AI & Compute Infrastructure — August 18, 2026

🇺🇸Velaura AI raised $110 million in Series A financing, valuing the low-power AI-compute developer above $1 billion. #

Velaura AI announced that it raised $110 million in Series A financing and that the round brought its valuation to more than $1 billion. The company describes itself as an AI-compute infrastructure developer working on ultra-low-power silicon and software technologies.

The announcement cited a “proven technology foundation” and multiple engagements with hyperscale cloud operators as factors behind investor interest. The supplied source does not identify the hyperscalers or provide technical specifications, performance measurements, deployment volumes, or commercial-contract values.

This is described as a completed Series A financing rather than a proposed facility or fundraising target. The source does not disclose participating investors, ownership dilution, security terms, liquidation preferences, board rights, or the planned allocation of the proceeds.

• Velaura AI
• Series A

🇺🇸Anthropic’s proposed revolving credit facility may exceed $10 billion; lead banks were asked to provide $1.25 billion each. #

Anthropic’s planned revolving credit facility is expected to exceed its approximately $10 billion target, according to reports citing people familiar with the financing. The facility was still being arranged, so the reported size represents an expected credit commitment rather than funded borrowing or cash already drawn by Anthropic.

The developer of Claude reportedly asked the most active lead banks to provide approximately $1.25 billion of commitments each. Banks in the next participation tier were encouraged to provide approximately $1 billion each, while less-active participants were expected to commit around $750 million or less.

Banks were reportedly competing for positions in the facility partly because higher participation and syndicate ranking can lead to larger fees and potentially more prominent roles in later capital-markets transactions. The reports connected that competition with Anthropic’s preparations for an initial public offering, but did not provide an IPO filing date, facility interest rate, maturity, collateral, covenants, or amount drawn.

🇨🇳ByteDance’s jumbo bank loan attracted more than $30 billion in lender orders. #

ByteDance (字节跳动) received more than $30 billion of lender orders for a jumbo bank loan, according to a report citing sources. The figure measures lender demand or orders and does not establish how much ByteDance ultimately borrowed.

The supplied report does not state the loan’s final size, currency, maturity, pricing, intended use, collateral, covenants, or syndicate composition. It also does not specify whether the lender orders were binding commitments or preliminary indications.

The update concerns private bank financing rather than an equity financing, valuation event, or public securities offering. No information was provided on dilution because a bank loan does not, based on the supplied report, involve an announced issuance of ByteDance shares.

🇺🇸Thirty-year Treasury yields reached their highest since 2007, while monthly federal interest costs remained above $100 billion. #

The 30-year U.S. Treasury yield reached its highest level since 2007 during the week. One market update reported that the yield touched 5.34% intraday before partially retreating, amid elevated oil prices, geopolitical concerns, inflation worries, and government-debt concerns.

Federal interest spending increased from $76 billion in January to $107 billion in May. The reported monthly total subsequently remained above $100 billion, linking the rise in long-term borrowing costs with an already elevated federal interest-expense burden.

The long-duration bond selloff was not limited to the United States. Reports said 30-year French yields reached their highest since 2008, German yields rose to levels associated with 2011, U.K. long-term yields approached 6%, and comparable Japanese yields were near historical highs.

• Thirty-year Treasury yields

Cerebras CS-4 combines three WSE-3 Turbo wafer-scale engines, 750 PFLOPS, and 129.6 PB/s, with third-quarter shipments planned. #

Cerebras announced the CS-4 as a new AI accelerator system built with three WSE-3 Turbo wafer-scale engines. The company specified aggregate AI compute performance of 750 petaflops, or PFLOPS, and memory bandwidth of 129.6 petabytes per second, or PB/s.

A wafer-scale engine is a large processor implemented across substantially more silicon area than a conventional individually packaged chip. The report identifies the WSE-3 Turbo as the processing engine inside CS-4 but does not provide power consumption, memory capacity, system price, manufacturing yield, or rack-level deployment details.

Cerebras expects CS-4 shipments to begin in the third quarter of 2026. That is a planned shipment schedule; the supplied report does not state that commercial-volume shipments have already begun or identify customers receiving the first systems.

• Cerebras CS-4

Cerebras claims CS-4 exceeded 4,400 tokens per user-second on GPT-OSS-120B and reached 30 times GPU-based inference speed. #

Cerebras said CS-4 generated more than 4,400 tokens per user per second in a GPT-OSS-120B test. The company also claimed performance as high as 30 times that of graphics processing unit-based inference systems.

These are vendor-reported benchmark claims from Cerebras. The supplied report does not identify the exact GPU system, batch size, precision, latency threshold, software stack, power envelope, or other test conditions needed for an independent like-for-like comparison.

Cerebras separately said CS-4 can deliver up to 10 times the throughput per watt of CS-3. It also said the system can support models with more than 50 trillion parameters, although the source does not describe the model architecture, partitioning method, or memory configuration behind that capacity claim.

• Cerebras
• CS-4
• GPT-OSS-120B

Avicena began shipping 1-terabit-per-second LightBundle evaluation kits supporting up to 335 microLED channels operating at 3 gigabits per second. #

Avicena began shipping 1-terabit-per-second LightBundle evaluation kits to customers and partners developing next-generation AI infrastructure. These are evaluation kits intended for technical assessment, not an announcement of full-volume production deployment.

The LightBundle kit supports as many as 335 microLED optical channels, with each channel operating at 3 gigabits per second. The design therefore uses a large number of comparatively low-speed parallel optical channels to provide terabit-class connectivity.

The shipment follows Avicena’s earlier introduction and demonstration of the technology. The supplied announcement does not provide kit pricing, link distance, power consumption, error rates, packaging dimensions, customer identities, or a volume-production timetable.

• Avicena

Coherent began customer sampling of 300-millimeter high-thermal-conductivity silicon carbide substrates for AI semiconductor thermal-management applications. #

Coherent began sampling 300-millimeter high-thermal-conductivity silicon carbide, or SiC, substrates to leading AI-semiconductor partners. The announced stage is customer sampling rather than confirmed high-volume production or completed customer qualification.

The substrates are intended to address thermal-management requirements in artificial intelligence and high-performance computing semiconductor applications. Coherent characterized the development as an expansion of its scalable materials platform.

The 300-millimeter figure refers to substrate diameter, not semiconductor process-node size or data-center capacity. The source does not disclose customer names, thermal-conductivity measurements, wafer pricing, production capacity, qualification schedules, or commercial shipment dates.

• Coherent

MLPerf Client v2.0 added image-generation, agentic-AI, and updated large-language-model tests for benchmarking AI PCs. #

MLCommons released MLPerf Client v2.0, expanding its benchmark suite for measuring AI performance on personal computers. MLCommons is an open engineering consortium focused on machine-learning performance and transparency.

The release adds benchmark categories for image generation and agentic-AI workflows. It also updates the suite’s large language model tests, broadening coverage beyond the AI workloads included in the previous client benchmark.

The update is a benchmarking release, not a ranking of a particular central processing unit, graphics processing unit, or neural processing unit. The supplied announcement does not include vendor scores, system configurations, power results, or comparisons among specific AI PCs.

• MLPerf Client v2.0

Northwest Power and Conservation Council proposed 11 gigawatts of generation and 5 gigawatts of storage by 2032, with $2.3 billion in fixed costs. #

The Northwest Power and Conservation Council proposed adding 11 gigawatts of generation and 5 gigawatts of storage by 2032. The portfolio forms part of the council’s proposed Ninth Power Plan.

The council estimated that the portfolio would carry $2.3 billion in fixed costs in 2032. The supplied source does not provide a technology-by-technology breakdown of the generation, storage duration, energy output, transmission requirements, or treatment of variable operating and fuel costs.

The figures describe a proposed regional planning portfolio, not generation and storage capacity already operating or under construction. The source also does not state that all 16 gigawatts have secured permits, financing, interconnection agreements, or final investment approval.

TerraPower plans to announce a second nuclear plant serving a data center this year, targeting a 2027 groundbreaking. #

TerraPower expects to announce its next nuclear project during 2026, with the plant planned to serve a data center. Chief Executive Officer Chris Levesque said the company is targeting a 2027 groundbreaking for the second project.

The Bill Gates-backed company was described as the only U.S. company currently building a utility-scale nuclear power plant. The planned data-center project would be separate from that first plant and remains at the pre-announcement stage described by management.

Levesque declined to identify the data-center customer. The source also does not provide the project’s location, generating capacity, reactor configuration, contract structure, power-delivery schedule, construction cost, financing, or regulatory status.

• TerraPower

Pennsylvania ordered data centers to obtain local approval, meet water standards, fund added electricity costs, and secure power. #

Pennsylvania Governor Josh Shapiro issued an executive order imposing new requirements on data-center development. Projects must obtain local-government approval before seeking state permits and must comply with water-conservation standards.

Developers are also required to cover the incremental electricity costs created by their projects and address their own power-supply requirements. The measures distinguish responsibility for added data-center load from general electricity costs borne by other users.

Shapiro said numerous data-center projects were entering Pennsylvania and that some developers were disregarding community interests. The report also noted that New York had pursued a suspension of up to one year for environmental permitting of large data centers, while Texas had paused approvals for some new projects and initiated a review.

• Pennsylvania

State Grid and China Southern Power Grid signed a framework covering supply security, major projects, markets, and innovation. #

State Grid and China Southern Power Grid signed the Strategic Cooperation Framework Agreement on Implementing the 15th Five-Year Plan and Accelerating New Power-Grid Construction (《落实“十五五”规划 加快新型电网建设战略合作框架协议》). The agreement followed an August 18 meeting between the two grid operators.

The framework covers joint work on electricity-supply security and development of a new power system (新型电力系统). It also includes major power projects and broader construction of a new power grid (新型电网).

The companies additionally agreed to advance development of the national unified electricity market (全国统一电力市场) and energy-and-power technology innovation. The source describes a cooperation framework and does not disclose project-level investment, construction capacity, schedules, procurement awards, or binding power-delivery commitments.

• State Grid
• China Southern Power Grid

Amazon plans to raise its Louisiana data-center investment from $12 billion to $18 billion, adding a third campus. #

Amazon announced plans to increase its Louisiana data-center investment from $12 billion to $18 billion. The expansion includes construction of a third data-center campus in the state.

The reported figures describe Amazon’s planned investment rather than capital expenditure already completed. The source does not break the total down among land, buildings, servers, networking equipment, power infrastructure, cooling systems, or workforce costs.

No capacity figures were provided for the third campus. The source also does not state its location, power demand, number or type of accelerators, construction schedule, expected completion date, or electricity-supply arrangements.

• Amazon

China Unicom reported intelligent-computing capacity above 45 EFLOPS while continuing to optimize its data-center footprint. #

China Unicom reported intelligent-computing capacity of more than 45 exa floating-point operations per second, or EFLOPS, during its half-year results briefing. The source does not state the numerical precision used for that capacity figure or identify the accelerator and server mix behind it.

The operator said it was continuing to optimize its data-center layout as part of its computing-network strategy. It also said it was working with China Telecom to build what it described as the largest jointly constructed and shared 5G network.

Management separately said China’s traditional communications market had entered a period of competition for existing users. China Unicom is therefore reducing the number of mobile-service packages on sale and simplifying tariffs while focusing on maintaining and increasing customer value.

• China Unicom

Anhui will coordinate quantum, general, intelligent, and supercomputing resources while offering small businesses elastic compute and AI compute vouchers. #

The Anhui provincial government issued implementation guidance calling for coordinated development of computing clusters and edge-computing resources. It plans a unified scheduling platform integrating quantum computing (量算), general-purpose computing (通算), intelligent computing (智算), and supercomputing (超算).

The policy supports elastic computing capacity and token-related services for small and medium-sized enterprises. It also calls for implementation of an AI computing-voucher policy, which would provide policy-backed access to compute rather than announce a specific new accelerator deployment.

Anhui additionally plans a quantum–artificial intelligence cross-disciplinary innovation center (量子人工智能交叉融合创新中心). The source does not specify the platform’s aggregate capacity, voucher budget, eligible hardware, participating data centers, launch schedule, or allocation mechanism.

• Anhui

Google’s Governance Agent project uses BigQuery column-level lineage to propagate trusted descriptions, tags, classifications, and quality metadata downstream. #

Google’s Governance Agent project is built on Google Cloud Knowledge Catalog, BigQuery, and column-level lineage. It uses lineage records to determine where a downstream column originated and then carries existing governance metadata forward instead of requiring people to recreate it manually.

The project propagates column descriptions, governance tags, data classifications, and quality-related metadata from trusted upstream datasets. Google presented the approach as a response to the loss of context that occurs when raw tables are joined, filtered, reshaped into views, and then used to create additional downstream assets.

The underlying problem is that data meaning, personally identifiable information classifications, and quality context often do not travel with transformed data. As a result, a small number of core datasets may be thoroughly documented while downstream tables and views become progressively less documented even when the underlying data has not deteriorated.

The project is described as an automation approach to keeping governance metadata current in the background. The source presents Google’s architecture and intended workflow, but does not provide production-scale performance, customer adoption, operating cost, or accuracy measurements.

• Google’s Governance Agent project
• BigQuery

OpenAI paused code-capable frontier inference and two weeks of reinforcement-learning training; its largest program awaits safety validation. #

OpenAI said it immediately paused frontier-model inference tasks capable of executing code in its research cluster following the Hugging Face incident. It also slowed frontier scaling and paused reinforcement-learning training for its latest deployed model for two weeks.

The company’s largest frontier reinforcement-learning training program remained paused pending completion of safety validation. Chief Executive Officer Sam Altman said OpenAI acted because rapidly developing model capabilities needed corresponding alignment, safety, and monitoring standards.

OpenAI plans to monitor how its most capable unreleased models solve problems and use online tools, with a goal of notifying safety teams within 30 minutes after concerning behavior is detected. It also added controls intended to prevent certain models from connecting to the internet for higher-risk tasks.

The company further said it would require stricter sandboxing when training and evaluating models. These statements describe OpenAI’s disclosed safety actions and targets; the supplied sources do not identify the paused models, compute-cluster size, training run cost, or scheduled date for resumption.

• OpenAI

Mandiant’s Agentic Vulnerability Discovery Harness found more than 100 true-positive critical vulnerabilities in two days and contributed to 12 CVEs. #

Mandiant said its Agentic Vulnerability Discovery Harness, or AVDH, found more than 100 true-positive critical vulnerabilities in two days during a recent incident-response investigation involving stolen corporate repositories. The company said the work took a fraction of the time required for a manual review.

Mandiant has used AVDH for 10 months across environments containing tens of millions of lines of code. It has executed thousands of pipelines and produced tens of thousands of findings, with the framework used in proactive reviews, penetration tests, red-team operations, and incident response.

The work uncovered dozens of assignable flaws in widely used web extensions and open-source projects and resulted in 12 assigned Common Vulnerabilities and Exposures, or CVEs. The source specifically identifies CVE-2026-13242 and CVE-2026-55803 and says another dozen flaws were in active disclosure.

AVDH combines multiple AI agents with a structured orchestration layer and human subject-matter expertise intended to validate findings skeptically. Mandiant also reported finding a remote-code-execution vulnerability in a client web application, and said AVDH can operate alongside CodeMender’s continuing scans as a two-layer defense.

Box is integrating Gemini Multimodal Embeddings 2 to index text, images, document pages, spreadsheet tables, and charts together. #

Box and Google Cloud are integrating Gemini Multimodal Embeddings 2 into Box’s Agentic Platform. The technology places text, raster images, document pages, rendered spreadsheet tables, and visual charts in a unified multimodal vector space.

The integration is intended to preserve visual and spatial relationships that can be lost when documents are converted into flat text. Examples include maintaining the connection between column headers and data in financial tables, interpreting clinical visuals, and following the logic of multi-page flowcharts.

The system is also intended to support retrieval across mixed enterprise formats, such as a PDF policy, a spreadsheet tracking log, and a presentation deck. This extends text-based retrieval-augmented generation by allowing agents to search visual and structured content alongside narrative text.

Google and Box said enterprises store financial models, clinical-trial protocols, merger-and-acquisition due-diligence materials, engineering schematics, and compliance playbooks in Box. The source describes the architecture and planned capabilities but does not provide an integration-release date, pricing, embedding dimensions, retrieval-accuracy results, or customer deployment figures.

• Box
• Gemini Multimodal Embeddings 2