, investing across the AI stack. #
Micron Technology launched the Micron Ventures Paradigm Fund with $250 million in investment capacity. The company described it as Micron Ventures’ third fund and its largest vehicle to date.
The fund is intended to partner with companies developing AI technologies. Its investment remit spans the full AI technology stack, including model architectures and other layers of the AI ecosystem.
The announcement establishes the size and mandate of the investment vehicle but does not identify individual portfolio investments in the supplied source. It also does not disclose investment pacing, ownership targets, or fund-deployment deadlines.
🇺🇸Axe Compute (Nasdaq: AGPU), which retains a legacy oncology unit, reported $3.2 million in compute revenue and more than $3 billion in signed contracts. #
Axe Compute reported $3.2 million of second-quarter revenue, representing its first full quarter of compute-service revenue and an increase of more than 90 times sequentially. All of that revenue came from Axe Compute Access, which rents already-online third-party graphics processing unit capacity; revenue from its dedicated Axe Compute Build contracts begins only when deployments go live.
The company said it signed more than $2.8 billion of new contracts, bringing its 2026 signed contracted value above $3 billion. Management’s stated expected annualized revenue run rate of more than $696 million upon full deployment is forward-looking rather than recognized revenue, as are the chief executive’s target of signing another $2 billion of contracts before year-end and associated deployment expectations.
Axe Compute recorded a $17.2 million net loss, including a $13.1 million non-cash loss on digital assets, and approximately negative $4.9 million of adjusted earnings before interest, taxes, depreciation and amortization. It ended the quarter with $21.9 million in cash, up from $6.9 million on March 31, and reported $60.8 million of customer prepayments.
The Nasdaq-listed company sources GPU capacity and coordinates hardware, colocation, networking, storage, power infrastructure and financing for customers. It continues to evaluate strategic alternatives for Helomics, its legacy oncology drug-discovery operation and proprietary tumor-sample biobank, which is outside its core compute-infrastructure business.
- Axe Compute (Nasdaq: AGPU)
SK hynix spent ₩18.33 trillion on equipment during the first half of 2026, up 72.7% year over year. #
SK hynix reported consolidated cash spending of ₩18.3288 trillion on purchases of tangible assets during the first half of 2026. That compared with ₩10.6157 trillion in the corresponding 2025 period, an increase of 72.7%.
Research and development spending also increased to ₩6.0428 trillion from ₩3.0456 trillion, representing 98.4% year-over-year growth. Of the first-half total, ₩5.8163 trillion was allocated to continuing development costs.
The company was reported to be expanding production facilities as demand increased for high-bandwidth memory used in AI systems, server dynamic random-access memory and enterprise solid-state drives. The disclosed ₩18.3288 trillion represents equipment-related tangible-asset spending, not revenue or data-center capacity.
Lightmatter launched an Open Compute Project co-packaged optics workstream involving 19 companies. #
Lightmatter launched the Open Silicon Photonics for AI Systems initiative as an official workstream within the Open Compute Project. The initiative concerns co-packaged optics architecture for AI systems.
The coalition comprises nine new participants and 10 founding members, for 19 companies in total. Companies identified in the announcement include Celestica, Dell Technologies, Flex, Foxconn Interconnect Technology and GlobalFoundries.
The workstream is an industry architecture initiative rather than a disclosed commercial system deployment. The supplied source does not state a production schedule, capital commitment, optical bandwidth target or customer order value.
Supermicro displayed the ASG-4116S-NU160R, a 4U server accommodating 160 NVMe SSD bays. #
Supermicro displayed the ASG-4116S-NU160R on the FMS 2026 show floor. It is a 4-rack-unit storage server with 160 bays for Non-Volatile Memory Express solid-state drives.
The 160-bay count describes storage-device density within one server chassis. It does not represent 160 GPUs, accelerators, compute nodes or racks.
The supplied report identifies the exhibited server and its physical storage density but does not provide aggregate capacity in terabytes or petabytes. It also does not disclose pricing, customer orders, availability timing or measured throughput.
- Supermicro
- ASG-4116S-NU160R
Li Auto is exploring an early-stage custom cloud-inference chip using the dataflow architecture of its autonomous-driving chip. #
Li Auto is exploring the development of a proprietary cloud-inference chip, according to multiple sources cited by LatePost (晚点). The project remains at an early stage and is expected to use the same dataflow architecture as the company’s autonomous-driving chip.
The proposed processor is a cloud inference accelerator, distinct from the in-vehicle autonomous-driving chip whose architecture it would reuse. Cloud inference processors can handle selected workloads otherwise run on GPUs, including autonomous-driving model data processing, testing and simulation, as well as requests to large language models.
The report distinguishes inference from model training that involves parameter updates. Those parameter-update workloads would still require training processors, rather than being assigned to the proposed inference chip.
Shengyi Technology reported RMB 19.03 billion in first-half revenue, with AI servers and data-center switches forming its principal growth line. #
Shengyi Technology (生益科技) reported first-half 2026 revenue of RMB 19.026 billion, an increase of 50.05% year over year. Net income attributable to listed-company shareholders was RMB 3.287 billion, up 130.42%, while total profit rose 129.33% to RMB 4.221 billion.
The company identified AI servers and data-center switches as the principal growth line during the period. It cited strong demand for high-layer-count boards, advanced high-density interconnect products and packaging substrates, alongside higher upstream raw-material prices, as factors increasing industry output value and profitability.
At the end of the reporting period, total assets were RMB 42.398 billion, up 29.38% from year-end, and equity attributable to listed-company shareholders was RMB 18.502 billion, up 10.63%. The company also said it was deepening cooperation with major domestic and international AI-server, data-center and compute-chip customers.
Digi Power X targets 15 megawatts of IT load in December and another 25 megawatts in March, conditional on financing. #
Digi Power X said construction of its Tier III AI data center in Columbiana, Alabama, remained on schedule. Phase 1 is expected to deliver 15 megawatts of information-technology load in December 2026, followed by an additional 25 megawatts in Phase 2 in March 2027, for as much as 40 megawatts of IT load.
The additional 25 megawatts in Phase 2 is expressly conditional on Digi Power X securing adequate financing. The company said it was in advanced discussions with lenders and had engaged Goldman Sachs to assist in syndicating debt financing, but the supplied disclosure did not state that financing had closed.
The campus is covered by a colocation and master services agreement with Cerebras Systems. The initial 10-year term carries an approximately $1.1 billion contract value, while the approximately $2.5 billion potential value assumes the exercise of one seven-year extension and remains subject to the contractual conditions described by the company.
The dedicated on-site substation for Phase 1 was complete, grid interconnection had been finalized and a utility power-delivery agreement was in place. These infrastructure milestones do not mean the targeted 15-megawatt IT deployment had commenced commercial operation as of the filing.
Federal Energy Regulatory Commission (FERC) approved PowerTransitions’ 1.2-gigawatt New York plant acquisition and TransAlta’s Colorado purchases totaling 319 megawatts. #
The Federal Energy Regulatory Commission approved PowerTransitions’ proposed acquisition of a 1.2-gigawatt gas-fired power plant in New York. PowerTransitions intends to use the acquired facility for an “energy campus.”
FERC also approved TransAlta’s acquisition of two gas-fired power plants in Colorado. The two Colorado facilities have combined generating capacity of approximately 319 megawatts.
The approvals concern acquisitions of power-generation assets, not equivalent amounts of installed GPU capacity, data-center IT load or contracted compute. The supplied report does not provide purchase prices or state that either buyer had converted the plants into data-center infrastructure.
- TransAlta’s Colorado purchases
TerraVolt (GEDC) contracted 55,000 MMBtu per day for its planned Idaho campus and delivered a $6 million letter of credit. #
TerraVolt Holdings entered into a natural-gas supply agreement in April 2026 for 55,000 million British thermal units per day. The fuel is intended for a planned on-site-powered data-center campus in Southeast Idaho connected to the Northwest Natural Gas Pipeline.
The company paid the fuel supplier a $3,832,500 natural-gas reservation fee in May. In August, it delivered a letter of credit with a maximum drawable amount of $6 million to secure its obligations under the supply agreement.
TerraVolt describes its planned offering as a Physical Infrastructure-as-a-Service platform combining behind-the-meter power, construction-ready data-center sites, utilities and fiber connectivity. The targeted customers include hyperscalers, neocloud providers and colocation operators, but the filing does not state that the Idaho campus was operational or serving tenants.
The company said it was negotiating with multiple landowners. Its planned generation approach uses on-site natural-gas turbines and reciprocating engines in a location where that form of production is permitted by local and state building codes.
Quinbrook Supernode’s first two phases reached commercial operation at 520 megawatts/1,858 megawatt-hours, while third-phase financing closed at A$469 million. #
Quinbrook’s Supernode battery-storage project placed its second phase into commercial operation, bringing Phases 1 and 2 to fully operational status. Their combined power and energy ratings are 520 megawatts and 1,858 megawatt-hours, respectively.
Financing for Phase 3 closed at A$469 million. Once that phase is completed, the project’s total storage capacity is expected to exceed 3 gigawatt-hours.
Contemporary Amperex Technology Co. Limited, or CATL (宁德时代), is the project’s sole battery-energy-storage-system supplier. Phase 3 will exclusively use CATL’s TENER (天恒) energy-storage system, and the companies also plan to develop an eight-hour-duration storage solution.
The 520-megawatt figure is the storage project’s power rating, while 1,858 megawatt-hours is its stored-energy capacity. Neither figure represents data-center IT load or GPU-cluster capacity.
Fermi signed its first definitive Project Matador tenant lease on August 9; initial private-grid AI campus operations target 2027. #
Fermi entered into a definitive lease with the first tenant for Project Matador on August 9, 2026. The company said vertical construction was expected to begin following execution of that lease.
Commercial operations for the first private-grid-powered AI data-center campus are targeted to commence in 2027. This is a company target rather than an operating result or a statement that the campus was already delivering capacity.
As of June 30, Fermi had not begun revenue-generating activities. Its work through that date consisted of formation, commercial engagement, infrastructure procurement, leasing, preliminary site development and marketing for Project Matador.
Fermi expects substantially all future revenue to come from leases with hyperscalers, neocloud providers, semiconductor companies and other compute-intensive tenants. Its disclosed energy-generation and facility-square-footage figures are unaudited and were outside the review scope of its independent registered public accounting firm.
Digi Power X generated $1.1 million from five weeks of bare-metal GPU rentals using 0.6 megawatts of deployed capacity. #
Digi Power X reported approximately $1.1 million of GPU bare-metal rental revenue from roughly five weeks of second-quarter operations. This was the company’s first AI-compute revenue and formed part of total quarterly revenue of approximately $6.6 million.
The company invested approximately $30 million in GPU infrastructure representing about 0.6 megawatts of deployed AI-compute capacity. Its B300 bare-metal infrastructure had operated at 100% uptime since May 2026, according to management’s filed press release.
Digi Power X’s 24-month bare-metal rental agreement with SubQ AI became effective on May 15 and has an expected total contract value of approximately $19.6 million. The service provides exclusive access to NVIDIA Blackwell GPUs through Digi Power X’s NeoCloudz GPU-as-a-Service platform at its AI data center.
The filing says the deployment follows NVIDIA reference architecture and operates in a facility engineered to Rated 3 standards, with redundant utility feeds, N+1 uninterruptible power supplies and a cooling distribution unit-and-chiller-based two-loop cooling design. Digi Power X’s stated plan to add approximately 10 megawatts during 2027 is a future expansion plan, not currently deployed capacity.
Google Cloud extended Compute Flexible Committed Use Discounts to G2 NVIDIA L4 and G4 RTX Pro 6000 virtual machines. #
Google Cloud made Compute Flexible Committed Use Discounts available for G2 virtual machines using NVIDIA L4 GPUs and G4 virtual machines using NVIDIA RTX Pro 6000 GPUs. The program allows customers to receive committed-spend discounts on these GPU-backed virtual-machine families.
The flexible commitment can cover G2 and G4 GPU virtual machines alongside general-purpose compute, Google Kubernetes Engine, Cloud Run and other eligible services under one spend commitment. Google said customers can also adapt across virtual-machine families and migrate between regions.
This is a cloud-pricing and commitment-program expansion, not an announcement of new physical GPU inventory or data-center megawatts. The supplied announcement does not specify a discount percentage or minimum commitment value.
Alibaba Cloud launched Lingjun Zhenwu M890 Supernode (灵骏真武 M890 超节点) instances in Wulanchabu, supporting models exceeding 2 trillion parameters. #
Alibaba Cloud launched Lingjun Zhenwu M890 Supernode (灵骏真武 M890 超节点) instances, with the first offering made available for sale in Wulanchabu. The source describes it as China’s first supernode-form compute platform to run models containing more than 2 trillion parameters.
Kimi K3 and Qwen3.8-Max had already been made available externally through the M890 instance. The disclosure identifies model-scale support but does not provide an accelerator count, accelerator model, rack count, electrical load or networking-bandwidth figure.
A supernode is a tightly integrated compute-system form rather than a single GPU or conventional standalone server. The source’s “more than 2 trillion parameters” refers to supported model size, not the number of accelerators or tokens processed.
China’s National Supercomputing Internet (国家超算互联网) added DeepSeek-V4-Pro and DeepSeek Harness to its 100,000-card-class heterogeneous compute pool. #
China’s National Supercomputing Internet (国家超算互联网) added the production release DeepSeek-V4-Pro-0813 and the DeepSeek Harness agent framework. The platform describes itself as the country’s first 100,000-card-class pool integrating supercomputing and intelligent-computing resources.
The pool provides computing resources across the model lifecycle, including training, fine-tuning, evaluation and deployment. Users can download DeepSeek models and source code, conduct private deployments, run distributed inference and undertake secondary agent development through the platform.
The “100,000-card-class” description refers to a heterogeneous domestic compute-resource pool. The source does not identify all devices as GPUs or provide a breakdown by GPU, neural processing unit, accelerator generation, server or cluster.
- National Supercomputing Internet (国家超算互联网)
- DeepSeek-V4-Pro
- DeepSeek Harness
Alibaba open-sourced Qwen3.8-27B, a native multimodal dense model with 27 billion parameters, under the Apache 2.0 license. #
Alibaba released the Qwen3.8 model series for developers, research institutions and companies to download, deploy and use. Qwen3.8-27B is a native multimodal dense model containing 27 billion parameters.
Alibaba said the model’s overall capabilities exceed those of Qwen3.7-Plus and highlighted performance in coding and office-productivity scenarios. These are company-reported comparisons; the supplied sources do not include independent benchmark tables.
Qwen3.8-27B was released under the Apache 2.0 license. Alibaba also said a quantized version can run smoothly on a consumer-grade graphics card and that worldwide downloads of Qwen models had exceeded 3 billion.
The 27-billion figure describes model parameters, not training tokens, GPU count or serving capacity. “Dense” distinguishes the architecture from a mixture-of-experts model in which only selected parameter groups are activated for each input.
The California Public Utilities Commission approved Waymo expansions in San Francisco and Los Angeles, plus launches in Sacramento and San Diego. #
Waymo said the California Public Utilities Commission approved an expansion of its autonomous ride-hailing service in the San Francisco Bay Area. The regulator also approved an expansion in Los Angeles.
The commission separately approved Waymo service launches in Sacramento and San Diego. Waymo is the autonomous-driving subsidiary of Alphabet, rather than a separate listed company.
The approvals concern geographic authorization for autonomous ride-hailing service. The supplied messages do not provide launch dates, vehicle counts, fleet-compute specifications or operating-hour restrictions for the newly approved areas.
- California Public Utilities Commission
- Waymo
SpaceXAI made its Grok 4.6 model available through GitHub Copilot for software-development workflows. #
SpaceXAI launched Grok 4.6 within GitHub Copilot. The integration makes the model available through a software-development assistant used for coding workflows.
The announcement identifies model availability through GitHub Copilot rather than the launch of a new compute cluster or a standalone hardware product. No GPU type, inference throughput, context-window size or pricing was included in the supplied source.
The source also does not specify whether access is universal across all GitHub Copilot plans or limited by region, subscription tier or preview status. Accordingly, the verified update is the addition of Grok 4.6 to the Copilot environment.
- SpaceXAI
- Grok 4.6
- GitHub Copilot
