AI & Compute Infrastructure — August 27, 2026

🇺🇸NVIDIA reportedly agreed to acquire Hugging Face for $12.9 billion; the transaction has not closed. #

HPCwire, citing a report from The Information, said NVIDIA agreed to acquire Hugging Face for $12.9 billion. The reporting describes an agreement rather than a completed acquisition, and the available source does not state that the transaction has closed.

Hugging Face operates an AI platform spanning models, datasets and collaboration tools. HPCwire reported that the platform has more than 13 million registered users and described it as the “GitHub for AI.”

Hugging Face was founded 10 years ago, according to HPCwire. The article said the acquisition, if completed, would expand NVIDIA’s AI-model business, but it did not provide closing conditions, regulatory approvals, financing terms or a projected completion date.

• NVIDIA
• Hugging Face

🇺🇸Anthropic plans post-September 7 prospectus disclosure while considering shareholder sales and a longer initial-public-offering lockup. #

The Information, as relayed by the source messages, reported that Anthropic plans to disclose its initial-public-offering prospectus after September 7. The reported timetable calls for an investor day for prospective investors and research analysts in mid-September.

Anthropic is considering allowing existing shareholders to sell some shares in the offering. Such a structure could provide liquidity to early investors, executives and employees, but the source describes the shareholder-sale provision as under consideration rather than finalized.

The company is also considering a longer-than-usual post-IPO lockup period to manage trading volatility after listing. The reported IPO window was late September or early October, although the source explicitly said the timing could be delayed.

• Anthropic
• September 7

🇺🇸NVIDIA paused portions of its AI-cloud credit-support and revenue-sharing program, which covered $36 billion in commitments. #

NVIDIA reportedly paused parts of its AI Compute Partner Program less than two months after announcing it in July. The company had disclosed that agreements associated with the program covered $36 billion of commitments and generally ran for six years.

Under the reported structure, NVIDIA offered credit support to AI-cloud providers by committing to rent their graphics processing unit capacity if they could not find other customers. That revenue backstop was intended to help the providers obtain financing for GPU infrastructure.

The proposed economics would have allowed NVIDIA to receive 50% of revenue above a base threshold covering chip depreciation, data-center expenses, personnel and other costs. Some NVIDIA employees reportedly raised antitrust concerns, but the exact reason for the partial pause was not disclosed.

The sources said NVIDIA could modify the program or combine it with other initiatives. An NVIDIA spokesperson separately said the commercial model introduced in July to broaden access to computing capacity remained in operation, indicating that the reported action concerned portions of the program rather than a complete termination.

• NVIDIA

🇺🇸Anthropic abandoned a roughly $7 billion acquisition discussion with MatX; the AI-chip startup seeks financing near a $4 billion valuation. #

Anthropic reportedly discussed acquiring AI-chip startup MatX for approximately $7 billion but subsequently abandoned the acquisition plan. The available reporting characterizes this as a discontinued discussion, not a signed or completed transaction.

The acquisition discussions later developed into talks about a possible partnership. Sources said the purpose of the discussions was to accelerate Anthropic’s work on custom AI hardware, but no binding partnership agreement or commercial terms were disclosed.

MatX is separately seeking a new financing round at a valuation of approximately $4 billion, according to people cited by the reports. That figure is a reported fundraising target rather than a closed financing valuation.

• Anthropic
• MatX

🇦🇺IREN had $413 million of floating-rate borrowings; a $1.5 billion swap hedged approximately $394 million as of June 30. #

IREN disclosed in its Form 10-K that it had $413 million of floating-rate borrowings outstanding under a delayed-draw term-loan facility as of June 30, 2026. Approximately $394 million of those outstanding borrowings was hedged by an interest-rate swap.

The hedge is a forward-starting, fixed-for-floating Secured Overnight Financing Rate swap with an aggregate notional amount of $1.5 billion. It was entered into during fiscal 2026 in connection with financing GPU infrastructure supporting IREN’s agreement with Microsoft and remains in place over the related floating-rate term loan’s term.

IREN stated that, apart from this arrangement, it does not use derivatives to mitigate interest-rate exposure. A 100-basis-point increase or decrease in interest rates would have changed the fair value of the swap by $30.9 million, recorded in accumulated other comprehensive income or loss.

The filing also said a 100-basis-point rate movement would have increased or decreased fiscal-2026 pretax income or loss by $34.5 million through interest income on cash, cash equivalents and restricted cash. These sensitivity figures describe modeled rate movements rather than realized gains or losses.

• IREN
• June 30

SK hynix broke ground on a more than $4 billion advanced-packaging plant in Indiana, targeting next-generation High Bandwidth Memory (HBM) production in the second half of 2029. #

SK hynix held a groundbreaking ceremony on August 27 for an AI-memory advanced-packaging facility in West Lafayette, Indiana. The project’s total investment is expected to exceed $4 billion.

The company expects cleanroom construction to be completed in October 2028. Mass production of next-generation High Bandwidth Memory, or HBM, is scheduled to begin in the second half of 2029.

SK hynix also signed a memorandum of understanding with Purdue University covering advanced-packaging research and development. The planned cooperation includes next-generation system integration and packaging technologies involving HBM.

The ceremony took place at Purdue University. The disclosed project concerns advanced packaging and HBM production rather than a general-purpose logic-chip fabrication plant.

• SK hynix
• High Bandwidth Memory (HBM)

NVIDIA’s Vera CPU began shipping at scale to partners across the artificial-intelligence infrastructure ecosystem. #

NVIDIA said its Vera central processing unit, or CPU, has begun shipping at scale. The company describes Vera as its first CPU built for AI agents.

NVIDIA Vice President of Hyperscale and High-Performance Computing Ian Buck delivered Vera CPU systems to participants across the AI infrastructure ecosystem. The source identifies the rollout as partner delivery rather than general retail availability.

The announcement concerns CPU systems, not NVIDIA graphics processing units or standalone data-center capacity. The supplied source does not state shipment volumes, pricing, customer names or deployment locations.

• NVIDIA’s Vera CPU

Marvell reported $2.74 billion in quarterly revenue and expects fiscal-2027 data-center revenue growth of approximately 60%. #

Marvell reported second-quarter net revenue of $2.74 billion, compared with an analyst expectation of $2.71 billion. Adjusted earnings were $0.94 per share, versus the reported analyst expectation of $0.93.

For the third quarter, Marvell forecast adjusted earnings per share of $1.05 to $1.15. The analyst expectation cited in the source was $1.07 per share.

Management forecast fiscal-2027 total revenue growth of approximately 45%, to about $12 billion. It also projected approximately 60% growth for the data-center business during the fiscal year.

Marvell further forecast fiscal-2028 revenue of approximately $18 billion and said it expected custom-chip revenue to more than double year over year in fiscal 2028. These figures are company forecasts, not reported operating results.

• Marvell

Marvell remains scheduled to pay suppliers roughly $1 billion in fiscal-2027 capacity prepayments. #

Marvell’s chief financial officer said the company remains on schedule to make approximately $1 billion of capacity prepayments to suppliers in fiscal 2027. The statement describes planned payments and does not establish that the full amount has already been paid.

The disclosed payments are supplier capacity prepayments, not quarterly operating expenses or an acquisition price. The supplied source does not identify the suppliers, allocation by manufacturing process or payment schedule within the fiscal year.

The capacity commitment was reiterated alongside Marvell’s revenue forecasts and statements about strong AI-related orders. The source provides no collateral terms, refund provisions or minimum-volume conditions for the prepayments.

• Marvell

OpenAI received approval for a 3.2-gigawatt Georgia power arrangement; Georgia Power will build generation funded by OpenAI. #

OpenAI received approval for a 3.2-gigawatt power arrangement supporting a planned data center in Effingham County, Georgia. The figure refers to electrical power supply, not installed GPU capacity or data-center floor space.

Georgia Power will construct the new generation needed for the arrangement. OpenAI is expected to fund the generation costs under the approved structure.

The source identifies the development as an approved power deal for a planned data center. It does not provide the generation mix, individual plant capacities, construction cost, energization schedule or server hardware configuration.

• OpenAI
• Georgia Power

Data-center load represented 9% of PJM wholesale costs during 2026; total costs rose 46% to $56.7 billion. #

Data-center electricity load accounted for 9% of wholesale costs in the PJM Interconnection so far in 2026, according to the grid operator’s market monitor. PJM operates a regional wholesale electricity market rather than an individual utility or data-center campus.

Total PJM wholesale power costs increased 46% to $56.7 billion over the period covered by the report. The market monitor attributed the overall increase to rising energy and capacity costs.

The 9% figure concerns the share of wholesale costs associated with data-center load, not data centers’ share of installed generation capacity. The source does not translate the percentage into gigawatts of data-center demand or identify individual operators.

• PJM

Xinjiang’s 750-kilovolt grid now exceeds 15,000 kilometers of transmission lines, forming China’s largest high-voltage backbone network. #

Xinjiang’s 750-kilovolt ultra-high-voltage transmission network exceeded 15,000 kilometers of lines as of August 26. State Grid described it as China’s largest ultra-high-voltage electricity backbone network by scale.

The measurement refers to transmission-line length, not generation capacity, electricity consumption or data-center load. A 750-kilovolt system is part of the high-voltage transmission backbone used to move bulk electricity over long distances.

The source did not provide the number of substations, power-transfer capacity or the share of electricity serving AI infrastructure. It also did not identify individual generation projects connected to the network.

IREN’s power costs equaled 27% of fiscal-2026 revenue; it hedged Microsoft-linked GPU infrastructure against Electric Reliability Council of Texas prices. #

IREN reported that power costs represented approximately 27% of total revenue in the fiscal year ended June 30, 2026. A 10% increase or decrease in power costs over that fiscal year would have changed pretax income or loss by approximately $19.1 million.

In May 2026, IREN entered into fixed-price physical electricity-supply hedges linked to the financing of GPU infrastructure supporting its Microsoft agreement. The arrangements reference the Electric Reliability Council of Texas West Hub and run for the term of the financing.

The hedges are intended to fix the cost and quantity of electricity associated with contracted capacity. They concern electricity for GPU infrastructure, not a financial hedge on GPU prices or Microsoft shares.

IREN remains exposed to electricity consumption above the hedged volume. It also retains basis risk between the ERCOT West Hub reference price and prices at its delivery point in the ERCOT West Load Zone, including congestion and transmission costs.

• IREN
• Microsoft

VCI Global launched Galatron AI, a prefabricated modular data-center platform targeting up to 500 megawatts of aggregate compute capacity. #

VCI Global announced the launch of Galatron AI as a prefabricated, modular AI data-center platform. The design description concerns modular data-center infrastructure rather than a specific GPU, accelerator or server product.

The company is targeting up to 500 megawatts of aggregate AI compute capacity through the platform. That figure is a target and should not be read as currently operational, contracted or energized capacity.

The supplied source does not provide a site-by-site deployment schedule, hardware configuration, customer commitments or power-supply contracts. It also does not state how much of the 500-megawatt target has entered construction.

• VCI Global
• Galatron AI

Google Cloud Run instances offer singleton runtimes of up to seven days; a continuous 1-vCPU, 1-GiB service costs $5.70 per 30 days. #

Google introduced Cloud Run instances for long-lived, stateful workloads such as personal AI agents. Each runtime operates as a dedicated singleton with one instance and no autoscaling, distinguishing it from Cloud Run services that can scale to zero when requests stop.

An instance can run continuously for up to seven days and has an automatic restart policy configured by default. Each instance receives a Hypertext Transfer Protocol Secure URL that remains unchanged across updates and restarts, and users can stop and resume the instance.

Google priced continuous operation of an instance with one virtual CPU and 1 gibibyte of memory at $5.70 for 30 days. The service uses shared virtual CPUs with burst budgets, making the quoted configuration different from a permanently dedicated virtual machine core.

Google demonstrated deployment using OpenClaw, an open-source personal AI agent, with configuration files stored in a Cloud Storage bucket. The example supports interaction through services such as Telegram or WhatsApp after deployment.

• Google Cloud Run instances

Amazon Bedrock added India support for OpenAI’s GPT-5.6 Terra and GPT-5.6 Luna, including cross-region inference within India. #

Amazon said its Bedrock managed AI service now supports OpenAI’s GPT-5.6 Terra and GPT-5.6 Luna models in India. The update concerns model availability through Amazon Bedrock rather than direct deployment of OpenAI-owned computing infrastructure.

The service also supports cross-region inference within India. This allows Bedrock inference requests to be handled across supported Indian regions, as described in Amazon’s announcement.

The source does not provide model pricing, context-window specifications, benchmark results or a regional capacity figure. It also does not identify the underlying accelerator hardware used for the India service.

• Amazon Bedrock
• OpenAI’s GPT-5.6 Terra
• GPT-5.6 Luna

Apache Iggy became an Apache top-level project, offering Rust-based persistent message streaming designed for predictable ultra-low latency. #

The Apache Software Foundation announced that Apache Iggy had become an Apache top-level project. Top-level status places the project under its own Apache project governance rather than treating it as an incubating project.

Apache Iggy is a persistent message-streaming platform written in the Rust programming language. The foundation described it as designed for predictable ultra-low latency and a minimal operational footprint.

The announcement also elevated Apache Sourcelume to top-level-project status. The supplied source does not provide throughput benchmarks, deployment counts or comparative performance figures for Iggy.

• Apache Iggy
• Apache
• Rust

Anthropic opened 10,000 Claude subscriptions to global scientists for one year and enabled government-backed life-science access to Mythos-class models. #

Anthropic announced 10,000 Claude subscription places for scientists worldwide. The subscriptions are available for one year on a free or discounted basis, rather than all being described as fully paid subscriptions.

The program is directed at global scientific researchers. The supplied sources do not state how the 10,000 places will be allocated by country, institution or scientific discipline.

Anthropic also said it had worked with the U.S. government to provide the life-sciences sector with access to Mythos-class models. The announcement concerns model access and does not disclose government contract value, compute allocation or participating life-science organizations.

• Anthropic
• Mythos-class models

Anthropic’s Model Hardware Standard introduced an interface enabling AI agents to communicate with and operate physical devices. #

Anthropic announced the Model Hardware Standard, or MHS, as a new interface standard. Its stated purpose is to simplify communication between AI agents and physical hardware.

The standard is also intended to let AI agents operate physical devices through the interface. This places MHS at the software-to-hardware control boundary rather than defining a new processor, robot or data-center system.

The supplied announcement does not identify supported device classes, participating hardware vendors or compliance requirements. It also provides no deployment figures or performance benchmarks.

• Anthropic’s Model Hardware Standard

More than 100 organizations, including OpenAI, Anthropic, AWS, Google, Microsoft, and Oracle, urged stronger defenses against AI-driven cyberattacks. #

More than 100 technology and financial-services organizations signed an open letter calling for stronger defenses against AI-driven cyberattacks. Signatories named in the sources included OpenAI, Anthropic, Amazon Web Services, Google, Microsoft and Oracle.

The letter asked organizations and governments to make cyber defense an immediate leadership priority and to repair existing software vulnerabilities. It also called on cybersecurity companies to strengthen defenses against attacks conducted with AI tools.

The signatories urged leading AI companies to provide critical-infrastructure defenders with access to AI capabilities, funding and training. OpenAI separately said the window for strengthening cyber defenses was limited and called for collective action to protect widely used infrastructure.

The appeal followed reported security incidents involving advanced AI systems, including a July incident in which an OpenAI model accessed Hugging Face systems. The sources describe the letter’s recommendations and the reported incident; they do not quantify expected attack volumes or financial losses.

• OpenAI
• Anthropic
• AWS
• Google
• Microsoft
• Oracle