AI & Compute Infrastructure — September 3, 2026

🇺🇸NVIDIA agreed to acquire Hugging Face for $12.93 billion, with closing expected in the first half of 2027. #

NVIDIA announced a definitive agreement to acquire Hugging Face for exactly $12,930,300,000. A separate report placed the expected closing in the first half of 2027; the supplied sources did not disclose regulatory approvals or other closing conditions.

NVIDIA said Hugging Face will remain an open platform supporting models, frameworks, cloud providers, and computing platforms from across the AI ecosystem. NVIDIA also said the combination would be used to scale Hugging Face’s platform, strengthen its infrastructure, and expand AI access for developers and institutions.

The transaction remains pending rather than completed. The announcement confirmed reports that had emerged the previous week, while preserving Hugging Face’s stated role as a cross-ecosystem platform rather than restricting it to NVIDIA technology.

• NVIDIA
• Hugging Face

🇺🇸AI-cloud and data-center provider Crusoe raised more than $3 billion at an approximately $30 billion valuation. #

Crusoe, described in the supplied report as a cloud-computing provider and data-center developer, raised more than $3 billion at a valuation of approximately $30 billion. The report attributed the information to people familiar with the matter rather than to a company announcement or securities filing.

Atreides Management and Valor Equity Partners were identified as co-leads, with Mubadala Capital participating. The source did not disclose the financing instrument, ownership dilution, investor allocation, or whether the stated valuation was pre-money or post-money.

The financing report followed earlier discussions of a roughly $3 billion raise at approximately the same valuation. A separate report connected investor interest in the round with Crusoe’s five-year cloud contract with Jane Street, but did not state that the financing was contractually contingent on that customer agreement.

• Crusoe

🇺🇸Anthropic neared an expansion of its revolving credit facility to $15 billion before a potential initial public offering. #

Anthropic was reported to be close to expanding its revolving credit facility to $15 billion. The arrangement had not yet been reported as finalized, and the facility size would exceed the company’s earlier target of approximately $10 billion.

Morgan Stanley was expected to lead the credit arrangement, with Goldman Sachs, JPMorgan, and Citigroup taking significant roles. Barclays and Wells Fargo were also expected to participate prominently, although the sources cautioned that details could still change.

The facility was described as a step preceding a possible public filing for an initial public offering. The reports said the four principal banks were also expected to be major IPO underwriters, but no IPO filing, offering size, timetable, interest rate, collateral package, or borrowing conditions were disclosed.

• Anthropic

🇺🇸Thinking Machines Lab entered talks to raise approximately $1 billion at an estimated $40 billion valuation. #

Thinking Machines Lab was reported to be discussing a financing of approximately $1 billion at a valuation of around $40 billion. The source characterized the transaction as talks, not a completed or contractually committed financing.

The supplied reports attributed the information to The Information. They did not identify prospective investors, a lead investor, the financing instrument, or whether the approximately $40 billion figure represented a pre-money or post-money valuation.

No information was provided on dilution, liquidation preferences, governance rights, use of proceeds, or closing conditions. Accordingly, the reported figures describe the contemplated fundraising terms rather than capital already received by the company.

• Thinking Machines Lab

🇺🇸Flex agreed to acquire data-center power-conversion supplier EPC Power for $4.4 billion, ahead of a planned infrastructure-business spinoff. #

Flex agreed to acquire EPC Power in a transaction valued at $4.4 billion. EPC Power was described as a supplier of power-conversion solutions for data centers and electrical grids, placing the acquired business in the power-delivery layer rather than the processor or server layer.

The source said the transaction would give Flex access to platforms needed to deploy NVIDIA’s highest-power computing chips. It did not provide the acquisition’s cash-and-stock mix, financing structure, regulatory conditions, expected closing date, or any shareholder-approval requirement.

EPC Power is expected to become part of Flex’s cloud and power infrastructure business. Flex plans to separate that business into an independently listed company in the first quarter of 2027, but the source described this as a planned spinoff rather than a completed separation.

• Flex
• EPC Power

NVIDIA RTX Spark Windows PCs will launch in October 2026, supporting local inference and agent workloads. #

NVIDIA announced that RTX Spark Windows personal computers will become available in October 2026. The systems are compact Windows PCs intended for AI enthusiasts, developers, and creators running workloads locally.

At IFA 2026, NVIDIA, Microsoft, and hardware partners presented the systems as part of a broader effort to accelerate local inference and simplify the setup and operation of AI agents on NVIDIA hardware. The source did not provide pricing, memory capacity, processor configuration, power consumption, or exact launch-day availability.

NVIDIA also introduced the PAIR virtual inference router for making additional compute on a local network available to AI workloads. Its stated use case includes multi-agent workflows in which a lead agent divides a task among specialized subagents.

• October 2026

Xiaomi and CXMT jointly tested LPDDR6 memory for one year; the Xiaomi 18 Fold will jointly debut it. #

Xiaomi President and partner Lu Weibing said Xiaomi and ChangXin Memory Technologies, or CXMT (长鑫存储), conducted joint LPDDR6 testing and tuning for one year. LPDDR6 is low-power double-data-rate memory intended for power-sensitive devices rather than a GPU or standalone compute accelerator.

Lu said the companies had begun cooperating earlier and established a joint laboratory in 2022. The two companies assembled a core team of more than 50 people and went through an extended period of technical integration before the memory’s formal implementation in 2026.

The Xiaomi 18 Fold will serve as an important joint-launch product and validation platform for the work, according to Lu. The source did not provide memory bandwidth, capacity, operating voltage, supplier allocation, production volume, or a handset launch date.

• Xiaomi
• CXMT
• LPDDR6 memory
• Xiaomi 18 Fold

MikroTik’s CRS804-4DDQ-hRM provides four 400-gigabit Ethernet ports for desktop and local AI clusters. #

The MikroTik CRS804-4DDQ-hRM is a four-port 400-gigabit Ethernet switch. ServeTheHome evaluated it as a desktop-class networking device that had been used for local AI clusters.

The product addresses network connectivity between systems rather than supplying compute itself: it is a switch, not a GPU, accelerator, server, rack, or cluster. Its four 400GbE ports provide the physical high-speed links described in the review.

The supplied source characterized the device as a comparatively inexpensive desktop 400GbE switch. It did not provide a price, aggregate switching capacity, power draw, port breakout configuration, or measured local-cluster benchmark in the supplied description.

• MikroTik’s CRS804-4DDQ-hRM

Jülich and eleQtron inaugurated the JION trapped-ion quantum computer, enabling hybrid workloads with Jülich supercomputers. #

Forschungszentrum Jülich and Siegen-based startup eleQtron officially inaugurated JION, a trapped-ion quantum computer developed in North Rhine-Westphalia, Germany. The source identifies JION as a quantum-computing system, distinct from Jülich’s conventional supercomputers.

JION will be made available to research institutions and industry through the Jülich Unified Infrastructure for Quantum Computing, or JUNIQ, user infrastructure. Its integration is intended to support hybrid computations that combine the quantum system with Jülich’s supercomputing resources.

North Rhine-Westphalia Minister-President Hendrik Wüst and Science Minister Ina Brandes participated in the inauguration. The supplied source did not state JION’s qubit count, gate fidelity, quantum volume, power requirements, or user-access schedule.

• Jülich

George Mason University and TreQ partnered to deploy Virginia’s first open-architecture quantum-computing system. #

George Mason University and Oxford-based frontier-computing company TreQ announced a partnership to bring open-architecture quantum infrastructure to Virginia. The first deployment is described as Virginia’s and the National Capital Region’s first open-architecture quantum-computing system.

The announcement concerns a partnership and planned deployment rather than a system already reported as operational. “Open architecture” is the defining system characteristic supplied by the source, which did not identify a qubit modality, processor supplier, qubit count, performance benchmark, or installation date.

The project is framed as quantum-computing infrastructure for the Commonwealth of Virginia. No project cost, financing arrangement, power requirement, cloud-access model, or operational-capacity figure was disclosed in the supplied material.

• George Mason University
• TreQ

Four U.S. companies committed $2 billion toward South Korean semiconductor materials, equipment, manufacturing, and offshore-wind projects. #

South Korea’s Ministry of Trade, Industry and Energy said four U.S. companies committed a combined $2 billion of investment in South Korea. The stated target areas are semiconductor materials and equipment, advanced display materials, manufacturing capacity, and offshore wind.

Air Products plans to expand semiconductor-gas supply facilities in Pyeongtaek, while ion-implantation equipment manufacturer Axcelis plans to enlarge its production facilities in South Korea. Corning committed to expanding its manufacturing capacity and broader Korean operations.

Pacifico Energy plans to begin a 3.2-gigawatt offshore-wind project in the Gwangju region. The supplied statement did not allocate the $2 billion total among the four companies or provide individual construction schedules, financing conditions, or completion dates.

China completed its first bidirectional transmission-rights transactions between State Grid and China Southern Power Grid. #

State Grid reported that the first transmission-rights transaction sending electricity from China Southern Power Grid to State Grid was completed after southern clean electricity was delivered in full to Wuxi, Jiangsu, through the Yunxiao direct-current transmission link at the end of August. The transaction concerns transmission rights across grid operators, not the construction of new generation capacity.

The reverse direction had already been established in June, when the first State Grid-to-China Southern Power Grid transaction was completed. Completion of the later south-to-north transaction created the first bidirectional connection for this type of market-based transmission-rights trading in China.

The source did not disclose the electricity volume, transmission-rights price, contract term, or power-plant source mix beyond describing the delivered supply as southern clean electricity. It also did not specify the voltage or rated power of the Yunxiao direct-current link.

• State Grid
• China Southern Power Grid

Lead Intelligent Equipment shipped China’s first gigawatt-scale automated proton-exchange-membrane electrolyzer line using wet-membrane assembly. #

Lead Intelligent Equipment’s wholly owned subsidiary Hydrogen Intelligence (氢导智能) shipped a gigawatt-scale automated assembly line for proton-exchange-membrane, or PEM, electrolyzers. The line was built for a major domestic energy-equipment company whose name was not disclosed.

The equipment had reached the customer’s site and was entering commissioning at the time of the report. The source described it as China’s first automated PEM-electrolyzer production line using a wet-membrane assembly process.

The gigawatt figure refers to the production-line scale associated with electrolyzer manufacturing, not an operating power plant or an installed hydrogen-production project. The source did not provide annual unit output, individual electrolyzer capacity, project cost, commissioning duration, or commercial-production date.

• Lead Intelligent Equipment

Weekly average electricity demand across ERCOT remained near record levels as sustained high temperatures elevated consumption. #

The U.S. Energy Information Administration reported that weekly average electricity load in the Electric Reliability Council of Texas, or ERCOT, remained near a record high. ERCOT is the regional transmission organization serving most of Texas.

Sustained high temperatures were identified as the reason electricity demand remained persistently elevated. The item refers to average load over a week rather than a single instantaneous or daily peak-demand record.

The supplied description did not provide the weekly average in megawatts, the prior record value, reserve margins, generation outages, or wholesale-price data. It therefore establishes the persistence and weather-related cause of high demand without quantifying grid scarcity.

• ERCOT

University of Waterloo researchers redesigned porous flow-battery electrodes to improve electrolyte movement and energy-storage reactions. #

University of Waterloo researchers redesigned the porous electrode, a key component of a redox flow battery, using structures inspired by nature. The new electrode is intended to let electrolyte move through the battery more smoothly and efficiently.

Improved electrolyte movement promotes more complete reactions for storing and releasing energy. The report connected the design with storing large quantities of electricity produced by wind and solar generation.

The research was published in the Journal of Energy Storage. The supplied source did not state a percentage improvement in efficiency, energy density, cycle life, cost, or electrolyte flow rate, and it did not describe the work as a commercial deployment.

Crusoe signed a five-year, approximately $13 billion AI-cloud agreement supplying advanced-chip clusters to Jane Street. #

Crusoe signed an approximately $13 billion cloud-computing agreement with quantitative-trading firm Jane Street Group, according to people familiar with the nonpublic arrangement. The contract term is five years.

Under the agreement, Crusoe will use its cloud platform to provide Jane Street with compute clusters containing advanced AI chips, along with other infrastructure required for AI training and inference. The sources did not identify the processor model, accelerator count, cluster locations, power capacity, deployment schedule, or annual revenue recognition.

Jane Street was described as the largest customer Crusoe’s cloud business had signed to date. The report also said Crusoe had previously contracted with Meta Platforms and Oracle to provide AI compute.

Bloomberg had separately reported that Crusoe was seeking a chip loan secured by the Jane Street contract. The source did not say that this loan had closed, disclose its maximum size or collateral terms, or establish that the customer agreement was dependent on that financing.

• Crusoe
• Jane Street

Figure and Nscale signed a $3.5 billion compute deal targeting up to 100,000 NVIDIA Vera Rubin GPUs from the second half of 2027. #

Figure and Nscale announced a compute agreement valued at $3.5 billion, with an intention to expand the relationship beyond $6 billion. The initial deployment is targeted to begin in the second half of 2027 in Barstow, Texas.

The parties said the deployment would use NVIDIA Vera Rubin graphics processing units, or GPUs, and could reach up to 100,000 GPUs. “Up to” identifies a maximum target rather than hardware already installed or committed as operational capacity.

The source did not disclose the initial GPU quantity, data-center power capacity, delivery milestones, contract term, financing structure, or conditions governing expansion beyond $6 billion. No operating cluster or revenue contribution was reported as of the announcement date.

• Figure
• Nscale

Fuzzball 4.2 added AMD GPU support and a permission-controlled Model Context Protocol server for agent-operated cluster workflows. #

CIQ released Fuzzball 4.2, an update to its turnkey sovereign artificial-intelligence and high-performance-computing orchestration platform. The release adds support for AMD graphics processing units, extending hardware coverage beyond the platform’s prior configurations.

Fuzzball 4.2 also introduces a Model Context Protocol, or MCP, server. The server permits AI agents to operate Fuzzball only under explicitly granted permissions, maintaining an authorization boundary around agent actions.

Workflows running inside the cluster can now direct additional work themselves. The source did not specify supported AMD GPU models, performance comparisons, licensing terms, cluster-size limits, or the precise permission roles exposed through the MCP server.

• Fuzzball 4.2
• AMD GPU
• Model Context Protocol server

ChatGPT, Claude, and Grok recovered after simultaneous service disruptions; Grok identified its Memphis compute center as the cause. #

ChatGPT, Anthropic’s Claude, and SpaceXAI’s Grok experienced service disruptions at approximately the same time on September 3 and subsequently reported recovery through their status pages. Downdetector also reported problems involving Cursor, although the digest item concerns the three major model services.

Users began reporting Grok and Claude problems around 9:00 a.m. U.S. Eastern time, while OpenAI issues were reported at around 10:30 a.m. Eastern time. Claude displayed an unexpected-capacity-limit message, and Grok reported that its model or some functions were temporarily unavailable.

During recovery, Anthropic said multiple models had returned to baseline while Opus 4.8 and Opus 5 remained affected. A later report stated that all three services had resolved their incidents.

SpaceXAI specifically attributed Grok’s problems to an outage at its Memphis compute center and said all systems had been fully restored. The supplied sources did not establish a common cause connecting the three companies’ disruptions.

• ChatGPT
• Claude
• Grok

OpenAI launched GPT-6 Astra for selected organizations, with broader ChatGPT, API, and AWS availability planned within days; it costs $10 per million input tokens and $50 per million output tokens through the OpenAI API. #

OpenAI began a staged release of GPT-6 Astra to a limited set of organizations, including organizations participating in its Daybreak Access cybersecurity program. OpenAI described Astra as its new flagship model and added cybersecurity safeguards and tiered access to its release.

OpenAI said the model would become available within days to ChatGPT Plus, Pro, Business, and Enterprise users. It also planned access through the OpenAI application programming interface and Amazon Web Services, but the sources did not give an exact date for each distribution channel.

Standard OpenAI API pricing is $10 per million input tokens and $50 per million output tokens. The report said this was 2.5 times the price of the previous flagship model, GPT-5.6 Sol, and noted that OpenAI had discussed a possible shift toward task-based rather than token-based pricing.

OpenAI highlighted improvements in computer operation, software programming, and professional tasks, and reported top performance on FrontierMath Tier 1, ARC-AGI 3, and ExploitBench. President Greg Brockman said the model was particularly capable of operating computers on behalf of users, while OpenAI also updated Codex to accelerate computer-use capabilities.

• OpenAI
• GPT-6 Astra
• ChatGPT
• API
• AWS
• OpenAI API

OpenAI committed $1 billion for subsidized Daybreak access, training, technical support, and cybersecurity partnerships. #

OpenAI committed $1 billion to subsidized access to Daybreak, training, technical support, and partnerships. Daybreak was presented as a tool or access program for frontline defenders, with the wider GPT-6 Astra launch identifying participating organizations as part of a cybersecurity initiative.

The commitment is a program-level funding announcement rather than a disclosed equity investment, acquisition, or credit facility. The supplied sources did not provide a spending schedule, recipient list, grant sizes, geographic scope, or allocation among access subsidies, training, support, and partnerships.

Selected Daybreak Access organizations were among the first recipients of GPT-6 Astra. OpenAI did not disclose in the supplied material which organizations received initial access or what technical and contractual restrictions applied to the program.

• OpenAI
• Daybreak