🇫🇷Mistral AI raised $3.5 billion in Samsung Electronics-led Series D financing, reaching a valuation above $24 billion. #
Paris-based generative-AI startup Mistral AI announced a $3.5 billion Series D fundraise, according to Crunchbase News. The report places its valuation at more than $24 billion, rather than exactly $24 billion, and says the valuation nearly doubled.
Samsung Electronics led the financing. A separate financialjuice message reports that Samsung will integrate Mistral AI across its semiconductor operations, establishing a stated operational component alongside the investment.
The source describes the round as money raised, not a maximum financing facility or a fundraising target. However, the supplied material does not specify Samsung’s investment amount, ownership percentage, security terms, closing conditions, or Mistral’s allocation of the proceeds.
- Mistral AI
- Samsung Electronics
🇺🇸AI-coding startup Cognition AI raised $2 billion in Series E financing at a $48 billion valuation, led by a16z and Accel. #
Cognition AI announced on September 8 that it had raised $2 billion in Series E financing at a $48 billion valuation (已在 E 轮融资中筹集 20 亿美元，估值达到 480 亿美元). The report describes an announced completed fundraising, rather than a proposed financing facility.
New investors a16z and Accel led the round, with existing investors Founders Fund, General Catalyst, and Avenir participating (本轮融资由新投资者 a16z 和 Accel 领投，现有投资者 Founders Fund、General Catalyst 和 Avenir 参投). Cognition said its annualized revenue increased from $492 million to nearly $900 million since its previous financing in May (自 5 月上一轮融资以来，其年化收入已从 4.92 亿美元增至近 9 亿美元).
That revenue figure is explicitly annualized, not reported full-year revenue, and remains a company statement relayed by Telegram. An earlier alert described the amount raised as more than $2 billion (融资额超过 20 亿美元), while the later, fuller report used $2 billion; neither provides share pricing, dilution, investor allocations, or financing conditions.
🇺🇸Qualcomm issued Amazon a warrant covering 25 million shares maximum at $161.26/share, expiring 2036; 3.75 million shares vest upon issuance. #
Qualcomm, identified in the source as QCOM, issued Amazon, identified as AMZN, a warrant covering up to 25 million shares. The Telegram report states a price of $161.26 per share and an expiration year of 2036, without providing the precise expiration date.
A separate message says the warrant portion covering 3.75 million shares vests upon issuance. That is a vesting provision, not evidence that Amazon has exercised the warrant or received all 25 million underlying shares.
The warrant announcement appeared alongside reporting of a multigenerational product agreement between the companies. The supplied messages do not disclose the remaining vesting schedule, purchase-linked conditions, settlement mechanics, or a percentage dilution calculation, and they do not include the underlying securities filing.
🇺🇸GlobalFoundries finalized a definitive $375 million award agreement with the U.S. Department of Commerce for quantum semiconductor research. #
GlobalFoundries announced that it had finalized a definitive agreement for a $375 million award with the Department of Commerce’s CHIPS Research and Development Office. The September 8 announcement concerns research and development funding, rather than a commercial chip-purchase contract.
The stated purpose is to accelerate the build-out of GlobalFoundries’ Quantum Technology Solutions business. The announcement also identifies strengthening the American quantum-semiconductor ecosystem as an objective, but the supplied excerpt does not quantify manufacturing capacity or research milestones.
The transaction stage is a finalized award agreement, not merely a preliminary proposal. The excerpt does not establish that the full $375 million has already been disbursed, nor does it provide payment schedules, matching-investment requirements, repayment provisions, or other award conditions.
🇺🇸The U.S. expanded 50% additional tariffs to selected Canadian steel and aluminum products, effective September 15. #
Telegram reporting attributes to the White House an adjustment to the coverage of an additional 50% ad valorem tariff—a duty assessed by value—on selected Canadian products (调整对加拿大部分产品征收 50%额外从价关税的适用范围). The stated implementation time is 12:01 a.m. U.S. Eastern time on September 15 (美国东部时间 9 月 15 日凌晨 0 时 01 分).
The product descriptions include selected steel products, specified aluminum products, and certain base-metal products (部分钢铁产品、特定铝产品、部分贱金属产品). The broader announcement also covers categories including paper products, golf carts, certain outboard motorboats, and cheese or substitutes not made from cow’s milk (部分纸制品、高尔夫球车及部分舷外机动艇、非牛乳制成的奶酪及奶酪替代制品).
Importantly, the reported change adds some products while removing the previous additional tariff from others (同时取消部分加拿大产品此前适用的 50%额外关税). The source says the attached proclamation lists determine exact coverage; those lists are not supplied, so the material does not establish which individual data-center construction materials or equipment imports are affected.
Qualcomm and AWS announced multigeneration collaboration on custom AI-inference silicon and optical connectivity extending to 1.6 terabits per second. #
Qualcomm Technologies announced a multigeneration collaboration with Amazon to enable customized silicon at scale for large AI data centers. The HPCwire announcement identifies AI inference as the compute workload, rather than stating that the agreement covers model-training hardware.
The companies are also working on optical connectivity extending up to 1.6 terabits per second, as well as future-generation solutions. That figure describes connectivity speed, not accelerator compute throughput, GPU count, or total data-center capacity.
Separate Telegram reporting says Qualcomm expects Amazon revenue to begin in the December quarter and reports management’s confidence in a $5 billion data-center target. These are management expectations, not recognized revenue; the supplied messages do not specify the target’s period, committed order value, or the hardware configuration behind the collaboration.
Qualcomm’s chief financial officer also said the company has another hyperscaler engagement, meaning another engagement with a large-scale cloud operator. The source does not name that counterparty or provide its commercial terms, so it cannot be combined with the Amazon announcement to establish an additional quantified contract.
Arm clarified that C2-Ultra’s maximum instructions-per-cycle improvement over C1-Ultra is 7% under specified parameters, excluding memory-bandwidth gains. #
A September 8 editor’s note from Chips and Cheese says Arm contacted the publication to clarify the comparison between its C2-Ultra and C1-Ultra CPU cores. The corrected maximum instructions-per-cycle, or IPC, improvement is 7%, using the parameters referenced in the article’s endnotes.
The clarification explicitly says the memory-bandwidth improvement does not factor into that IPC improvement. Consequently, the source reports two distinct aspects of the design comparison rather than a single combined performance percentage.
The 7% is a maximum under specified conditions, not an average across workloads or a measured improvement for a complete server. The supplied excerpt does not reproduce the endnote parameters, benchmark results, clock frequencies, or memory-bandwidth figure, preventing a more detailed numerical comparison.
Semiconductor research center imec presented three-metal-level niobium-titanium-nitride superconducting circuits and 30-nanometer interconnects at the Applied Superconductivity Conference. #
Imec’s September 8 announcement describes superconducting technology being presented that week at the 2026 Applied Superconductivity Conference, abbreviated ASC. The reported work includes three-metal-level, or 3ML, niobium-titanium-nitride, or NbTiN, circuits and 30-nanometer interconnects.
These figures describe different structural features: three metal levels in the circuit technology and a nanometer-scale interconnect dimension. The excerpt does not identify 30 nanometers as a complete semiconductor process-node designation or provide a system-level performance result.
Imec also describes a new superconducting digital program bringing together foundries, system companies, and hyperscalers. Its stated application areas include high-performance computing, AI, quantum computing, photonics, and neuromorphic computing; the supplied material does not specify production volumes, commercial availability, or measured power savings.
SkyWater Technology launched SkyWater Quantum Solutions, a merchant foundry business covering semiconductor process development, integration, and manufacturing. #
SkyWater Technology announced the launch of SkyWater Quantum Solutions on September 8. The company describes it as a dedicated merchant-supplier offering focused on semiconductor process development, integration, and manufacturing.
The stated application scope extends beyond quantum computing to quantum networking and sensing. The announcement therefore concerns semiconductor development and manufacturing services, rather than the launch of a named quantum computer or cloud-compute service.
Dr. Mihir Bhaskar will lead the business as senior vice president and general manager. The supplied excerpt gives no customer contracts, manufacturing-capacity figures, capital-expenditure commitments, or revenue results associated with the launch.
- SkyWater Technology
- SkyWater Quantum Solutions
NextEra secured a $1.9 billion U.S. Department of Energy loan for the Duane Arnold nuclear-plant restart. #
Utility Dive reports that NextEra secured a $1.9 billion Department of Energy loan for restarting the Duane Arnold nuclear plant. The financing concerns a plant restart, not construction of a newly named nuclear facility.
The article’s supplied description places this transaction within a broader federal financing effort. It says the Department of Energy under the Trump administration has financed three plant restarts and made other conditional commitments to bring nuclear power online.
The report describes this particular loan as secured, while separately identifying other commitments as conditional. However, the supplied excerpt does not disclose disbursement status, interest rate, maturity, collateral, plant capacity, restart timing, or a data-center electricity offtake agreement.
Google backed a West Virginia solar-storage project planned with 100 megawatt-hours of Eos zinc-based storage alongside lithium-ion batteries. #
Utility Dive reports that Google is backing a solar-and-storage project at a former West Virginia coal mine. MN8 is developing the project, separating the developer’s role from Google’s reported backing.
The planned storage configuration includes 100 megawatt-hours of long-duration, zinc-based storage from Eos Energy Enterprises, alongside lithium-ion batteries. The 100-megawatt-hour figure applies to the Eos storage component, not necessarily to the combined storage installation.
Megawatt-hours measure stored energy, not the solar plant’s generating capacity or the batteries’ discharge-power rating. The supplied excerpt does not provide those power ratings, the lithium-ion storage capacity, an operating date, or the financial and contractual form of Google’s support.
Google–Blackstone’s Project Braid, an AI-cloud venture, faces reported data-center construction delays, including a site affected by missing transformers. #
Financial_Express reports, citing people familiar with the matter, that Project Braid is experiencing data-center construction delays. It describes the venture as involving $5 billion invested by Blackstone and plans to rent compute resources equipped with Google AI chips in 2027 (该项目由黑石投资 50 亿美元，计划于 2027 年向客户出租搭载谷歌 AI 芯片的计算资源).
At the Cheyenne, Wyoming site, Google reportedly canceled an arrangement under which AI data-center developer Crusoe would build the facility, took over the project, and reapplied for relevant permits (取消该安排并接手项目，重新申请相关许可). The report attributes that decision to sources saying Google had lost confidence in Crusoe’s ability to deliver on schedule.
Crusoe is identified separately as a developer involved in OpenAI and Oracle’s Texas Stargate project (得州建设的旗舰“星际之门”数据中心项目); Project Braid is not that project. Another Braid site was reportedly affected by missing required transformers (另一个数据中心选址因缺少所需变压器受到影响), while the report also identifies equipment shortages and permitting delays as construction bottlenecks.
Project leadership said work remained underway and reiterated a plan for 500 megawatts of “compute capacity” next year (项目仍在推进，并计划明年提供 500 兆瓦计算容量). This is a management deployment plan—not commissioned capacity—and the report does not define whether the megawatt figure represents IT load or another electrical-capacity boundary, or document the funding’s closing and drawdown terms.
- Google
- Blackstone
- Project Braid
The Electric Reliability Council of Texas recorded weekly average peak loads exceeding its 2023 record for six consecutive weeks. #
Utility Dive reports that the Electric Reliability Council of Texas, or ERCOT, recorded weekly average peak loads above a record set in 2023. The report attributes the finding to the U.S. Energy Information Administration, or EIA.
The observation covers six weeks beginning in late July, with every week in that period exceeding the earlier benchmark. The metric is specifically weekly average peak load, not a statement that a new instantaneous system peak occurred every week.
The supplied description does not give the load values in megawatts, the size of the increase, or a sector-by-sector demand breakdown. It therefore does not establish how much of the record load came from data centers, nor does it report outages or inadequate generation reserves.
AWS made OpenAI’s GPT-6 Astra accessible through Amazon Bedrock in its September 8 announcement. #
A September 8 financialjuice message attributes to an Amazon blog the announcement that OpenAI’s GPT-6 Astra is now accessible on Amazon Bedrock. The wording reports present availability, rather than a planned future integration.
A related Amazon-attributed message says OpenAI also introduced new enterprise plugins for ChatGPT Work to extend Astra’s browser-use capabilities into commonly used business applications (为 ChatGPT Work 推出全新企业插件，可在常用商业应用中拓展 Astra 的浏览器使用能力). That is a separate product statement; the supplied material does not establish that these plugins are features of the Bedrock offering.
The available sources do not specify Bedrock regions, pricing, quotas, deployment modes, or the hardware serving Astra. Although the Telegram alert cites a blog, the underlying Amazon blog URL is not included, so the source attribution remains to the supplied Telegram messages.
- AWS
- OpenAI’s GPT-6 Astra
- Amazon Bedrock
- September 8
Quantum Cyber bought an AI-compute cluster based on NVIDIA’s A100 GPU hardware. #
A September 8 financialjuice alert reports that Quantum Cyber bought an NVIDIA A100 AI-compute cluster. The wording identifies a hardware purchase, rather than merely an intention to acquire compute equipment.
The named hardware is NVIDIA A100 GPU hardware, while the acquired asset is described at the cluster level. The message does not disclose the number of GPUs, servers, or racks, and Quantum Cyber’s name does not establish that this is quantum-computing hardware.
The source gives no purchase price, financing terms, seller, installation location, deployment date, or customer contract. It also supplies no ticker, exchange, or legacy-business description, so the purchase alone does not establish a public-company pivot or an operating GPU-as-a-Service business.
- Quantum Cyber
- NVIDIA’s A100 GPU hardware
Sweden inaugurated Arrhenius, a new supercomputer located at Linköping University, on September 8. #
The European High Performance Computing Joint Undertaking, or EuroHPC JU, and the National Academic Infrastructure for Supercomputing in Sweden, or NAISS, inaugurated Arrhenius on September 8. The announcement identifies it as a new EuroHPC supercomputer located in Sweden.
The unveiling took place at Linköping University during a ribbon-cutting ceremony. Anders Dam Jensen, identified as EuroHPC JU’s executive director, attended the event.
The supplied excerpt establishes the inauguration and institutional participants but does not specify the system’s processors, accelerators, node count, memory, or interconnect. It also provides no benchmark performance, power consumption, acquisition cost, or user-access schedule, so no quantitative ranking or capacity comparison is supported.
DeepSeek announced Flash API price reductions effective September 10, noon Beijing time, with cache-hit pricing falling by up to 60%. #
Financial_Express relays a DeepSeek Open Platform announcement that revised Flash-series application programming interface, or API, pricing begins at noon Beijing time on September 10, 2026 (将于北京时间 2026 年 9 月 10 日 12:00 起，调整 flash 系列定价). The maximum reported reductions are 60% for cache hits, 33.33% for cache misses, and 11.11% for output (缓存命中、缓存未命中及输出价格最高降幅分别为 60%、33.33%和 11.11%).
The quoted off-peak prices are RMB 0.02 for input cache hits, RMB 1 for input cache misses, and RMB 4 for output (空闲时段输入缓存命中单价 0.02 元、输入缓存未命中单价 1 元，输出单价 4 元). Crucially, the supplied message does not identify the billing denominator, so these figures cannot safely be labeled as prices per million tokens or per request.
Peak-period pricing is twice off-peak pricing (高峰时段价格为空闲时段价格的 2 倍). Peak periods are Monday through Friday, 9:00–12:00 and 14:00–18:00 Beijing time, with all other times designated off-peak (高峰时段是指北京时间周一至周五 9:00-12:00、14:00-18:00（其余为空闲时段）).
Meta launched Muse, a personal AI agent for U.S. users, offering free access and $20/$100 monthly subscriptions. #
Meta launched Muse, a personal AI agent, with a free version and subscriptions priced at $20 and $100 per month (提供免费版本以及每月 20 美元和 100 美元两档订阅服务). Meta’s AI leader, Alexandr Wang—the person referenced in the source as Alexander Wang (亚历山大·王)—said the agent is powered by the Muse Spark family of foundation models (由 Meta 的 Muse Spark 系列基础模型驱动).
Reported tasks include making appointments, filling spreadsheets, and monitoring home cameras (预约、填写电子表格、监控家庭摄像头). The September 8 rollout covers U.S. users on iOS, Android, and the Muse web service, while AI-glasses access is described as coming soon rather than already available (Muse 正在美国面向 iOS、Android 和 muse.ai 用户推出，并将很快登陆 AI 眼镜).
Wang said Muse operates in an isolated environment within Meta’s compute infrastructure, does not read users’ actual passwords or payment information, and requests confirmation before sensitive actions (运行在 Meta 计算基础设施中的独立隔离环境内，不会读取用户实际密码或支付信息，并会在执行敏感操作前请求确认). Users can choose whether interactions may be used for model training; Meta says it removes key personally identifying information when users opt in (用户可选择是否允许 Meta 使用与 Muse 的互动内容训练 AI 模型，若选择允许，Meta 将删除关键个人身份信息后用于模型改进).
Meta also announced a later-2026 plan for Muse Confidential VM, with VM meaning virtual machine. The stated design encrypts the entire virtual machine, including personal data and Muse conversations, using a key held only by the user (整个虚拟机（包括用户的个人数据以及与 Muse 的对话）都将使用仅由用户持有的密钥进行加密); this is a planned feature, not evidence of protection already deployed at launch.
OpenAI released ChatGPT Images 2.5 to all ChatGPT, ChatGPT Work, and Codex users across desktop, mobile, and web. #
A September 8 financialjuice message attributes to OpenAI the announcement that Images 2.5 is available to all ChatGPT, ChatGPT Work, and Codex users. The stated platform coverage is desktop, mobile, and web, without a narrower paid-tier restriction in that message.
A fuller Telegram report says the new image engine improves generation speed by up to 50% over the previous generation (新模型生成速度较上一代提升最高 50%). It also reports improved preservation of people’s and pets’ original features and more precise editing through sketches or circled image regions (能够更好保持人物、宠物原始特征，并支持用户通过草图、圈选区域等方式直接指定修改位置).
Reported uses include converting photographs into different artistic styles, designing logos, and modifying tattoo concepts (将照片转换为不同艺术风格、设计 Logo、修改纹身方案). A later message says GPT-Image-2.5 Sunburst and GPT-Image-2.5 Flare are available through the API (GPT-Image-2.5 Sunburst 和 GPT-Image-2.5 Flare 模型已通过 API 提供).
The supplied material does not include the test conditions behind the maximum speed improvement, image-resolution limits, or API prices. One earlier Telegram message uses the alternative labels “DALL·E 2.5” and “ChatGPT Enterprise,” whereas the financialjuice alert and later fuller report use Images 2.5 and ChatGPT Work; the available material does not reconcile that inconsistency.
- OpenAI
- ChatGPT Images 2.5
- ChatGPT
- ChatGPT Work
- Codex
Japanese telecom operator KDDI reduced its Buffmee AI app’s response latency by 38% and improved time-to-first-token by nearly 18%. #
Google Cloud’s customer account reports that KDDI reduced total application response latency by 38% and improved time-to-first-token, or TTFT, by nearly 18% for Buffmee. These are separate application-performance metrics, and the report says the work brought the service to KDDI’s target response performance.
Buffmee is a consumer retrieval-augmented generation, or RAG, application built around the concept of “AI that helps you grow.” It grounds responses in more than 100 sources, including books, magazines, and web media, and supports information search, summarization, personalized learning, and exploration of hobbies.
The development team faced latency problems while grounding a large collection of proprietary content, including books and magazines. It used the Gemini Enterprise Agent Platform Evaluation Service, an LLM-as-a-Judge approach—using a large language model to evaluate outputs—and the Rule of Hundreds framework to replace labor-intensive manual testing with automated evaluation.
KDDI constructed hundreds of automated evaluation tests and a benchmark dataset to assess answer reliability across use cases, reporting a 25% improvement in groundedness scores. Those results come from Google Cloud’s customer case study; the supplied excerpt does not give absolute latency values, the groundedness scoring scale, or evidence that errors or hallucinations were eliminated.
