AI & Compute Infrastructure — August 19, 2026

🇺🇸Commodity Futures Trading Commission (CFTC) opened public consultation on compute futures; CME, ICE, and Architect plan contracts subject to approval. #

The U.S. Commodity Futures Trading Commission (CFTC) began seeking public input on futures contracts tied to computing capacity. The consultation asks how compute differs from commodities and derivatives already regulated by the agency.

CME Group, Intercontinental Exchange (ICE), and Architect Financial Technologies have announced plans to offer compute-related contracts if they receive regulatory approval. The proposed market is intended to let end users and other participants hedge risks such as energy shortages that affect available AI computing capacity.

The consultation also addresses how variables affecting compute prices should be standardized. One identified issue is the construction of the reference-price indexes that would be used to settle compute futures contracts.

CFTC Chairman Michael Selig described the consultation as a first step toward establishing clearer rules for a U.S. compute market. No source provided a launch date or confirmed regulatory approval for any of the planned contracts.

• CME
• ICE
• Architect

🇺🇸Datavault AI’s filed agreement conditions a $120 million upfront payment on closing to fund 48,000 NVIDIA H200 GPUs. #

Datavault AI Inc. filed an executed term sheet under which Scilex Holding Company would contribute $120 million in cash at the transaction’s closing. The filing describes the payment as conditional on closing, rather than as financing that had already been received.

Datavault must use the money exclusively for its proposed “Quantum-Ready Edge Network” across an estimated 100 U.S. cities. Eligible uses include purchasing 48,000 NVIDIA H200 graphics processing units (GPUs), build-out, equipment, directly attributable working capital, and reasonable overhead.

The agreement says Available Infrastructures has 24,000 H200 GPUs in its warehouse and that another 24,000 would be purchased at a pre-fixed per-GPU price. It attributes an approximately $2.4 billion current market value for all 48,000 GPUs and projected annual revenue potential of $10 billion to $100 billion to Available Infrastructures; these are filed third-party estimates, not reported Datavault operating results.

The payment is to be secured by the GPUs in stock and those to be purchased until Scilex has received $180 million of the specified payments. Datavault would pay Scilex 30% of network gross revenue until payments reach $250 million, then 15% until cumulative payments reach $1.2 billion, followed by 5% of network revenue for the network’s lifetime.

• Datavault AI’s filed agreement

🇺🇸ChronoScale (CHRN) classified its legacy exoskeleton business as held for sale, targeting fiscal 2027 divestiture to focus on GPU cloud. #

ChronoScale Holdings Corp. reported that its board concluded on May 29, 2026, that the Legacy Ekso business met the accounting criteria to be classified as held for sale. The board committed to a divestiture plan intended to leave ChronoScale focused solely on its cloud business.

Legacy Ekso designs, develops, and markets robotic exoskeletons and related services for rehabilitation, personal mobility, and industrial applications. Most of that unit’s sales have come from enterprise healthcare products used for neurological rehabilitation in clinical settings.

ChronoScale expects to complete the divestiture during fiscal 2027, but the filing does not report a completed sale or identify a buyer. Its cloud operation currently supplies GPU computing for artificial intelligence, machine learning, rendering, and other high-performance computing workloads from third-party colocation facilities.

Management also described a planned four-layer cloud platform spanning infrastructure, compute, AI-platform capabilities, and implementation services. The filing explicitly says the deployment and expansion of these layers are forward-looking and were not reflected in operating results for the fiscal year ended May 31, 2026.

• ChronoScale (CHRN)

🇺🇸EigenQ and Silicon Valley Acquisition Corp. confidentially submitted a draft Form S-4 for their proposed quantum-technology merger. #

EigenQ, Inc., described as a quantum-technology company, and Silicon Valley Acquisition Corp. announced the confidential submission of a draft registration statement on Form S-4 to the U.S. Securities and Exchange Commission. Silicon Valley Acquisition Corp. is a publicly traded special purpose acquisition company.

A Form S-4 is the registration document associated here with the companies’ proposed business combination. The submission is a draft under SEC review and does not mean the registration statement has become effective or that the merger has closed.

The source describes the transaction as a proposed merger intended to take EigenQ’s quantum-technology business public. It does not provide a completed transaction value, closing date, final exchange ratio, or confirmation that SEC review has been completed.

• EigenQ
• Silicon Valley Acquisition Corp
• Form S-4

🇯🇵Japan’s July exports increased 23.2%, while semiconductor-equipment exports rose 49.1% year over year. #

Japan’s exports increased 23.2% year over year in July, exceeding the reported 19.9% consensus estimate and marking a fifth consecutive acceleration. The reported growth rate was the fastest since October 2022.

Semiconductor-equipment export value rose 49.1% from a year earlier. Exports to the United States increased 22%, exports to the European Union rose 19.1%, exports to Asia increased 24.5%, and exports to China rose 25.8%.

Imports grew 27.8% year over year, above the reported 26.5% estimate and at the fastest rate since November 2022. Petroleum import value increased 87.8%, while crude-oil import volume rose 5.5% and liquefied natural gas imports fell 4.8% to 5.014 million metric tons.

Japan recorded an unadjusted trade deficit of ¥634.5 billion, compared with the reported ¥680 billion deficit estimate and a revised ¥409.9 billion deficit in June. The seasonally adjusted goods-trade balance was a deficit of ¥686.009 billion.

• Japan’s July exports

Fractile reached a preliminary agreement to sell $250 million of AI inference chips to Anthropic for 2027 use. #

Fractile and Anthropic reached a preliminary agreement under which Fractile would sell approximately $250 million of AI chips to Anthropic. The chips are expected to enter use in 2027, and the parties reportedly plan to expand the scale of their relationship.

Fractile was founded in 2022 and focuses on AI inference chips, which execute trained models to generate answers and perform other model-serving tasks. The source distinguishes these products from chips primarily used to train AI models.

The chip agreement is preliminary, not described as a completed sale or recognized revenue. No deployment volume, per-chip price, delivery schedule, or final contract conditions were provided.

Separately, Fractile was reported to be in advanced financing discussions to raise approximately $600 million at a $6.5 billion pre-money valuation. Three months earlier, it had completed a $220 million financing led by Accel, Founders Fund, and Factorial Funds at an approximately $1 billion post-money valuation.

• Fractile
• Anthropic

TFC Optical Communication’s co-packaged-optics products entered mass production while it developed higher-density, higher-channel-count optical connections with customers. #

TFC Optical Communication said supporting products for co-packaged optics (CPO) had entered mass-production delivery. It reported that product yields were sufficient for scaled manufacturing and that it was preparing capacity around customer demand forecasts.

Products currently being delivered include fiber-array units (FAUs) for CPO applications and external laser source (ELS) products. The company cited its optical-coupling, precision-manufacturing, and volume-delivery capabilities as the basis for product performance, yield, and delivery support.

Its research and development team is working with customers on next-generation optical connections with higher density and more channels. The source does not quantify production volume, customer commitments, channel counts, or revenue from these CPO-related products.

TFC also said shortages of certain materials had constrained increases in active optical-device production during the first half. It added new suppliers, and reported that the shortages had gradually eased from the third quarter of 2026, with overall supply conditions improving.

Dier Laser’s through-glass-via equipment continued generating orders and revenue; laser-modification and etching systems received bundled orders. #

Dier Laser (帝尔激光) said its laser-modification equipment for through-glass vias (TGVs) continued to receive orders and generate revenue. It also reported that customers had begun placing bundled orders covering both laser-modification and etching equipment.

The company is additionally developing etching and automated optical inspection (AOI) equipment around its TGV product line. AOI equipment remained in active customer validation, while the laser-modification and etching systems had progressed to bundled orders.

Dier Laser said its TGV equipment covers both wafer-level and panel-level applications, and some glass-substrate customers had placed repeat orders. It identified current applications including glass-substrate vias for advanced packaging, fiber-array units, and passive glass components for optical communications.

The company described TGV and related applications as still being in the industrial-introduction stage. It said remaining glass-substrate production difficulties were concentrated more in downstream metallization, film and material-combination yields, and microcracking than in its laser-modification process, which it said met customers’ yield and technical requirements.

• Dier Laser’s through-glass-via equipment

IBM joined and cooled two cryogenic modules within one environment, targeting modular systems linking hundreds of quantum chips. #

IBM announced that it had successfully joined and cooled two cryogenic modules in a single environment. Cryogenic modules maintain the ultra-low temperatures needed for the company’s quantum-computing hardware.

The architecture is designed around modular, shared, ultra-cold infrastructure rather than treating each module as an isolated system. IBM’s stated objective is to link hundreds of quantum chips through this modular approach.

The work is directed toward a more powerful quantum computer capable of addressing large workloads and supporting fault-tolerant quantum computing. The source reports the two-module cooling result but does not state that a system containing hundreds of connected quantum chips has already been completed.

• IBM

Itochu plans roughly 10 Japanese data centers by 2030, each providing approximately 50 megawatts of power capacity. #

Itochu plans to enter data-center development by investing several hundred billion yen in approximately 10 Japanese facilities by 2030. Each planned data center is expected to provide around 50 megawatts of power capacity.

The proposed facilities would be located in the Tokyo metropolitan area, the Osaka region, Kyushu, and other Japanese markets. Itochu plans to bring one or two projects into operation annually and lease them to customers including U.S. technology companies.

Itochu would acquire the land and oversee construction, treating the projects as part of its real-estate business. It ultimately plans to sell the completed data centers to third parties to improve asset-use efficiency.

The company may also work with East Japan Railway Company (JR East), which owns land, operates power-generation assets, and has relationships with electric utilities. The cited report estimates Japan’s data-center services market at ¥5.65 trillion in 2030, approximately 30% above its 2025 level.

• Itochu

Hyperscale Data recorded a $2.3 million impairment while reallocating Michigan data-center power from Bitcoin mining to AI and colocation. #

Hyperscale Data Inc. (NYSE American: GPUS) recorded a $2.3 million impairment charge for Bitcoin-mining equipment at its Michigan data center during the three and six months ended June 30, 2026. The accounting review followed execution of a master services agreement with a third-party customer for AI compute, neocloud hosting, and colocation services.

The company expects to allocate a significant portion of the Michigan facility’s available electrical capacity to that customer’s deployments. As a result, it expects to substantially wind down Bitcoin mining at the location and redirect power capacity toward AI compute, neocloud, and colocation.

The filing says the impairment did not reflect physical deterioration of the mining equipment. It arose because of the expected reduction in use and uncertainty over whether the affected equipment would be redeployed, relocated, or sold.

Management concluded that the asset group’s carrying amount was not recoverable and reduced it to estimated fair value. Fair value was based primarily on observable market data for comparable Bitcoin-mining equipment, using an orderly-liquidation-value approach.

• Hyperscale Data

International Electrotechnical Commission (IEC) approved a proposal for the first international standard covering data-center microgrid monitoring and energy-management systems. #

The International Electrotechnical Commission (IEC) approved a project proposal titled Technical Specification for Data Center Microgrid Monitoring and Energy Management Systems (《数据中心微电网监控与能量管理系统技术规范》). The source describes it as the first IEC international-standard project focused on coordination between computing capacity and electricity systems.

The project will address joint monitoring and energy management for computing and power infrastructure in data-center microgrids. Its stated purpose is to provide technical guidance for an emerging operating environment that previously lacked a unified standard.

A Chinese expert will serve as project leader, with experts from France, Germany, Spain, and other countries participating. Approval of the proposal establishes the standard-development project; it does not mean a final international standard has already been published.

• International Electrotechnical Commission (IEC)

Compute-Power Coordination Technology Joint Research Institute (算电协同技术联合研究院) opened in Guangzhou, targeting green-power interconnection, direct-current distribution, storage, and operations. #

The Compute-Power Coordination Technology Joint Research Institute (算电协同技术联合研究院) was inaugurated in Guangzhou on August 19. It was jointly established by the National Innovation Center for Advanced Energy Storage, China Southern Power Grid Research Institute, China Southern Power Grid Power Technology, Guangdong HEC Technology Holding, and Chindata.

The institute is focused on data-center energy systems and power distribution. Its stated operating priorities are high reliability, low energy consumption, green-electricity use, and intelligent autonomous operation.

The five participants signed a cooperation agreement covering four research areas. These are direct green-power connections and source-grid-load-storage coordination, advanced direct-current distribution and power-quality optimization, new energy-storage devices and integrated equipment, and intelligent operations and reliability management for computing centers.

The organization aims to become a domestic platform for compute-power coordination research, commercialization of technical results, and professional training. The source reports the institute’s establishment and research agenda but does not provide a funding amount or deployment schedule.

Electric Reliability Council of Texas (ERCOT) issued an extreme-weather notice from Thursday through Tuesday as extreme heat was forecast. #

The Electric Reliability Council of Texas (ERCOT) announced an extreme-weather event notice covering Thursday through the following Tuesday. ERCOT attributed the notice to forecasts of extreme heat.

The forecast specifically identified the north-central and south-central regions. The source did not provide a projected peak-load figure, reserve margin, or expected generation shortfall.

The announcement was an extreme-weather notice rather than a reported order for customer load shedding. No service interruption or emergency curtailment was identified in the supplied source.

ChronoScale deployed 6,144 NVIDIA H100 GPUs across Colorado, Minnesota, and Utah, all dedicated to customer Together AI. #

ChronoScale reported 6,144 deployed NVIDIA H100 graphics processing units as of May 31, 2026. The GPUs operate in multiple clusters across Colorado, Minnesota, and Utah.

ChronoScale does not own the underlying data-center properties; it rents space in third-party colocation facilities and installs company-owned computing equipment. It has secured colocation-provider contracts for space and energy supporting the cloud operation.

All of the currently deployed GPU capacity is provided to Together AI under a master terms-of-service agreement. Together AI receives dedicated access, and ChronoScale charges a fixed per-GPU, per-hour rate billed and paid monthly, although the filing says those usage fees are subject to change.

ChronoScale reported additional deployable capacity at each of its three locations beyond the capacity currently contracted. Management said it intended to offer that capacity through dedicated GPU-as-a-Service and a planned usage-based managed-inference offering called Token Factory, but those statements concern intended future services.

• ChronoScale
• Together AI

Together AI generated 99.5% of ChronoScale’s fiscal-year revenue; its renewed contract runs 12 months, then renews every 60 days. #

ChronoScale’s filing says its cloud business accounted for approximately 99.5% of total revenue for the fiscal year ended May 31, 2026. Together AI was the cloud business’s only customer at that date and used GPU capacity across all three ChronoScale colocation sites.

The master terms-of-service agreement was originally entered into in December 2023 and most recently renewed effective March 1, 2026. Its renewed initial term is 12 months, after which it automatically renews for successive 60-day periods unless either party gives the required advance termination notice.

Under the agreement, Together AI receives access to dedicated NVIDIA H100 GPU infrastructure across clusters totaling approximately 6,144 GPUs. Services are priced at fixed per-GPU, per-hour rates and are billed and paid monthly, with the filing noting that these fees are subject to change.

ChronoScale’s filing therefore documents material reliance on one customer for its reported cloud operations. Management discussed plans to market additional capacity to other customers, but the filing did not report that this prospective diversification had already occurred as of May 31.

• Together AI
• ChronoScale

Biren Technology’s Wenzhou developer-cloud cluster has operated for months, offering standardized integrated training-and-inference services through China Unicom partners. #

Biren Technology (壁仞科技) said its developer-cloud intelligent-computing cluster in Wenzhou had already been running for several months and was operating well. The cluster was formally presented at an August 18 event titled “Chips Gather at the Oujiang, Computing Starts the Future” (“芯聚瓯江·算启未来”).

China Unicom’s Zhejiang and Wenzhou branches, ZTE, Youyun Technology, and Biren jointly built the cluster. The deployment uses Biren’s domestically developed general-purpose GPU computing foundation together with carrier data-center and network resources.

Biren and China Unicom Wenzhou also signed a strategic cooperation agreement. They plan to combine the GPU infrastructure, operator network and machine-room resources, and cloud-platform operations capabilities to provide standardized integrated AI training-and-inference computing services.

The target users are AI companies, research institutions, and developers. The source does not provide the number or model of accelerators, cluster performance, power capacity, service pricing, or contracted customer volume.

• China Unicom

xAI made Grok 4.6 generally available on Amazon Bedrock for cloud customers. #

xAI announced that Grok 4.6 had reached general availability on Amazon Bedrock. The status indicates that the model is available through Amazon’s managed cloud AI service rather than remaining in a limited preview.

The announcement concerns access to the Grok 4.6 model through Bedrock. It does not identify the underlying accelerator type, GPU count, data-center location, or infrastructure capacity used to serve the model.

The supplied source also does not state pricing, supported context length, performance measurements, regional availability, or service-level terms. No separate hardware deployment or infrastructure transaction was announced in the cited message.

• Grok 4.6
• Amazon Bedrock

OpenAI previewed Private Safety Processing with zero-data-retention compatibility; early customers include Microsoft and Databricks, with a September rollout planned. #

OpenAI previewed Private Safety Processing, a system intended to strengthen safeguards while remaining compatible with zero data retention. It is designed to detect risk patterns spread across multiple interactions rather than examining only an isolated prompt and response.

OpenAI said longer and more autonomous AI workflows can produce risks through combinations of interactions. The system is intended to protect against attacks by malicious users and misaligned AI agents without giving OpenAI personnel access to the underlying raw customer content.

The architecture does not retain customer prompts or model responses, and OpenAI personnel cannot view them. Enterprise customer data will not be used for model training unless the customer explicitly opts in.

Microsoft and Databricks were identified as early customers participating in testing. OpenAI expects a broader rollout in September and plans to publish a technical white paper alongside the release.

• OpenAI
• Private Safety Processing
• Microsoft
• Databricks

SandboxAQ made AQPotency generally available on Claude through Model Context Protocol for scoring candidate-molecule potency against disease targets. #

SandboxAQ announced the general availability of AQPotency on Anthropic’s Claude through Model Context Protocol (MCP). MCP provides the connection through which Claude can invoke the AQPotency capability.

SandboxAQ describes AQPotency as a Large Quantitative Model (LQM) for predicting how well a potential drug may work. For a specified disease target, it scores how strongly candidate molecules are likely to act on that target.

The announcement concerns model access for drug-discovery workflows, not a new computing cluster or chip deployment. The supplied source does not provide benchmark accuracy, training-compute requirements, pricing, or the number of molecules that can be evaluated per unit of time.

• SandboxAQ
• AQPotency
• Claude
• Model Context Protocol

MiniMax launched MiniMax Design, using its H3 multimodal video model for task decomposition, model invocation, generation, editing, and delivery. #

MiniMax (稀宇科技) launched MiniMax Design, which it describes as a harness that turns multimodal-model capabilities into a production workflow. A user provides a creative request, after which the system interprets the goal and decomposes it into tasks.

MiniMax Design can invoke relevant models and skills, process source materials, generate content, edit the results, and produce the final deliverable. The workflow is therefore presented as an orchestration layer rather than only a standalone content-generation model.

MiniMax’s native multimodal video model H3 supplies the underlying capabilities for the product. The company said H3 had demonstrated multimodal understanding, video generation, video editing, and reference-based control.

MiniMax identified video-editing tasks as a particularly relevant use case because they require understanding, control, and modification of existing material. The source does not provide pricing, model size, context specifications, accelerator requirements, or service-performance metrics.

• MiniMax
• MiniMax Design