AI & Compute Infrastructure — August 31, 2026

🇺🇸Anthropic signed a $35 billion cloud agreement with GPU-cloud provider Lambda; NVIDIA holds the lease for Hut 8’s Texas site. #

Anthropic signed a reported $35 billion cloud-computing agreement with Lambda, an NVIDIA-backed cloud provider. The capacity will come from a data center under development by bitcoin miner and data-center developer Hut 8 in Nueces County, Texas.

NVIDIA reportedly signed an agreement with Hut 8 several weeks earlier to secure the data-center capacity and holds the lease for the site. Lambda will use the facility to deploy chips purchased from NVIDIA, which is also an investor in Lambda; the amount Lambda will pay NVIDIA for use of the space was not disclosed.

The arrangement lets Lambda serve Anthropic without directly procuring the data-center space. The reporting characterized Anthropic as non-investment-grade and said it had been seeking additional cloud-capacity agreements after experiencing compute-supply constraints earlier in the year.

• Anthropic
• Lambda
• NVIDIA
• Hut 8

🇺🇸AI-infrastructure provider Nscale closed approximately $3 billion in financing for deployments in Ward County, Texas, and Madison County, North Carolina. #

Nscale closed approximately $3 billion in financing tied to two artificial-intelligence deployments. The identified locations are Ward County, Texas, and Madison County, North Carolina.

The source describes the financing as closed rather than a targeted fundraise or maximum facility size. It does not provide a breakdown of the approximately $3 billion between the two deployments.

No financing conditions, lenders, collateral, interest rates, customer commitments, hardware configuration, power capacity, or deployment schedule were disclosed in the supplied reports. The sources also do not specify whether the funding consists of debt, equity, project finance, or a combination.

• Nscale
• Ward County, Texas
• Madison County, North Carolina

🇮🇱CollPlant Biotechnologies signed a definitive agreement to acquire photonic-computing company LightSolver, with closing expected this week. #

CollPlant Biotechnologies signed a definitive agreement to acquire LightSolver, an Israeli deep-technology company developing a computing architecture based on lasers. The transaction is intended to bring CollPlant into the supercomputing and photonic-computing markets.

The announced transaction remains pending: the definitive agreement has been signed, while closing was expected during the same week as the Aug. 31 announcement. The source therefore does not establish that the acquisition had already closed.

The consideration was described as equity-centric and including an upfront closing component. The supplied article excerpt does not disclose the purchase price, share count, ownership dilution, closing conditions, earnouts, assumed liabilities, or post-closing governance.

• CollPlant Biotechnologies
• LightSolver

🇪🇺The European Commission designated ChatGPT a very large online search engine, triggering stricter Digital Services Act requirements. #

The European Commission designated ChatGPT as a “very large online search engine” under the European Union’s Digital Services Act. The designation places the service under the law’s stricter regulatory requirements.

In the same announcement, the Commission designated Reddit and Roblox as “very large online platforms.” All three services are U.S.-based, but ChatGPT’s classification differs because it was designated as a search engine rather than an online platform.

The supplied sources do not enumerate the individual compliance measures, deadlines, user thresholds, or potential penalties attached to ChatGPT’s designation. They establish the classification and the resulting application of stricter Digital Services Act oversight.

• European Commission
• ChatGPT
• Digital Services Act

EuroHPC Joint Undertaking signed Bull to deliver Finland’s €387.8 million LUMI-AI, using AMD MI430X GPUs and sixth-generation 256-core EPYC CPUs. #

The EuroHPC Joint Undertaking and the LUMI AI Factory consortium selected Bull to deliver LUMI-AI under a €387.8 million procurement contract. The system will be hosted at CSC – IT Center for Science’s data center in Kajaani, Finland.

LUMI-AI will use AMD Instinct MI430X graphics processing units and sixth-generation AMD EPYC central processing units with 256 CPU cores each. These are distinct compute components: the MI430X devices are GPU accelerators, while the EPYC processors are host CPUs.

The project is described as a next-generation artificial-intelligence supercomputer for European research and industry. The supplied excerpts do not state the number of GPUs or CPUs, system power, rack count, memory capacity, interconnect configuration, measured performance, or installation date.

• EuroHPC Joint Undertaking
• Bull
• AMD MI430X GPUs

CXMT is producing HBM3E (High-Bandwidth Memory 3E), a memory generation used by leading AI processors. #

ChangXin Memory Technologies, commonly known as CXMT, is producing High-Bandwidth Memory 3E, or HBM3E, according to reporting attributed to The Information. The development was characterized as a breakthrough in advanced memory chips.

HBM3E is an advanced type of high-bandwidth memory used in many leading AI processors. It is memory rather than a graphics processing unit, neural processing unit, or complete accelerator.

The supplied reports do not specify production volume, customer qualification, yields, stack height, memory capacity, bandwidth, manufacturing process, shipment timing, or customers. They also do not state whether production is sampling, qualification output, or high-volume manufacturing.

• CXMT
• HBM3E (High-Bandwidth Memory 3E)

Amazon EC2 R9g and R9gd instances entered general availability, using Graviton5 CPUs with up to 25% better compute performance. #

Amazon Web Services made its Amazon Elastic Compute Cloud R9g and R9gd instance families generally available. Both are powered by AWS Graviton5 central processing units.

AWS says the new instances deliver up to 25% better compute performance than the preceding R8g generation. This is a vendor-provided “up to” performance comparison rather than an independently reported benchmark result.

The memory-focused instances are positioned for databases, in-memory caches, and real-time analytics. The supplied source does not provide instance sizes, memory-to-vCPU ratios, local-storage specifications for R9gd, regional availability, or pricing.

• Graviton5 CPUs

MediaTek joined the NVIDIA NVLink Fusion ecosystem as both companies expanded edge-to-cloud AI-platform collaboration. #

MediaTek joined NVIDIA’s NVLink Fusion ecosystem as the companies expanded their existing collaboration on edge-to-cloud AI computing platforms. NVIDIA chief executive Jensen Huang described the relationship as a major engineering alignment and MediaTek as one of NVIDIA’s largest partners.

The announcements also said NVIDIA would invest $3.5 billion through convertible bonds issued by MediaTek. The supplied reports do not provide the bonds’ interest rate, maturity, conversion price, conversion ratio, collateral, covenants, or resulting potential dilution.

MediaTek separately said Alphabet, described as a long-term AI-infrastructure partner, participated in its record $3.9 billion overseas convertible-bond offering. The sources do not clarify the allocation among investors beyond NVIDIA’s reported $3.5 billion investment.

• MediaTek
• NVIDIA NVLink Fusion

AMD released ROCm 10 and made its AI-native ROCm.AI development and optimization environment generally available. #

AMD released ROCm 10, marking ten years of its ROCm software stack. ROCm is AMD’s software platform for programming and operating supported GPU-accelerated computing hardware.

AMD also moved ROCm.AI into general availability after first introducing it at Advancing AI 2026. The company describes ROCm.AI as an AI-native software experience intended to accelerate development and performance optimization on AMD hardware.

The announcement says ROCm.AI brings together three core developer experiences, although the supplied article excerpt does not identify them individually. It also does not provide benchmark results, supported-model lists, hardware-compatibility details, or comparisons with competing software environments.

• AMD
• ROCm 10
• ROCm.AI

Saudi PIF-backed HUMAIN and digital-infrastructure company DataVolt began building Oxagon’s planned data center; the initial 100 megawatts are expected online in 2028. #

HUMAIN, which is backed by Saudi Arabia’s Public Investment Fund, expanded its cooperation with digital-infrastructure company DataVolt. The companies began construction of a planned AI data center at Oxagon, the industrial area under development within NEOM.

The first phase is expected to provide 100 megawatts of capacity and enter service in 2028. The figure refers to planned data-center capacity, not the number or aggregate rated power of installed GPUs.

The supplied source does not state the project’s total planned capacity beyond the initial phase, capital cost, financing structure, customer contracts, hardware suppliers, cooling architecture, or whether the 2028 target represents construction completion or operational availability.

• HUMAIN
• DataVolt

HUMAIN, AMD, and Cisco plan up to 250 megawatts of Saudi AI infrastructure from 2027, using MI400-series GPUs. #

HUMAIN, AMD, and Cisco announced that their joint venture plans to deploy up to 250 megawatts of AI infrastructure in Saudi Arabia beginning in 2027. The next expansion phase is scheduled to start that year.

The deployment is planned around AMD Instinct MI400-series graphics processing units. The 250-megawatt figure describes the upper limit of planned infrastructure capacity, not committed GPU shipments or currently operating data-center load.

The joint venture also remains on a stated plan to deploy up to 1 gigawatt of AI infrastructure by 2030. Both “up to” figures are forward plans; the supplied sources do not disclose financing, customers, site locations, GPU quantities, procurement commitments, or construction milestones.

• HUMAIN
• AMD
• Cisco
• MI400-series GPUs

ChronoScale plans a 50-megawatt North American deployment with Microsoft, featuring NVIDIA GB300 NVL72 systems and liquid cooling. #

ChronoScale announced plans with Microsoft for a 50-megawatt AI-compute deployment in North America. The announcement describes a planned deployment rather than capacity already operating.

The project is expected to use NVIDIA GB300 NVL72 systems. NVL72 denotes a rack-scale accelerated-computing system, while the announced 50 megawatts refers to deployment capacity rather than a count of GPUs, racks, or servers.

Advanced liquid-cooling infrastructure is planned to support the high-density AI workloads. The source does not disclose the specific location, number of NVL72 systems, investment amount, financing, binding capacity commitment, commissioning date, or Microsoft’s precise contractual role.

• ChronoScale
• Microsoft
• NVIDIA GB300 NVL72 systems

Regional grid operator PJM Interconnection removed Oklo’s 750-megawatt mixed-technology project from its study cycle; Oklo requested federal reinstatement. #

PJM Interconnection removed Oklo’s 750-megawatt mixed-technology project from the regional transmission organization’s interconnection-study cycle. An interconnection study evaluates a project’s ability to connect to the grid; removal is not the same as cancellation of an operating plant because the project had not reached that stage.

PJM reportedly told Oklo that the company had not demonstrated the project could ride through a sudden decline in grid voltage. The 750-megawatt figure applies to the mixed-technology project as a whole and should not be treated as the capacity of a single nuclear reactor.

Oklo asked the Federal Energy Regulatory Commission to direct PJM to reinstate the project. The supplied source does not report that FERC had granted the request or that the project had returned to the study cycle.

• PJM Interconnection
• Oklo

Iran’s Darahu combined-cycle plant commissioned a 290-megawatt steam-turbine unit, representing a €240 million investment. #

Iran’s Darahu combined-cycle power plant formally commissioned a steam-turbine generating unit. The unit has installed generation capacity of 290 megawatts.

The reported investment for the unit was €240 million. The capacity describes electricity-generation capability and is not a data-center or AI-compute power allocation.

The supplied report does not identify the turbine vendor, fuel source for the combined-cycle plant, project financing, construction start, efficiency, expected annual generation, or grid-connection details.

• Iran’s Darahu combined-cycle plant

AWS and Azure launched multicloud interconnect previews across U.S. East, U.S. West, Sydney, and Frankfurt regions. #

Amazon Web Services and Microsoft Azure introduced preview services for private multicloud connectivity between their platforms. Azure described its Multicloud Interconnect for AWS as a cloud-native networking service built on open interoperability standards.

The preview became available in U.S. East, U.S. West, Asia Pacific (Sydney), and Europe (Frankfurt). These are cloud-region locations rather than a single global production launch.

The stated purpose is to simplify private connectivity and reduce operational complexity for multicloud and AI workloads while providing high-performance links between the two cloud environments. The supplied sources do not provide bandwidth tiers, latency measurements, service-level agreements, pricing, or a date for general availability.

• AWS
• Azure

BigQuery Graph became generally available with native Graph Query Language, petabyte-scale traversals, and no extract-transform-load requirement. #

Google Cloud made BigQuery Graph generally available after offering it in preview. The service brings graph capabilities directly into the BigQuery data warehouse, with International Organization for Standardization-standard Graph Query Language alongside SQL.

Graph traversals run natively without requiring extract-transform-load, or ETL, into a separate graph database. Google says the service reaches petabyte scale without the memory bottlenecks associated with a scale-up database and operates under existing row- and column-level security.

BigQuery Graph can call BigQuery ML and AI functions in the same query. Google identified uses including fraud and threat detection, supply-chain digital twins, identity resolution, customer profiles, knowledge graphs, Graph Retrieval-Augmented Generation, network topology, service dependencies, and data lineage.

Thales Cybersecurity Products said it uses property-graph modeling for multi-hop traversals and anomaly clustering in its threat-detection pipeline. It also combines the graph output with Gemini models to generate threat narratives, according to the customer statement included by Google.

• BigQuery Graph
• Graph Query Language

Google Cloud Run added dedicated singleton instances; continuous one-vCPU, one-gibibyte operation costs $5.70 for 30 days. #

Google Cloud introduced dedicated singleton compute runtimes for Cloud Run. Unlike runtimes that shut down when idle, these instances are designed to remain active continuously.

Google stated that operating one Cloud Run instance with one virtual central processing unit and one gibibyte of memory continuously for 30 days costs $5.70. The figure is tied to that specific resource configuration and operating period.

Google positioned the instances as infrastructure for personal AI agents that require continuously available compute at low cost. The supplied source does not provide regional price differences, networking or storage charges, accelerator support, service-level terms, or limits on the singleton configuration.

• Google Cloud Run

Diraq plans to deploy a silicon-spin quantum computer inside an Equinix Sydney data center, a shared commercial facility. #

Diraq and Equinix announced plans to deploy a Diraq quantum computer at an Equinix data center in Sydney, Australia. Diraq’s system is based on silicon-spin quantum-computing technology.

The companies describe the planned installation as the world’s first silicon-spin quantum computer to operate inside a shared commercial data center. This is a planned deployment and a company announcement, not confirmation that the system is already operational.

The shared commercial setting distinguishes the installation from a dedicated laboratory or privately operated research facility. The supplied source does not disclose qubit count, error rates, system performance, deployment date, customer-access model, power consumption, or commercial terms.

• Diraq
• Equinix Sydney data center

Anthropic temporarily reassigned 150 engineers to security and restarted external pre-release cybersecurity evaluations after imposing partner best-practice requirements. #

Anthropic temporarily shifted 150 engineers to security work following a security-breach incident. The company described the reassignment as temporary and said the engineers’ work focus had been redirected toward security matters.

Anthropic initially paused external cybersecurity testing of pre-release models. It then required organizations testing pre-release models with reduced cybersecurity protections to commit to a set of best practices, including operating requirements for partners with access to Claude Mythos 5.

After those best-practice commitments were implemented, Anthropic said it restarted external cybersecurity evaluations. The supplied sources do not describe the breach, identify affected systems, quantify damage, specify how long the reassignment will last, or enumerate the complete best-practice framework.

• Anthropic

The U.S. Defense Department added Grok Government and ChatGPT for unclassified military uses, including procurement research and supply-chain management. #

The U.S. Department of Defense announced the addition of Starshield AI’s Grok Government and OpenAI’s ChatGPT for unclassified military use. The deployment expands the number of AI assistants available through the department’s platform.

The department said users would receive features including deep-thinking reasoning, adaptive reasoning modes identified as automatic, fast, and expert, customizable workspaces, persistent projects, and reusable “playbooks.” The playbooks are intended to capture and distribute institutional knowledge.

The Defense Department identified procurement-market research and supply-chain management for logistics personnel as example use cases for Grok Government. The supplied source does not provide contract values, user counts, model versions, hosting arrangements, data-retention terms, or a schedule for broader adoption.

• U.S. Defense Department
• Grok Government
• ChatGPT

Galbot Store, Hong Kong’s first fully autonomous robotic retailer, begins operations September 1 without remote staff control. #

Galbot Store held its Hong Kong launch ceremony on Aug. 31 and was scheduled to begin formal operations on Sept. 1. The store is described as Hong Kong’s first fully autonomous robotic retail outlet and Galbot’s first location outside its home market.

At the demonstration, the humanoid robot store manager “Xiao Gai” (小盖) received orders, selected goods, and delivered them without remote control by staff. The operator described the process as relying on the robot’s autonomous perception and decision-making capabilities.

Hong Kong Investment Corporation co-organized the launch. Hong Kong Financial Secretary Paul Chan said the government supports embodied-intelligence deployments while emphasizing rules, standards, and safety, and referred to links between Hong Kong research, talent and capital and the Greater Bay Area’s manufacturing capabilities.

• Galbot Store
• September 1