AI & Compute Infrastructure — August 6, 2026

🇺🇸Alphabet sought up to $25 billion through ten bond tranches, attracting approximately $115 billion of orders. #

Alphabet launched a ten-part U.S.-dollar investment-grade bond offering and sought to raise as much as $25 billion. The transaction added a large block of corporate-bond supply to the market, with longer-dated U.S. Treasury yields reaching intraday highs after the offering was announced.

Orders reached approximately $115 billion, more than four times the contemplated maximum issuance. The reported demand also exceeded the order books for recent AI-related bond offerings from Amazon and SpaceX.

Alphabet offered a relatively high new-issue concession to attract investors. The sources describe the $25 billion figure as the maximum sought, not the final amount issued.

• Alphabet

🇺🇸A U.S. executive order set $21/kg polysilicon and $100/kg ingot/wafer floors, plus 15% tariffs on listed derivatives. #

President Donald Trump signed an executive order under Section 232 of the Trade Expansion Act of 1962 covering imported polysilicon and related products. The order established minimum import prices of $21 per kilogram for polysilicon and $100 per kilogram for polysilicon ingots and wafers.

The same program set minimum prices of $0.22 per watt for solar cells and $0.38 per watt for solar modules. A separate 15% ad valorem tariff applies to polysilicon ingots and other derivative products listed in the order’s annex.

The measures take effect at 12:01 a.m. U.S. Eastern Time on Dec. 4, 2026. The order also authorizes the Commerce Department to create a “return to America” incentive program under which qualifying companies that commit to U.S. production facilities and begin construction by Jan. 20, 2029, may seek Section 232 tariff exemptions for certain imported equipment and products.

🇦🇺Sharon AI (Nasdaq: SHAZ) held $1.9 billion cash after a $1.6 billion placement and $350 million convertible-notes offering. #

Sharon AI Holdings reported $1.9 billion of cash and cash equivalents as of June 30, 2026. The company said its near-term build-out was funded following an oversubscribed $1.6 billion private placement, a $350 million convertible-notes offering and accelerated receipt of $74 million from divesting Texas Critical Data Centers.

Second-quarter revenue was $1.9 million, up 412% from the corresponding 2025 period. The company reported a $430.4 million net loss, including $423.8 million of non-cash items, principally a $400.4 million fair-value loss on convertible notes caused by share-price appreciation; adjusted earnings before interest, taxes, depreciation and amortization were $0.6 million.

The filed 10-Q shows that Sharon AI generated $1,931,381 from GPU infrastructure services during the second quarter and no digital-asset-mining revenue. The company had ceased its Filecoin-related operations during 2025 to focus on its GPU-as-a-Service business, and its December 2025 business combination resulted in the current Nasdaq-listed Sharon AI Holdings structure.

• Sharon AI (Nasdaq: SHAZ)

🇺🇸Stripe entered exclusive negotiations for a cash-and-stock acquisition of OpenRouter, valuing the AI startup near $10 billion. #

Stripe was reported to be in exclusive negotiations to acquire OpenRouter, an AI startup. The proposed consideration consists of cash and stock.

The reported transaction value is approximately $10 billion. That figure is a proposed valuation associated with the negotiations, not a completed financing valuation or closed acquisition price.

No definitive agreement or transaction closing was reported in the supplied sources. The disclosed stage is therefore exclusive negotiations rather than a signed or completed acquisition.

• Stripe
• OpenRouter

🇬🇧Nscale, a UK AI-infrastructure startup, targets a September U.S. listing and claims $51 billion in contracts. #

Nscale was identified as a London-headquartered AI-infrastructure and cloud-services startup. It reportedly aims to list in the United States in September, but the supplied sources do not identify an exchange, ticker, offering size or filed registration statement.

The company claims that it has $51 billion of contracts. This is a company claim reported through Telegram and is not presented in the supplied material as an independently verified contracted-revenue figure.

Nscale’s second-quarter 2026 revenue was reported at more than $100 million, compared with $37 million in the first quarter. The source did not provide audited financial statements or further details about the composition, counterparties or recognition schedule of the claimed contracts.

• Nscale
• September U.S. listing

AMD announced plans to acquire Taalas to develop model-specific inference chips purpose-built for individual AI models. #

AMD announced that it will acquire Taalas. The stated objective is to develop future model-specific AI inference chips rather than broadly described, general-purpose accelerators.

The planned chips would be purpose-built to accelerate individual AI models. The announcement therefore concerns inference—the execution of trained models—and not model training infrastructure.

The supplied report did not disclose the purchase price, consideration structure, expected closing date or financing terms. It also did not state when the resulting model-specific chips would become commercially available.

• AMD
• Taalas

Samsung unveiled zHBM, a 3D-packaged high-bandwidth memory (HBM) design, plus two NAND offerings for solid-state drives. #

Samsung introduced zHBM at the Future of Memory and Storage Summit. The product was presented as a new form of high-bandwidth memory using three-dimensional packaging.

The announcement addressed what the source calls the AI “memory wall,” described as a major bottleneck for AI inference workloads. The product is a memory design, not a graphics processing unit or standalone AI accelerator.

Samsung also unveiled two NAND-flash offerings intended for solid-state drives. The supplied article summary does not provide their model names, capacities, interface speeds or availability dates.

• Samsung

NVIDIA tested at least three Rubin Ultra GPU variants, considering lower HBM capacity; final specifications remain undecided. #

NVIDIA has reportedly tested at least three variants of its next-generation Rubin Ultra graphics processing unit. Some test versions use less high-bandwidth memory than the configuration the company originally announced.

The Information attributed consideration of lower-memory versions partly to the possibility that NVIDIA may not obtain enough advanced HBM chips to support the original design at the required supply level. The sources state that reducing memory could affect performance and that customers running large AI models could need more GPUs than under the original configuration.

Rubin Ultra’s final specifications have not been decided. NVIDIA was reported to be planning to introduce the product around the end of 2027, but the supplied reporting characterizes the memory configuration as still under evaluation.

• NVIDIA
• HBM capacity

TSMC’s 3-nanometer capacity is expected to reach 180,000 wafer starts monthly in early fourth quarter, ahead of schedule. #

TSMC’s 3-nanometer process capacity is expected to increase to 180,000 wafer starts per month in early fourth quarter 2026. This figure refers to monthly wafer input at the fabrication stage, not finished chip output.

The expansion is expected to occur two to three months earlier than previously anticipated. The report attributed the faster schedule to strong orders from major customers including NVIDIA, AMD and Broadcom.

TSMC shares rose more than 2% alongside the report. No source provided a site-by-site allocation of the 180,000 monthly wafer starts or the proportion assigned to each customer.

SpaceX and NVIDIA are designing Starmind AI1 satellite payloads using Rubin GPUs and Vera CPUs. #

SpaceX announced on Aug. 5 that it was working with NVIDIA to develop and design Starmind AI1 satellite-computing payloads. NVIDIA also confirmed the collaboration through social media, according to the report.

Each Starmind satellite is planned to carry NVIDIA Rubin graphics processing units and Vera central processing units. SpaceX described the intended result as data-center-level computing capability in space.

The announcement concerns onboard satellite compute payloads rather than a terrestrial GPU cluster. The supplied sources do not state the number of satellites, GPU count per payload, electrical power per satellite or a deployment date.

• SpaceX
• NVIDIA
• Starmind AI1
• Rubin GPUs
• Vera CPUs

NRG neared a 1.2-gigawatt hyperscaler power deal while Texas maintained its data-center development pause. #

NRG was reported to be nearing a 1.2-gigawatt power agreement with a hyperscale customer. The cited material describes the transaction as near completion, not as a signed or operational supply contract.

The development occurred while Texas maintained a pause affecting data-center development. NRG President and CEO Robert Gaudette said the company’s customer-backed approach to adding capacity could operate in a more restrictive development environment.

The 1.2-gigawatt figure refers to electrical power capacity for the prospective hyperscaler arrangement. The supplied source does not identify the customer, generation assets, contract duration, pricing or energization schedule.

• NRG

BloombergNEF estimated Texas’s pause places 20% of the U.S. data-center pipeline at risk of delay. #

BloombergNEF estimated that the Texas pause puts 20% of the U.S. data-center development pipeline at risk of delay. The percentage concerns projects in the pipeline, not 20% of existing operational U.S. data-center capacity.

BNEF’s analysis linked the risk directly to the length of the moratorium. Its analysts said the longer the pause remains in force, the greater the risk to Texas’s data-center expansion.

The supplied source does not provide an absolute megawatt or gigawatt value for the affected 20%. It also does not state that all of the projects will be cancelled; the identified risk is delay.

• BloombergNEF

Duke Energy plans $10 billion in equity issuance while spending over $1 billion monthly to meet demand. #

Duke Energy plans to issue $10 billion of equity as it pursues generation and infrastructure spending. The source frames the equity program as part of the utility’s effort to capture a gas-generation growth opportunity.

President and CEO Harry Sideris said Duke was deploying more than $1 billion per month to meet record demand. The monthly figure describes the company’s spending pace rather than the amount of the planned equity issuance.

North Carolina advocates cited by Utility Dive said Duke had exaggerated demand-growth projections to justify its spending plan. The $10 billion remains a planned equity issuance; the supplied source does not state that the full amount has already been sold.

• Duke Energy

SpaceX plans a gas-fired power plant and large battery array for the Texas Terafab semiconductor facility. #

SpaceX plans to construct dedicated power infrastructure for Terafab, the large semiconductor manufacturing facility being developed with Tesla in Grimes County, Texas. Riley Trettel, who oversees SpaceX energy and data-center development, said the project would be self-sufficient.

The planned infrastructure includes a gas-fired power plant and what Trettel described as a very large battery array for energy storage. The source does not provide the generating capacity, storage capacity, equipment suppliers or construction timetable.

Tesla already produces utility-scale Megapack battery systems and operates a Megapack factory west of Houston. Elon Musk separately gave a rough estimate that approximately 25% of Terafab’s AI-compute output would serve Tesla Optimus and about 75% would support AI spacecraft, explicitly characterizing that allocation as approximate.

• SpaceX
• Terafab

Germany’s grid regulator proposed one-time renewable interconnection fees and annual €4–€7/kW charges beginning in 2029. #

Germany’s Federal Network Agency published a draft reform of the country’s grid-financing system. The proposal would require wind and solar generators to pay grid charges for the first time.

Renewable projects would face a one-time interconnection fee and, from 2029, annual charges of €4 to €7 per kilowatt. The unit is euros per kilowatt of capacity, not a charge per kilowatt-hour of generated electricity.

The draft is intended to distribute grid costs more evenly and better align electricity supply and demand. The regulator also linked the proposal to persistent grid congestion and efforts to reduce electricity costs for industrial and commercial users.

Sharon AI reported 212 megawatts of secured capacity and a six-year, $4.9 billion collaboration with NVIDIA for up to 40,000 GB300 GPUs. #

Sharon AI disclosed that secured AI Factory capacity increased by 80 megawatts to 212 megawatts. The company also reported a six-year strategic compute collaboration with NVIDIA valued at $4.9 billion and covering up to 40,000 NVIDIA GB300 GPUs.

Additional contracts disclosed by management included a five-year, $950 million take-or-pay agreement with a global technology company having a major Asia-Pacific presence. Sharon AI also reported a five-year, $373 million take-or-pay contract with a global AI platform for a deployment of 2,048 NVIDIA B300 GPUs.

The company said total contract value was approximately $8.8 billion as of Aug. 6, 2026. It also said more than 64,000 NVIDIA GPUs were expected to be deployed by mid-2027 and that a 600-petabyte VAST Data AI Operating System deployment was sized to support approximately 100,000 GPUs; these deployment figures are company expectations rather than completed installations.

The information was included in a company press release filed with the U.S. Securities and Exchange Commission. The filing proves that Sharon AI disclosed the contracts, capacity and forecasts, but does not independently verify future deployment or revenue recognition.

• Sharon AI
• NVIDIA

Mirendil is live on TPU v5P accelerators; NVIDIA accelerated systems will support pre-training and post-training next. #

Mirendil selected Google Cloud’s AI Hypercomputer infrastructure for model pre-training and post-training. The environment combines Google tensor processing units—AI accelerators known as TPUs—with NVIDIA’s full-stack accelerated-computing platform.

Mirendil is already operating a cluster of TPU v5P chips, while NVIDIA AI accelerated-computing systems are scheduled to come online subsequently. Google Cloud did not disclose the number of TPU chips or NVIDIA GPUs in its announcement.

Google and Mirendil jointly designed the compute, storage, networking and control-plane environment. They also developed a system using managed training clusters in the Gemini Enterprise Agent Platform to provision and manage both TPU and GPU environments.

Mirendil said its work includes end-to-end training workflows, from initial model pre-training through post-training, as well as reinforcement learning. A separate Telegram report characterized the Google Cloud agreement as worth upward of $100 million.

• Mirendil

Amazon Bedrock AgentCore runtime instances provide persistent managed EC2 compute, GPU support, and sessions lasting up to 14 days. #

Amazon Web Services announced runtime instances for Amazon Bedrock AgentCore. The service provides persistent, managed Amazon Elastic Compute Cloud infrastructure for production AI agents.

The instances support graphics processing units and multi-agent collaboration. Runtime sessions can remain active for as long as 14 days, distinguishing the service from short-lived invocation environments.

AWS described the instances as persistent compute for production deployments. The supplied announcement does not state supported GPU models, instance sizes, regional availability or pricing.

• EC2 compute
• GPU support

Google Cloud’s MedPerf uses A3 machines with NVIDIA H100 GPUs inside hardware-isolated trusted execution environments for private medical-AI evaluation. #

Google Cloud is collaborating with MLCommons on MedPerf, an open-source platform launched in 2023 to standardize medical-AI evaluation. MedPerf uses Google Cloud Confidential Space to create a secure environment for testing proprietary models against real-world patient data.

The workloads run inside hardware-isolated trusted execution environments, or TEEs. These special virtual machines encrypt memory while it is in use and harden the operating system so the hospital, research institution, other participants and Google cannot view the model code or patient data during evaluation.

For GPU-accelerated inference, MedPerf uses Google Cloud A3 machines with NVIDIA H100 GPUs. The configuration combines Intel Trust Domain Extensions on the central processing unit with NVIDIA Confidential Computing on the GPU, extending protection beyond CPU memory to model weights and patient data processed by the accelerator.

Before patient data enters the workload, the system provides cryptographic proof that approved code is running on genuine confidential-computing hardware in a properly hardened environment. The technology is being used by the Federated Tumor Segmentation initiative to validate models against private brain magnetic-resonance-imaging data from multiple locations.

• Google Cloud’s MedPerf
• A3 machines
• NVIDIA H100 GPUs

OpenAI made GPT-5.6 Luna ChatGPT’s default model, giving free users unlimited text chat and a Think button. #

OpenAI said it would switch ChatGPT’s default model for all users to GPT-5.6 Luna during the week. Compared with the previous default, GPT-5.5 Instant, Luna reportedly produced responses containing at least one factual error 62% less often.

Free and Go-plan users receive unlimited text chat. They also gain a new Think button that lets them select a higher level of reasoning for more complex questions.

Plus and Pro users receive an updated GPT-5.6 Sol model designed to improve factual reliability and provide more focused answers. OpenAI said Sol uses more compact formatting, removes unnecessary detail and maintains a more consistent tone across everyday conversation, research, advice, planning, writing and decision tasks.

Plus and Pro accounts also receive a slider controlling how much reasoning the model applies to an answer. OpenAI said the updated Sol model and slider became available to those subscribers on Aug. 6.

• OpenAI
• GPT-5.6 Luna
• ChatGPT
• Think button

Disney began limited AI-search tests on ESPN and Disney+, supporting natural-language questions, recommendations, text, and voice. #

Disney began beta testing AI-powered search and content-discovery functions on ESPN and Disney+. The tests are limited rather than a full release across all users.

ESPN Search lets selected users ask sports questions in natural language. It generates answers, related content, statistics and recommendations using ESPN articles, videos, research data, internal knowledge bases and sports-data systems.

Disney+ is separately testing an AI content-discovery tool with a small user group. Users can describe what they want to watch through text, voice or predefined prompts, and the system recommends content based on their current interests rather than relying only on prior viewing history.

The sources do not identify the underlying AI models, cloud infrastructure or test-group size. They also do not provide a schedule for broader availability.

• Disney
• ESPN
• Disney+

Apache Fluss became a top-level Apache project, providing lakehouse-native streaming storage for real-time analytics and AI. #

The Apache Software Foundation announced Apache Fluss as a new top-level project. Fluss provides lakehouse-native streaming storage for real-time analytics and artificial-intelligence workloads.

Its designation as a top-level Apache project places it directly within the Apache Software Foundation’s project structure. The supplied announcement does not provide benchmark results, deployment counts or a commercial pricing model.

The foundation simultaneously announced Apache Pony Mail as another top-level project. Pony Mail is a web-based mail-archive browser designed to scale to millions of archived messages, while Fluss is the storage project directed at streaming analytics and AI.

• Apache Fluss
• Apache