AI & Compute Infrastructure — September 4, 2026

🇺🇸U.S. payrolls rose 162,000 in August; unemployment held at 4.1%, while labor-force participation reached 61.6%. #

U.S. nonfarm payrolls increased by 162,000 in August, compared with a forecast of 55,000 and a previous reading of negative 23,000. Private payrolls rose by 127,000, while government payrolls increased by 35,000.

The unemployment rate remained at 4.1%, matching both the forecast and previous reading. Labor-force participation increased to 61.6% from 61.4%, exceeding the 61.5% forecast.

Manufacturing employment increased by 16,000, versus a forecast of 5,000 and a previous gain of 5,000. Average hourly earnings rose 0.3% month over month and 3.1% year over year, while the average workweek increased to 34.4 hours.

Following the report, federal-funds futures indicated a 58.6% probability of a 25-basis-point September rate increase and a 41.4% probability of no change. The employment release therefore raised market expectations for tighter monetary policy, although several cited analysts said the next inflation report remained important to the decision.

• U.S. payrolls

🇺🇸Anthropic’s IPO prospectus is expected in late September, with marketing potentially beginning by mid-October. #

Sources cited by Reuters said Anthropic expected to make its initial public offering prospectus public in late September. Marketing for the offering could begin as early as mid-October, but the reports did not state that either milestone had been completed.

Separate reporting said Anthropic was close to selecting Morgan Stanley and Goldman Sachs for key underwriting roles. Morgan Stanley was described as the leading candidate for the lead-left position, while Goldman Sachs was expected to serve as stabilization agent; those assignments had not been finalized.

JPMorgan, Citigroup and Barclays, which had previously provided debt financing to Anthropic, were also reported as potential participants in the underwriting group. The source material contains differing reports about the possible filing and listing timetable, so the late-September prospectus and mid-October marketing dates remain reported expectations rather than filed facts.

• Anthropic’s IPO prospectus

🇬🇧Nscale is discussing $1.5 billion of convertible notes and about $2 billion from NVIDIA; terms remain unsettled. #

London-headquartered AI cloud company Nscale was reported to be discussing as much as $3.5 billion of pre-IPO financing. The proposed structure comprised up to $1.5 billion of convertible notes sold to investors and approximately $2 billion of financing from NVIDIA.

Third Point, led by Daniel Loeb, was reported to be preparing to lead the convertible-note investment, with Goldman Sachs assisting the fundraising. Discussions were continuing, and the size, investor group and other details could change; no final decision or closed financing was reported.

The report also said Nscale had previously considered an IPO that could raise approximately $3 billion. It described a large data-center campus under construction in Norway for Microsoft, another large project in West Virginia, and approximately 194,000 NVIDIA Vera Rubin graphics processing units purchased by the company.

Nscale reportedly presented investors with approximately $10.3 billion of total contract value while also describing a $45 billion compute-supply agreement with Anthropic. The same presentation cited approximately $18.1 billion of prospective annual revenue and $13.6 billion of adjusted earnings before interest, taxes, depreciation and amortization, but explicitly characterized those figures as illustrative estimates rather than formal guidance.

• Nscale
• NVIDIA

🇺🇸Blue Owl is considering a listed data-center real estate investment trust seeded with approximately $6.5 billion of owned assets. #

Blue Owl Capital was reported to be considering the creation and public listing of a data-center real estate investment trust, or REIT. Approximately $6.5 billion of data-center assets already owned by Blue Owl would serve as the proposed REIT’s initial portfolio.

The reported plan would use an initial public offering and subsequent share sales to raise capital for additional data-center acquisitions. The source did not report that the REIT had been legally formed, that an IPO filing had been submitted, or that a transaction had closed.

Blue Owl completed fundraising in May for its latest digital-infrastructure fund, which received $7 billion of capital commitments. Discussions concerning the REIT and IPO were still underway, and the proposed terms remained subject to change.

At the time of the report, Blue Owl had a market capitalization exceeding $18 billion, and its shares had declined by nearly 21% year to date. The shares rose as much as 1.2% following the report.

• Blue Owl

Global semiconductor-equipment billings reached $40.53 billion in the second quarter, increasing 23% year over year. #

Global semiconductor-manufacturing equipment billings totaled $40.53 billion in the second quarter of 2026. That represented an increase of 23% from the corresponding quarter of 2025.

The figures came from SEMI’s Worldwide Semiconductor Equipment Market Statistics report. SEMI is an industry association covering the global semiconductor and electronics design and manufacturing supply chain.

The disclosed amount refers specifically to semiconductor-equipment billings, not semiconductor revenue, chip shipments, fabrication capacity or AI-accelerator sales. The supplied source summary did not provide regional billings, equipment-category splits or a sequential growth rate.

AMD announced Threadripper Halo Station, combining Threadripper Pro CPUs with Instinct MI350P accelerators for AI developers. #

AMD announced Threadripper Halo Station at IFA 2026 as a high-end workstation for AI developers. The system combines AMD Threadripper Pro central processing units with AMD Instinct MI350P accelerators.

The components serve separate compute roles: Threadripper Pro is the workstation’s CPU platform, while Instinct MI350P is an accelerator intended for AI and high-performance computing workloads. The announcement described a workstation rather than a rack, server cluster or cloud service.

The supplied article summary did not disclose CPU core counts, accelerator quantities, memory capacity, system power, pricing or availability. It also did not state benchmark results or comparative performance figures.

• AMD
• Threadripper Halo Station
• Threadripper Pro CPUs
• Instinct MI350P accelerators

IBM Nighthawk r2, a 120-qubit quantum processor, exceeds 100,000 circuits per second through a new reset architecture. #

IBM made its 120-qubit Nighthawk r2 quantum processor available through the IBM Quantum Platform. IBM said the processor can execute more than 100,000 quantum circuits per second.

The throughput is attributed to a new qubit-reset architecture. According to IBM, the design provides as much as 25 times the circuit throughput of its Heron systems.

The figure measures circuit execution throughput, not classical floating-point performance or logical-qubit count. The source identifies 120 physical qubits but does not provide a logical-qubit total, error-correction configuration or application-level benchmark.

The performance statements are IBM’s reported measurements as relayed by HPCwire. The supplied material does not include independent benchmark verification.

• IBM Nighthawk r2

Global DRAM market revenue rose 55.9% quarter over quarter during the second quarter, reaching a record high. #

According to CFM Flash Market data, second-quarter global dynamic random-access memory, or DRAM, market revenue reached US$147.024 billion (1,470.24亿美元). That was a 55.9% quarter-over-quarter increase and a record high.

Manufacturers continued expanding sales of higher-value server DRAM and high-bandwidth memory, or HBM, in response to strong server demand. The report said the market experienced increases in both shipment volume and pricing.

Average selling prices for server and PC DRAM had approximately doubled quarter over quarter in the first quarter. Their second-quarter increase slowed to a range of 40% to 50%, while still supporting substantial revenue growth.

For the third quarter, manufacturers and some server customers had reportedly set upper bounds under long-term pricing agreements. The report expected DRAM average selling prices to continue rising sequentially at a slower rate, but that statement was a forecast rather than a reported third-quarter result.

• DRAM

Quantum Foundry Copenhagen plans a 5,300-square-meter European quantum-chip fabrication facility with the Novo Nordisk Foundation. #

Quantum Foundry Copenhagen and the Novo Nordisk Foundation announced plans to establish a 5,300-square-meter quantum-chip fabrication facility in Copenhagen. The announcement described it as a European facility for quantum-chip design, manufacturing and scaling.

The project concerns semiconductor-style fabrication infrastructure for quantum chips rather than a quantum computer or conventional processor fabrication plant. Its stated purpose is to expand Europe’s capability to design and manufacture next-generation quantum devices.

The source describes the facility as planned, not operational. The supplied summary does not disclose construction cost, production capacity, process technology, equipment vendors, opening date or customer commitments.

• Quantum Foundry Copenhagen
• Novo Nordisk Foundation

Oakville and Rocky View County imposed one-year data-center moratoriums; Manitoba rejected a proposed Winnipeg-area project. #

Oakville, Ontario, and Rocky View County, Alberta, each imposed a one-year pause on data-center development. Manitoba Premier Wab Kinew separately rejected a proposed project outside Winnipeg, saying its disadvantages outweighed its benefits.

A proposed project from Canadian entrepreneur and television personality Kevin O’Leary, described as having capacity of up to 7.5 gigawatts, also faced opposition. Protesters separately opposed a proposed Vancouver data center that had support from multiple levels of government.

Reported community concerns included electricity consumption, noise, environmental effects and the possibility of higher electricity prices. One Angus Reid Institute poll found that 68% of respondents would oppose a large AI data center near their home, while a Leger survey found four-fifths were concerned about possible electricity-price increases and environmental effects.

The Canadian government issued a framework calling for projects to produce long-term local benefits, reduce environmental effects and avoid transferring electricity costs to residents. Amazon, Meta and Anthropic were among the technology companies reported to have signed the associated principles; another report said Canada had five operating large data centers and 159 announced or under-construction projects.

• Oakville
• Rocky View County
• Manitoba

China’s revised electricity-safety regulations, effective January 1, 2027, add county-level outage controls and encourage grid investment. #

China published revised Regulations on Emergency Response, Investigation and Handling of Electric Power Safety Accidents (《电力安全事故应急处置和调查处理条例》), which will take effect on January 1, 2027. The revisions extend electricity-safety controls from a primary focus on urban grids to broader urban and rural coverage.

The regulations add county-level power-outage control provisions. China’s National Energy Administration said county grids are critical infrastructure connecting urban and rural development and supporting public services.

The agency identified grid architecture and equipment operating pressure as continuing issues in some remote counties. It said the revisions would encourage investment in county-grid construction, operations, maintenance and emergency-response systems.

The rules also call for greater integration of urban and rural standards for grid planning, construction, operation and maintenance. The control standards are intended to account for regional development conditions while balancing safety spending and development benefits.

European Union gas storage stood at 65% capacity following Thursday’s Gas Coordination Group meeting. #

The European Commission said the European Union’s Gas Coordination Group met on Thursday. Following that meeting, a Commission spokesperson reported that EU gas storage was at 65% of capacity.

The percentage describes aggregate storage fullness, not daily gas supply, consumption, import capacity or electricity-generation capacity. The source did not provide a country-by-country breakdown.

No target level, comparison with the previous year or forecast for the remainder of the storage season was included in the supplied message. The update therefore establishes the reported 65% level and the meeting context only.

• European Union
• Gas Coordination Group

Gazprom plans to maintain 73.3 billion cubic meters of natural gas in domestic storage for winter. #

Gazprom plans to have 73.3 billion cubic meters of natural gas in domestic storage for winter. The figure refers to stored gas volume inside Russia rather than annual production, exports or pipeline capacity.

The source frames 73.3 billion cubic meters as a planned winter storage level. It does not state that the entire volume had already been injected or that the target had been reached on September 4.

No facility-level distribution, storage-capacity percentage, injection schedule or comparison with prior winters was provided. The message also did not specify how much of the stored volume would be available for withdrawal during peak demand.

• Gazprom

Sharon AI signed a five-year agreement adopting Rafay Systems as its centralized AI-infrastructure orchestration and operations layer. #

SharonAI Holdings, whose common stock and warrants trade on Nasdaq under SHAZ and SHAZW, disclosed a five-year strategic agreement with Rafay Systems in an exhibit to a Form 8-K. Sharon AI said it would use Rafay as a centralized orchestration and operations layer across its AI Factory environments.

The platform is intended to standardize how accelerated-computing infrastructure is provisioned, governed, monitored and exposed to customers across locations, tenants and workloads. The agreement moves Sharon AI from individual cluster deployments toward a common operating framework.

Rafay’s disclosed capabilities include bare-metal provisioning, infrastructure lifecycle management, virtual machines, Kubernetes, GPU platform-as-a-service, AI development environments and token-metered inference services. Sharon AI also cited automation, observability, governance, secure multitenancy and GPU-utilization management.

The filing establishes that Sharon AI publicly disclosed the five-year agreement, but the supplied exhibit does not state its financial value, minimum spending obligations or termination terms. Statements that the collaboration is expected to improve utilization, reliability or service delivery are company expectations, not reported operating results.

• Sharon AI
• Rafay Systems

Google Cloud VM Extension Manager reached general availability, adding fleet-wide policies, drift detection, self-healing, and automated rollbacks. #

Google Cloud made Virtual Machine Extension Manager generally available for Compute Engine fleets. The service is designed to replace custom startup-script approaches for managing software extensions inside guest operating systems.

Administrators can define declarative, project-wide policies that specify the desired software state across regions and zones. The manager continuously checks for configuration drift and can automatically restore instances to the defined state.

The service also supports phased deployments across multiple zones and automated rollback when a rollout fails. Fleet-health information is centralized and integrated with Google Cloud Monitoring.

The announcement concerns guest-software lifecycle management on virtual machines, rather than provisioning physical servers or managing accelerator scheduling. Google did not provide pricing, fleet-size limits or measured operational savings in the supplied update.

Yahoo uses flexible virtual-machine rankings and Auto-Zone placement to maintain Managed Service for Apache Spark cluster provisioning. #

Yahoo uses flexible virtual-machine configurations in Google Cloud’s Managed Service for Apache Spark, formerly Dataproc. Instead of requiring one fixed machine type, Yahoo supplies a ranked list of acceptable VM shapes so the service can search for available capacity.

Auto-Zone placement must also be enabled to search across zones within a selected region. Users can do this by passing a region or an empty zone string, allowing cluster provisioning to move beyond a single zone’s capacity constraints.

For autoscaled clusters, the acceptable machine types should have similar CPU-core counts and memory capacities, even when they belong to different VM families. Google said a consistent CPU-to-memory ratio across primary and secondary workers is important because the smallest ratio determines effective container sizing.

The configuration supports Yahoo’s analytics environment following its migration of on-premises Hadoop and big-data estates to Google Cloud. Its purpose is to reduce provisioning delays and manual intervention when a specific VM shape has constrained capacity, rather than to guarantee that every requested machine type is always available.

• Yahoo
• Auto-Zone

On Google Cloud TPU v6e accelerators, Gemma 3 27B generation plateaued beyond 64 concurrent users; Gemma 3 12B scaled further. #

Google benchmarked Gemma 3 12B and Gemma 3 27B on a Google Kubernetes Engine Autopilot cluster connected to a single-host Tensor Processing Unit v6e node pool. The TPU configuration used a 2-by-2 chip topology, with vLLM from the vllm-project/tpu-inference stack serving the models.

For decode-heavy text generation, Gemma 3 27B reached a scaling limit after 64 concurrent users. At 128 users, its normalized throughput plateaued at 4.12 times the baseline, while Gemma 3 12B reached an 8.19-times normalized throughput multiplier.

For prefill-heavy classification workloads, the two models produced more similar scaling. At 128 concurrent users, peak normalized throughput was approximately 6.0 to 6.4 times baseline for both models, without saturating the TPUs in that test.

The tests covered 16, 32, 64 and 128 concurrent users. Google reported that saturation could appear as latency spikes and silent request drops, and recommended using end-to-end latency rather than standard CPU or memory metrics for scaling decisions; these recommendations apply to the described benchmark configuration.

OpenAI made GPT-6 Astra available to all Pro, Enterprise, and Business Premium users and through its API. #

OpenAI chief executive Sam Altman said GPT-6 Astra was available to all Pro, Enterprise and Business Premium users. He also said the model was available through OpenAI’s application programming interface, or API.

A separate OpenAI update described access through ChatGPT Work and Codex for users in those subscription categories. The source did not provide API prices, context-window size, benchmark results or model architecture details.

Altman said OpenAI would next begin expanding access to Plus and Business users. That statement described the next rollout stage; the supplied sources did not state that access for those two tiers was already complete.

• OpenAI
• GPT-6 Astra
• Pro
• Enterprise
• Business Premium
• API

Japan plans domestically developed AI satellites, building test units during fiscal 2027–2029 and launching from fiscal 2030 onward. #

Japan’s Ministry of Defense plans to develop satellites equipped with artificial intelligence to analyze space-surveillance data. The stated purpose is to accelerate decision-making connected with Japan’s ability to respond to attacks using long-range missiles.

The ministry plans to work with Japanese companies to build test satellites during fiscal years 2027 through 2029. Performance testing would initially take place on the ground.

Launches are planned for fiscal 2030 or later, after which the satellite functions would be tested in space. The source therefore describes a development and test schedule, not an existing operational constellation.

The onboard AI system is expected to be developed domestically in Japan. The supplied report did not specify satellite quantity, launch provider, orbital regime, processor hardware, budget or deployment capacity.

National Highway Traffic Safety Administration opened a review covering approximately 1,000 steering-wheel-free Tesla Cybercab vehicles. #

The U.S. National Highway Traffic Safety Administration opened a review concerning approximately 1,000 Tesla Cybercab vehicles. The agency was examining the certification process and technical data Tesla used to demonstrate compliance with federal motor-vehicle safety standards.

Cybercab does not permanently install conventional human-driving controls. The identified omissions include a steering wheel, brake pedal, accelerator pedal and rear-view mirrors.

The action was described in the source material as a review inquiry into certification and compliance, rather than a completed finding that the vehicles violated federal standards. NHTSA had also said it was in contact with Tesla regarding the Cybercab rollout and was evaluating the situation.

Tesla said users could access Cybercab service through the Tesla Robotaxi application within a limited area of Austin, Texas. The supplied sources did not state that NHTSA had ordered a recall, suspended the service or reached a final safety determination.

• Tesla Cybercab vehicles