🇺🇸NVIDIA signed nonbinding memoranda with six financiers targeting over $500 billion of third-party AI-infrastructure capital. #
NVIDIA signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. The proposed partnerships remain subject to definitive agreements, so the targeted capital is not a closed financing commitment.
The parties plan to establish independent compute-financing platforms and large-scale funding pools intended to provide financing to NVIDIA customers. NVIDIA said the initiative would seek to mobilize more than $500 billion of third-party capital for AI infrastructure over the long term.
NVIDIA CEO Jensen Huang said all of the money would come from third-party capital and that the arrangements would support a broad AI buildout. NVIDIA characterized the platform as a way to make its computing systems and full-stack AI infrastructure an investable asset class for global capital.
The disclosed target encompasses financing for AI infrastructure rather than a direct investment by NVIDIA or an immediately available $500 billion fund. The memoranda therefore establish a framework and fundraising objective, not finalized transaction terms or deployed capital.
🇺🇸Intel announced a proposed $15 billion underwritten public common-stock offering amid AI-compute and advanced-packaging investment. #
Intel Corporation, listed on Nasdaq under INTC, announced a proposed $15 billion underwritten public offering of common stock. The announcement described the transaction as proposed, meaning the source did not establish that the offering had priced or closed.
Intel linked the capital raise to what it described as sustained customer demand driven by investment in AI compute. It also identified physical AI, purpose-built silicon, advanced packaging, and external wafer manufacturing as emerging investment areas.
Subsequent reports citing people familiar with the matter said Intel was considering increasing the offering to approximately $20 billion and expected pricing of about $95 per share or higher. Those reports said demand exceeded $100 billion and that an overallotment option could lift proceeds above $20 billion, but also stressed that the size and pricing remained under discussion and could change.
The reported expansion was not part of Intel’s original announcement and was not confirmed by the company in the supplied sources. The filed stage supported here is therefore a proposed $15 billion underwritten offering, with the larger amount remaining a reported possibility rather than completed financing.
🇺🇸CoreWeave closed a $2.6 billion, roughly five-year delayed-draw term loan for contracted high-performance-computing deployments. #
CoreWeave announced the closing of a $2.6 billion delayed-draw term-loan financing. A delayed-draw structure allows borrowing under the facility according to its agreed draw conditions rather than requiring all proceeds to be funded at announcement.
The financing has a term of approximately five years. The supplied source did not disclose the interest rate, lenders, collateral package, draw schedule, or other borrowing conditions.
CoreWeave said it would use the proceeds to purchase and deploy high-performance-computing infrastructure dedicated to customer contracts. The company described the financing as supporting the continued expansion of its AI cloud platform and committed customer deployments.
🇸🇬SAIHEAT signed a definitive Canopy Wave merger; closing would give Canopy holders 54.19% economic and 78.44% voting interests. #
Nasdaq-listed SAIHEAT Limited, trading under SAIH, entered into a definitive merger agreement with Santa Clara-based AI inference and GPU-cloud company Canopy Wave. The agreement is signed but has not yet closed; completion remains subject to required approvals and other closing conditions.
The all-share consideration comprises an aggregate of 3,306,269 newly issued or issuable SAIHEAT Class A and Class B ordinary shares. The negotiated transaction values used were $60 million pre-money for Canopy Wave and $40 million pre-money for SAIHEAT, which the filing expressly said were not appraisals, valuation opinions, or indications of market value.
At closing, former Canopy Wave shareholders are expected to own approximately 54.19% of the combined company’s economic interests and 78.44% of its voting power. Canopy Wave would become part of the public company, which would be renamed Canopy Wave Holdings Inc. and is expected to trade on Nasdaq as CWAV, subject to approval.
Canopy Wave launched its AI-infrastructure and GPU-as-a-Service offerings in 2024 and reported more than $15 million of aggregate revenue since launch; it introduced Inference-as-a-Service in November 2025. Its platform includes GPU cloud infrastructure, orchestration software, OpenAI-compatible application programming interfaces, intelligent GPU scheduling, SOC 2 Type II certification, and a zero-data-retention policy, while its GPU operations use third-party infrastructure accessed through leases.
🇭🇰TVB proposed a 51%-owned compute-services venture; Gaw Capital may contribute up to HK$2 billion in equity. #
Television Broadcasts Limited’s group, commonly known as TVB, announced a proposal to establish a joint venture with Gaw Capital to provide computing-power services. The proposal represents a diversification into compute services rather than a disclosed operating result.
TVB would hold 51% of the joint venture’s ordinary voting shares, while Gaw Capital would hold 49%. The sources did not disclose the venture’s planned hardware, accelerator type, power capacity, data-center locations, or customer commitments.
The project is intended to proceed in phases over several years. Funding is expected to include up to HK$2 billion of equity from Gaw Capital, together with project-level bank financing and TVB’s internal resources; “up to” describes the maximum contemplated contribution, not equity already funded.
- TVB
- Gaw Capital
- HK$2 billion
Microsoft plans a September unveiling of the Maia 300 AI accelerator and seeks TSMC capacity exceeding 300,000 chips for 2027. #
Microsoft plans to unveil its next-generation Maia 300 internally developed AI accelerator as early as September, according to reporting attributed to The Information. The report described the effort as recovering after a slow start.
Microsoft has been discussing manufacturing capacity with Taiwan Semiconductor Manufacturing Company, or TSMC, for more than 300,000 Maia chips to be delivered during 2027. The sources describe capacity discussions rather than confirmed production or delivery of that quantity.
Microsoft reportedly believes Maia can run both its internally developed models and OpenAI models at lower cost. No price, process node, power rating, memory configuration, performance benchmark, or final TSMC supply commitment was provided in the supplied reports.
The planned September event is an unveiling, not a stated general-availability date. Similarly, the 300,000-plus figure refers to requested manufacturing capacity for 2027 rather than chips already fabricated or installed.
- Microsoft
- Maia 300 AI accelerator
- TSMC
Sony and TSMC plan ¥1 trillion for next-generation imaging-sensor production in Kumamoto, targeting mass production by 2029. #
Sony Group and TSMC plan to invest ¥1 trillion in a Kumamoto production facility, according to reports attributed to Nikkei. The disclosed amount concerns next-generation imaging-sensor production rather than general-purpose logic or AI-accelerator manufacturing.
The companies are targeting mass production of the next-generation imaging sensors by 2029. The supplied sources did not provide an interim construction schedule, production volume, wafer capacity, sensor specifications, or ownership split.
The reports describe a plan rather than completed capital expenditure or active mass production. No financing structure, subsidies, customer commitments, or definitive investment-agreement terms were included in the supplied material.
Google and Delft University of Technology demonstrated hardware enabling system-level failure analysis of 3D integrated-circuit packages with stacked DRAM. #
Researchers from Google and Delft University of Technology published a technical paper titled “Innovative FA Hardware Solution to Enable System-Level Debug of 3D ICs.” FA refers to failure analysis, while 3D ICs are three-dimensional integrated circuits.
The work presents new failure-analysis hardware and a sample-preparation solution for system-level analysis. It targets mobile system-on-chip package-on-package devices in which dynamic random-access memory, or DRAM, is stacked above the system-on-chip.
The system-level scope distinguishes the work from analysis limited to an isolated die or component. The supplied abstract did not provide performance figures, supported package dimensions, commercial availability, or a production deployment schedule.
Pasqal trapped individual atoms using laser light generated by a photonic integrated circuit for neutral-atom quantum computing. #
Pasqal announced that it had trapped individual atoms using laser light generated by a photonic integrated circuit, or PIC. The work was completed with Aeponyx, a Pasqal subsidiary.
The result applies to neutral-atom quantum computing, in which individual atoms form part of the computing hardware. The announcement specifically concerns atom trapping using light produced by an integrated photonic device, not a completed fault-tolerant quantum computer.
Pasqal characterized the result as addressing a central hardware barrier on the path toward fault-tolerant, industrial-scale systems. The supplied source did not disclose qubit counts, trapping fidelity, error rates, operating duration, fabrication process, or a commercialization timetable.
Anthropic, Macquarie Asset Management, and GIC formed Theseus Infrastructure to build U.S.-focused facilities leased under long-term agreements. #
Anthropic, Macquarie Asset Management, and Singapore sovereign wealth fund GIC established Theseus Infrastructure as a data-center development, operation, and leasing platform. Its initial geographic focus is the United States, and the facilities are intended to support Anthropic’s computing-capacity requirements.
The platform will develop purpose-built facilities and lease them to Anthropic under long-term agreements. Macquarie and GIC committed to provide most of the equity funding for each project, while the sources did not disclose total investment, megawatt capacity, site locations, or construction dates.
Anthropic also committed to cover any increases in residential electricity rates attributable to the facilities. The supplied material did not describe the mechanism for measuring or reimbursing such increases.
The announced structure separates facility ownership and financing through Theseus from Anthropic’s occupancy under leases. No specific project was identified as completed or operational in the supplied sources.
- Anthropic
- Macquarie Asset Management
- GIC
- Theseus Infrastructure
The PPL-Blackstone joint venture secured 5 gigawatts of gas turbines for Pennsylvania data centers. #
The joint venture between PPL and Blackstone secured 5 gigawatts of gas-turbine capacity for data-center projects in Pennsylvania. The figure refers to generation equipment capacity rather than data-center IT load, accelerator capacity, or completed electricity production.
PPL President and CEO Vincent Sorgi said he expected bilateral contracting to be the principal pathway for adding generation in the region. His comments came as PJM Interconnection planned a backstop reliability auction for late September.
The source did not disclose turbine manufacturers, individual unit counts, fuel-supply arrangements, project sites, construction dates, customer contracts, or commissioning schedules. Securing turbines therefore should not be conflated with having 5 gigawatts of operational generation or completed data centers.
Siemens’ gas-turbine backlog approached 70 gigawatts after 15 gigawatts of quarterly orders; transformer capacity will expand 50% by 2030. #
Siemens reported that its gas-turbine order backlog was approaching 70 gigawatts. The company booked 15 gigawatts of new gas-turbine orders during the quarter.
Executives said gas-turbine lead times were running at three years or longer. The backlog figure measures ordered turbine capacity and does not represent operating generation capacity already connected to power grids.
Siemens is also expanding manufacturing for grid equipment and plans to increase transformer manufacturing capacity by 50% by 2030. The source did not specify the baseline transformer output, factory-by-factory allocations, or the capital expenditure required for that increase.
The turbine and transformer figures refer to separate infrastructure categories: turbines generate electricity, while transformers change voltage levels for transmission and distribution. Neither figure directly measures data-center capacity.
Hungary’s Paks Nuclear Power Plant began restoring Unit 2 output after Danube water levels improved cooling conditions. #
Hungary’s Paks Nuclear Power Plant began increasing the output of Unit 2 on Aug. 10 after Danube River water levels rose and cooling-water conditions improved. The action was a restoration of reactor output rather than the startup of a newly constructed generating unit.
Hungarian Prime Minister Gábor Magyar said the Danube level was 19 centimeters above the low recorded one week earlier. The plant operator said this improvement allowed the restoration procedure to begin.
An earlier update said Unit 2 was expected to return to full output by 20:00 Greenwich Mean Time. The supplied sources did not provide Unit 2’s electrical capacity, the duration of its reduced-output period, or the exact output level reached by the end of the day.
- Paks Nuclear Power Plant
- Unit 2
Quake AI reported approximately 85% quarterly GPU utilization and signed a multi-year Together AI agreement for NVIDIA HGX B300 capacity. #
RUM Group reported approximately 85% utilization across Quake AI’s existing graphics-processing-unit estate during the quarter. Quake AI became the company’s cloud and AI-infrastructure business following the June 17 closing of Rumble’s acquisition of approximately 85% of Northern Data’s outstanding shares and the renaming of the parent as RUM Group.
The company signed a multi-year agreement under which Together AI will purchase dedicated GPU-cloud capacity powered by NVIDIA HGX B300 systems. The filing identifies the agreement and hardware platform but does not disclose its contract value, GPU count, deployment sites, pricing, or initial service date.
RUM Group reported second-quarter revenue of $40.4 million, up 58% sequentially and 61% year over year; Northern Data contributed $10.1 million from the acquisition date. The Rumble video business generated $30.3 million of quarterly revenue.
Management also identified approximately 250 megawatts of unmonetized capacity targeted for 2027 and described it as a potential annual run-rate opportunity exceeding $3 billion. That amount is a management estimate tied to assumptions referenced in a company presentation, not contracted revenue; the company also said it and Tether had mutually decided not to enter a previously contemplated agreement for up to $75 million of annual GPU services over two years.
- Quake AI
- Together AI
- NVIDIA HGX B300
Zhipu AI (智谱) activated more than 50,000 domestically produced compute chips to support rising inference demand. #
More than 50,000 domestically produced AI-compute chips have reportedly been activated for Zhipu AI (智谱) to address increasing inference demand. The source separately noted market reports from July 21 that Zhipu had built 1 gigawatt of domestic AI-compute infrastructure, but that capacity claim was presented as market information rather than an official confirmation.
Zhipu’s Model-as-a-Service platform reportedly had nearly 7 million registered application programming interface users, an increase of approximately 2 million since early July. The user base included 23,000 enterprise customers.
ZCode, Zhipu’s developer product positioned against Codex, surpassed 1 million users within one month of launch. The source also reported that Zhipu’s annual recurring revenue had increased fifteenfold during 2026.
An investor claimed that annual recurring revenue had reached $2 billion, but Zhipu officially denied that figure. The source subsequently cited a person close to the company as saying the actual number should be higher, but no official replacement figure was supplied; the disputed $2 billion claim therefore should not be treated as verified company revenue.
Google Cloud launched Developer Device Platform in public preview, enabling parallel testing across hundreds of physical and virtual devices. #
Google Cloud launched Developer Device Platform, or DDP, in public preview as a fully managed environment for testing mobile applications. It provides on-demand access to multiple hardware profiles across physical devices and high-concurrency virtual emulators and represents an evolution of Firebase Test Lab for cloud developers.
Its Device Streaming application programming interface lets developers remotely access a selected physical device or emulator for real-time interaction, testing, debugging, and performance monitoring. This allows developers and coding agents to work with device-specific characteristics such as foldable displays and differences between central processing units and graphics processing units.
The Device Run application programming interface supports tests embedded in continuous integration and continuous delivery pipelines. Tests can run in parallel across hundreds of devices, with results used to isolate device-specific failures.
Google described DDP as its first device platform built for agentic development. The company said integrations with Android Studio and the Android command-line interface were forthcoming, rather than already generally available in the public-preview release.
- Google Cloud
- Developer Device Platform
Meta released Muse Glimmer, a 30-billion-parameter open-weight dense model with a context window exceeding 120,000 tokens. #
Meta released Muse Glimmer as a 30-billion-parameter dense model. “Open-weight” means the model’s trained weights are being made available under the release’s applicable terms, while “dense” indicates that its parameters are not described as a mixture-of-experts architecture.
Muse Glimmer supports a context window of more than 120,000 tokens. Context-window size measures the amount of tokenized input and output the model can process in one interaction; it is separate from the model’s 30-billion-parameter count.
NVIDIA described Muse Glimmer as designed for local AI and agentic workflows on NVIDIA platforms. The supplied source did not provide benchmark scores, training-compute figures, licensing details, memory requirements, or a complete list of supported NVIDIA systems.
The U.S. Department of Energy launched an initiative seeking contributors for open-weight scientific foundation models under its Genesis Mission. #
The U.S. Department of Energy launched an initiative to produce open-weight foundation models designed specifically to accelerate scientific discovery. The effort forms part of the department’s broader Genesis Mission.
The department requested input from potential contributors across commercial companies, universities, and research institutions. The announcement is a request for participation and input, not a completed model release or awarded procurement contract.
The models are intended to constitute a new class of scientific foundation models. The supplied announcement did not disclose parameter counts, training datasets, computing allocations, licensing terms, contributor funding, or a release schedule.
OpenAI launched GPT-5.6-Cyber for approved defenders; testing showed responses to 95% of advanced cybersecurity-task requests. #
OpenAI expanded its Daybreak cybersecurity program and introduced GPT-5.6-Cyber, a model designed for advanced, authorized cybersecurity work. Access to its higher-risk capabilities is limited to approved users and includes additional controls and monitoring.
Daybreak has two disclosed access levels. Daybreak Blue provides GPT-5.6 Sol access without system-level cybersecurity safeguards, while Daybreak Red provides GPT-5.6-Cyber for exploit validation and advanced vulnerability research.
In testing, GPT-5.6-Cyber responded to 95% of advanced cybersecurity-task requests. The cited task categories included exploit-chain development, authentication bypass, and privilege escalation; the percentage measures response coverage for requests, not a 95% success rate at completing or validating the tasks.
OpenAI recommended that Codex Daybreak customers switch from “full access mode” to “automatic review mode” through application defaults and interface controls. It also said all individual Daybreak accounts would be required to use hardware security keys beginning Sept. 1, 2026, and separately stated that GPT-5.6-Cyber was not used in an attack on Hugging Face.
