Daily News Coverage: NVIDIA Plans Open-Weight AI Model to Counter DeepSeek, Kimi K3, Confirms Full Production of Groq 3 & NVIDIA Vera CPU Deployment to SpaceX
The CODEW Daily News Coverage | August 25, 2026
Good Morning! NVIDIA shares extended their losing streak to seven days — the longest since 2022 — as AI hardware stocks sold off amid rising server costs and political scrutiny. NVIDIA is reportedly investing $60 billion to build the world's most powerful open-weight AI model, while the company also announced full production of its Groq 3 LPX low-latency inference racks. Hugging Face is exploring a potential $13 billion sale, Anthropic's new models appeared in the API, and NVIDIA's Vera CPU will power SpaceXAI's next-generation agents. Meanwhile, Xiaomi launched three AI chips, and Alibaba released its Wan3.0 video model. Here's today's briefing from The CODEW's newsroom.
THE LEAD
NVIDIA Shares Extend Losing Streak to 7 Days as AI Hardware Stocks Sell Off
What happened: NVIDIA shares fell 2.91% on Monday, marking the seventh consecutive day of declines — the longest losing streak since September 2022. The broader semiconductor sector also sold off, with the Philadelphia Semiconductor Index dropping 2.7% as all 30 components declined. Micron Technology fell 5.83%, AMD dropped 3.49%, Broadcom declined 2.63%, and Intel fell 3.12%. The selloff was driven by multiple factors: NVIDIA's AI server price increases of over 15% to customers, rising political scrutiny of AI data centers, and concerns about the sustainability of AI infrastructure spending.
Why it matters: The seven-day losing streak reflects growing investor anxiety about the sustainability of the AI hardware boom. NVIDIA's server price hike — driven by surging memory costs — signals that even the dominant AI chipmaker cannot fully absorb rising component costs. Political scrutiny of AI data centers is intensifying ahead of midterm elections, with some analysts identifying regulatory risk as a growing concern. The memory sector's strength — with Samsung, SK Hynix, and Micron controlling most DRAM capacity — suggests the bottleneck has shifted from chips to memory.
Who is affected: AI chipmakers (NVIDIA, AMD, Intel, Broadcom, Micron), hyperscalers facing higher server costs, and enterprise AI buyers.
What to watch next: NVIDIA's upcoming earnings report later this week will be a critical test of whether the selloff reflects genuine concerns or a temporary pullback.
AI & Infrastructure
NVIDIA Plans $60B Open-Weight AI Model to Counter China's DeepSeek, Kimi K3
NVIDIA is planning to spend $60 billion to license technology from AI startup Poolside and build one of the world's most powerful open-weight AI models, according to reports. The move is a direct response to China's rising dominance in open-weight models, including DeepSeek and Kimi K3. Under the agreement, NVIDIA will pay $60 billion for technology licensing and invest an additional $10 billion at a $120 billion pre-money valuation. More than 100 Poolside engineers will join NVIDIA's Nemotron open-source project. The company recently released the lightweight Nemotron 3.5 Lightning and is developing a 1-trillion-parameter supermodel.
Why it matters: The move represents a strategic pivot for the U.S. AI industry. American labs have historically prioritized closed models, but Chinese open-source models have surged ahead. NVIDIA is now taking direct aim at this gap, positioning itself to compete with its own customers — OpenAI and Anthropic — by offering a lower-cost, customizable open alternative.
NVIDIA Announces Full Production of Groq 3 LPX Racks for Low-Latency AI Inference
NVIDIA announced that its Groq 3 LPX racks have entered full production. The Groq architecture integrates 500 megabytes of high-speed SRAM directly on the chip to reduce memory bottlenecks. Each LPX rack integrates 256 Groq 3 chips, processing 3,400 tokens per second. The racks will be deployed alongside Vera CPUs and Rubin GPUs in new cloud provider Nebius data centers. Groq chips are manufactured by Samsung, while NVIDIA's GPUs are produced by TSMC.
Why it matters: Low-latency inference is emerging as the next frontier in AI competition. As AI agents become more interactive, response speed directly impacts user experience. The market is heating up: AMD is integrating Cerebras chips into its rack systems, and OpenAI recently launched an "Ultrafast" mode with Cerebras support. Low-latency chips are not replacing GPUs but address specific workloads — the "decode" stage of model serving.
SpaceXAI Deploys NVIDIA Vera CPU for Next-Generation Agent Applications
NVIDIA announced that SpaceXAI, the artificial intelligence division of SpaceX, will deploy NVIDIA Vera CPUs to develop and operate next-generation AI agent applications. The chips will be used in multiple projects, including the Grok chatbot, handling taskssuch asn tool calling, code execution, data analysis, and model inference to reduce agent response times.
Why it matters: The deployment extends AI compute from terrestrial data centers to space applications, representing a significant milestone in the convergence of space technology and AI infrastructure. For enterprise buyers, it signals that AI agent performance is increasingly dependent on specialized silicon optimized for inference workloads.
Anthropic's New 'Marshmallow' and 'Melon' Models Appear in API; Hugging Face Explores $13B Sale.
Developers discovered two new Anthropic models in the company's API with internal codenames "marshmallow" and "melon" — both in early access preview. Early testers report Marshmallow's performance already surpasses the current Opus 5, suggesting the models may represent updates to the Opus and Sonnet product lines. The internal codename shift from animals to food items may indicate a new development branch. Anthropic's iteration cadence is approximately every two weeks, suggesting an official release could come as soon as next month.
Separately, Hugging Face, the world's largest AI community with over 3 million models and 1 million datasets, is exploring a potential sale that could value the company at more than $13 billion. The company has hired a bank to gauge buyer interest.
Why it matters: Anthropic's rapid iteration pace — two new models in API following the recent release of Mythos 5 — signals intensifying competition in the foundation model layer. Marshmallow's reported performance surpassing Opus 5 suggests Anthropic is maintaining its competitive position against OpenAI and Google. The Hugging Face sale exploration comes just weeks after Stripe's $7 billion+ acquisition of OpenRouter, indicating that AI value is rapidly flowing to the middleware layer — platforms that aggregate and distribute models — rather than just the model builders themselves.
NVIDIA in Talks to Invest in Perplexity at Over $300B Valuation
NVIDIA is in discussions to participate in AI search startup Perplexity's next funding round, which would value the company at more than $300 billion. Perplexity's annualized revenue has grown from under $250 million at the start of the year to over $750 million — a threefold increase in approximately eight months. The company plans to go public in 2028.
Why it matters: The potential investment — NVIDIA's latest in a series of strategic AI investments — would further entrench the "circular financing" model that has drawn regulatory scrutiny. Critics argue that NVIDIA's investments in its own customers artificially inflate demand. The International Bank for Settlements recently identified circular financing as one of three major global financial stability risks, while the Bank of England has warned that the pace of AI investment is historically unprecedented.
AI Operating System 'Omarchy' Gains Traction as Silicon Valley Bets on Next Harness
Linux distribution Omarchy released version 4.0, embedding an AI programming agent directly into the operating system. Users can control their computer using natural language commands, and the system has attracted over a thousand community-contributed plugins in ten days. The desktop has been rewritten as a single process with all configuration as text scripts, allowing AI to read and modify every component. The project's founder, DHH, along with Dell and Jack Dorsey, has formed the Omacom Foundation with $8 million in initial funding.
Why it matters: Omarchy represents a bet on a new layer of the technology stack — between the model and the hardware — that could become the operating system for the AI era. If successful, it could reshape how developers interact with computers and reduce dependence on traditional OS vendors.
Semiconductors & Hardware
Xiaomi Launches Three AI Chips: Xuanjie O3, O100, and D100
Xiaomi unveiled its next-generation Xuanjie family of three AI chips, representing over 21 billion yuan ($2.9 billion) in cumulative R&D investment over five years. The flagship Xuanjie O3 is a 3nm AI SoC with 24 billion transistors and an NPU optimized for large language models, scoring 5.22 million on AnTuTu. The Xuanjie O100 is the world's first wafer-level stacked AI accelerator chip with 1.22 TB/s bandwidth. The Xuanjie D100 is China's first 3nm autonomous driving high-compute chip. The O3 will debut in the Xiaomi 18 Fold.
Why it matters: Xiaomi's entry into high-end AI chips — particularly the wafer-level stacked O100 — signals China's ambition to develop indigenous AI silicon capabilities. The 3nm process and wafer-level stacking technology suggest Xiaomi is targeting frontier AI workloads previously dominated by U.S. suppliers.
SK Hynix to Use Intel EMIB for Next-Generation HBM as TSMC CoWoS Capacity Overflows
SK Hynix will adopt Intel's EMIB (Embedded Multi-Die Interconnect Bridge) advanced packaging technology for its next-generation HBM memory. The move comes as TSMC's CoWoS capacity is fully booked, causing orders to overflow to alternative suppliers. SK Hynix disclosed the 2.5D advanced packaging solution at the 2026 Hot Chips conference.
Why it matters: The shift represents a significant competitive development for Intel's foundry business, which has been working to establish itself as a credible alternative to TSMC. The CoWoS capacity crunch — now severe enough to push a major HBM supplier to adopt a competitor's packaging — highlights the severity of advanced packaging constraints in the AI supply chain.
Enterprise & Cloud
Alibaba Releases Wan3.0 Video Model with Native 30-Second Generation
Alibaba Cloud's Bailian platform opened the Wan3.0 video generation model to all developers. The model natively generates 30-second, 1080P cinematic-quality video with synchronized audio. It supports comprehensive reference inputs including images, text, audio, video, documents, and web pages, with up to 10 images, 5 videos, and 5 audio inputs combined.
Why it matters: Wan3.0 represents a significant advance in AI video generation capability, particularly in native audio-video synchronization and multi-modal input support. As enterprise demand for AI-generated video content continues to grow, Alibaba is positioning itself to compete with U.S. providers in the enterprise AI video market.
Cybersecurity
CISA Orders Federal Agencies to Patch Actively Exploited TrueConf Server Vulnerabilities
CISA has added two TrueConf Server vulnerabilities — CVE-2026-72529 and CVE-2026-72530 — to its Known Exploited Vulnerabilities (KEV) catalog. Both flaws enable unauthenticated remote code execution. Kaspersky reported that the hacker group HeadMare has been exploiting these vulnerabilities since July. Federal agencies have been ordered to remediate the flaws by August 25.
Why it matters: TrueConf Server is widely deployed in enterprise and government environments for video conferencing and collaboration. Active exploitation by a sophisticated threat group underscores the urgency of patching. Organizations using TrueConf Server should prioritize remediation immediately.
CakePHP Critical Vulnerability (CVSS 9.2) Published; Incus Container Manager Vulnerability Disclosed
A critical severity vulnerability (CVSS 9.2) affecting CakePHP and cakephp/database was published on August 25 (CVE-2026-77635). Additionally, a high-severity vulnerability in Incus container manager (CVE-2026-63125) affects all versions before 7.3.0. No public proof-of-concept is currently available, but the exploit logic is clear and carries the risk of rapid weaponization.
Why it matters: The CakePHP vulnerability (CVSS 9.2) is among the most severe disclosed this week. Organizations using CakePHP should prioritize assessment and patching. The Incus container manager vulnerability — while not yet listed in CISA's KEV catalog — poses a significant risk to containerized environments due to its clear exploit logic.
Funding & Markets
XPeng Robotics Raises $900M Series A at $6.3B Valuation, Setting Industry Record.
XPeng's robotics business completed a first-round funding of over $900 million at a post-money valuation of $6.3 billion, setting a new record for China's embodied AI industry. The round was led by IDG Capital with participation from GSR Ventures and strategic investments from Tencent and Alibaba. XPeng's IRON humanoid robot will enter mass production by the end of 2026.
Why it matters: The record-setting round reflects growing investor conviction in embodied AI and humanoid robotics. The participation of Tencent and Alibaba as strategic investors signals that China's largest technology companies see robotics as a critical AI frontier. The $6.3 billion valuation — comparable to many publicly traded robotics companies — suggests the IPO market for AI robotics could be robust.
Smart Ring Maker Oura Seeks Up to $3B IPO at Over $16B Valuation
Smart ring manufacturer Oura is seeking to raise up to $3 billion in an initial public offering at a valuation exceeding $16 billion. The company has not yet filed an S-1.
Why it matters: Oura's IPO plans reflect the broadening of the AI-enabled consumer hardware market beyond traditional tech companies. The $16 billion+ valuation would make Oura one of the most valuable consumer wearables companies, signaling investor confidence in the AI-health intersection.
What to Watch
- NVIDIA earnings: The company's upcoming earnings report will be a critical test of whether the seven-day losing streak reflects genuine concerns or a temporary pullback.
- Anthropic model releases: When Marshmallow and Melon officially launch — and whether they deliver on performance promises that surpass Opus 5.
- Hugging Face sale: Who emerges as the potential buyer and at what price — a $13B+ acquisition would be one of the largest AI infrastructure deals in history.
- TrueConf Server vulnerabilities: Whether federal agencies meet the August 25 CISA remediation deadline and whether exploitation activity expands.
- Xiaomi AI chips: How Xiaomi's Xuanjie chips perform in real-world deployments and whether they gain traction beyond Xiaomi's own devices.
Reviewed by Erwin Castro
on
Tuesday, August 25, 2026
Rating:
