Today marks a pivotal shift in AI regulation as the European Union brings ChatGPT under the Digital Services Act (DSA), alongside massive hardware bets from Nvidia and Together AI. From 305B multimodal models to $1B ad run rates, the AI landscape is scaling in both capability and compliance. We are seeing a transition from "theoretical" safety to "enforcement" safety, while the hardware war moves from general GPUs to specialized rack-scale factories and custom silicon.
EU AI Regulation: ChatGPT Under DSA
ChatGPT Designated as VLOSE
The European Commission has officially designated ChatGPT as a Very Large Online Search Engine (VLOSE) under the Digital Services Act (DSA) rulebook. This designation follows the reporting of 159 million monthly active users (MAU) within the EU, far crossing the 45 million threshold required for this status.
Compliance Deadlines and Obligations
ChatGPT, along with Reddit (57.2M MAU) and Roblox (48M MAU), now has until the end of November to complete comprehensive systemic-risk assessments and submit to independent third-party audits. The goal is to ensure these platforms are managing risks related to disinformation, illegal content, and algorithmic bias.
Severe Financial Penalties
The stakes for non-compliance are immense. Under the DSA, failure to meet these requirements can result in fines of up to 6% of global annual turnover. This moves AI regulation from a "wait and see" approach to a strict liability framework.
First Enforcement RFIs
The EU AI Office has already issued its first formal enforcement Requests for Information (RFIs) to OpenAI, Anthropic, and Google. These requests specifically target security monitoring protocols and training-content summary compliance. Misleading answers or non-response can carry fines up to €15 million or 3% of global turnover.
Source: aiweekly.co, European Commission
Open Weights War: DeepSeek V4-Vision & Tencent Hy4
DeepSeek-V4-Flash-Vision-Exp Release
DeepSeek has quietly released DeepSeek-V4-Flash-Vision-Exp on Hugging Face under an MIT license. This is a 305-billion parameter experimental multimodal model built on the V4-Flash architecture. It integrates advanced vision encoding to allow the model to "see" and reason over visual data with high precision.
Multimodal Agent Performance Gains
The model card reveals substantial gains in multimodal agent tasks. Specifically, it scored 36.5 on ApexBench Pass@1, a significant jump from the 26.2 scored by its predecessor, V4-Flash-0731. It also showed improvements on the "Agents' Last Exam" benchmark (27.3 vs 25.2).
Tencent's Hy4-Preview Powerhouse
Tencent's Hunyuan team has released Hy4-preview, a massive 770-billion parameter model. It uses a mixture-of-experts (MoE) architecture with 49 billion parameters activated per token and 256 routed experts. This allows it to maintain high efficiency despite its scale.
Massive Context Window Capabilities
Hy4-preview features a 1-million-token context window, enabling it to process entire codebases or long documents in a single prompt. It claims an impressive 92.3 on GPQA Diamond and 65.7 on SWE-bench Pro, positioning it as a top-tier open-weight rival to proprietary models like Claude Opus 4.8.
Source: aiweekly.co, Hugging Face
Hardware & Infrastructure: Nvidia-MediaTek & Saudi Mega-Cluster
Nvidia's Strategic $3.5B MediaTek Investment
Nvidia is purchasing $3.5 billion in convertible bonds from Taiwan's MediaTek. This is not just a financial investment but a technical tie-up: MediaTek will adopt NVLink Fusion. This technology allows customers to wire custom XPUs (cross-platform processing units) directly into Nvidia's rack-scale AI factories, diversifying the hardware available in these clusters.
Together AI's $5B Saudi Partnership
Together AI has inked a strategic partnership with Saudi state-backed HUMAIN to build a 250MW AI data center. The deal is expected to generate over $5 billion in gross annualized revenue in its first year. This move expands Together's compute capacity significantly, bypassing some of the regulatory and social headwinds facing domestic US buildouts.
China's HBM3E Breakthrough
ChangXin Memory Technologies (CXMT) has commenced low-volume production of HBM3E (High Bandwidth Memory). While current yields are low—approximately 25% on 8-high stacks—it marks a critical milestone in China's quest for AI hardware independence, closing the gap with global leaders SK Hynix and Samsung.
Global AI Capex Trends
The trend toward "AI factories" is accelerating. With companies like AM Intelligence ordering 9,000 Nvidia Vera Rubin systems for an $8 billion buildout in India, the focus has shifted from individual chips to massive, power-efficient clusters.
Source: Bloomberg, aiweekly.co, The New York Times
Legal Battles: Sony/Warner sue Anthropic & Pentagon vs Anthropic
Multi-Billion Dollar Copyright Lawsuit
Sony Music Publishing and Warner Chappell have filed a lawsuit against Anthropic, CEO Dario Amodei, and co-founder Benjamin Mann. They allege "tens of thousands" of copyrighted songs were torrented from Library Genesis and scraped from MusixMatch to train Claude. The suit seeks up to $150,000 per work, creating a potential multi-billion dollar exposure.
Victory Against the Pentagon Blacklist
A U.S. District Judge ruled that the Pentagon's designation of Anthropic as a "supply-chain risk" was unlawful. The judge described the action as First Amendment retaliation after Dario Amodei refused to allow the military to use Claude for unrestricted mass surveillance and fully autonomous weaponry.
Infostealer Malware Crisis
Anthropic is currently fighting an "infostealer" malware outbreak. Malware like Vidar and LummaC2 siphoned active login sessions from user PCs, allowing attackers to drain usage limits. Anthropic is signing out affected users and refunding unauthorized charges.
The Authenticity Debate
These events highlight the growing tension between AI labs and the "real world"—whether it's the intellectual property of artists or the security requirements of national governments. The "black box" nature of training data is now a primary legal liability.
Source: The Guardian, aiweekly. konusunda, Anthropic
Enterprise & Monetization: OpenAI's $1B Ad Run Rate & Custom Silicon
ChatGPT's $1B Advertising Milestone
OpenAI disclosed that its advertising operation has crossed a $1 billion annualized revenue run rate. This occurred less than 200 days after the February 2026 pilot began. The company is now guiding toward $2.5 billion in ad revenue for the year, diversifying its business model ahead of a projected IPO.
OpenAI's Custom Silicon: The Jalapeño Chip
SemiAnalysis revealed details of "Jalapeño," OpenAI's first custom inference chip taped out with Broadcom. Running on TSMC N3P, the B0 stepping hits 13.4 PFLOPs of MXFP4 at 700W. It is reported to be 1.5-1.9x more power-efficient than Nvidia's Rubin architecture for models like GPT-OSS.
Shift to Outcome-Based Pricing
OpenAI is testing "pay-per-outcome" pricing with major customers. Instead of paying for tokens, customers pay only when the AI successfully completes a specific task. This aligns AI costs directly with business value, though critics warn that attributing a complex business outcome to a single AI system is difficult.
The "Circular Financing" Pattern
Analysts are flagging a pattern where Nvidia invests in companies (like MediaTek) that then use that capital to buy more Nvidia hardware, potentially inflating growth metrics through a circular financial loop.
Source: aiweekly.co, SemiAnalysis, The Information
Robotics & Edge: Skild S1 & Apple M6/M5 Ultra
Skild S1: Zero-Shot Long-Horizon Tasks
Skild AI's S1 model can execute robotics tasks up to 10 minutes long from a single human video prompt. In tests, it achieved a 66% success rate on unseen tasks (like pancake flipping), compared to just 9% for language-prompted VLAs. It requires no fine-tuning to adapt to new environments.
Apple's 2nm M6 and M5 Ultra
Apple launched the M6 chip on a cutting-edge 2nm process. More impressively, the M5 Ultra offers 4.5x the AI GPU compute of the M3 Ultra. With unified memory speeds of 1.2TB/s, Apple is positioning the Mac as a high-end local AI workstation for developers.
Physics AI and the End of Transformers?
Accelerated Understanding Inc has unveiled an enterprise physics AI based on neural operators rather than Transformers. In tests, it ingested 5 trillion data points in a single prompt—roughly 5 million times the capacity of current flagship models from Google or Anthropic.
The Edge AI Convergence
From 2nm chips to robotics that learn from video, the "edge" is becoming just as capable as the cloud. This convergence allows for real-time, private AI that doesn't rely on a constant connection to a massive data center.
Source: aiweekly.co, WSJ
Frequently Asked Questions
What is the EU DSA designation for ChatGPT?
ChatGPT has been named a "Very Large Online Search Engine" (VLOSE). This means it is subject to the Digital Services Act's strictest rules on transparency, auditing, and risk management for platforms with over 45 million EU users.
How does DeepSeek-V4-Vision-Exp improve on previous models?
It is a 305B multimodal model that significantly enhances "vision-to-action" capabilities. It scored 36.5 on the ApexBench Pass@1, proving it can handle complex visual reasoning much better than the V4-Flash-0731.
Why is Nvidia investing $3.5 billion in MediaTek?
The goal is to integrate MediaTek's silicon with Nvidia's NVLink Fusion. This allows for the creation of custom "AI factories" where different types of processors can work together seamlessly at rack-scale.
What is the "Jalapeño" chip and why does it matter?
Jalapeño is OpenAI's custom inference chip. It matters because it reduces dependency on Nvidia and provides superior performance-per-watt, which is the single biggest cost driver for running LLMs at scale.
What are the risks of the "infostealer" malware affecting Claude?
The malware doesn't attack Anthropic's servers; it steals "session cookies" from users' browsers. This lets attackers bypass passwords and MFA to use the user's account and usage limits.
What is a "neural operator" in physics AI?
Unlike Transformers, which use attention mechanisms to find patterns in sequences, neural operators learn the underlying mathematical operators of a physical system. This allows them to handle massive amounts of data without the quadratic memory cost of Transformers.
Sources: aiweekly.co, Bloomberg, SemiAnalysis, The Guardian