1. In 2024, Microsoft received $493.8 million in revenue share payments from OpenAI, increasing to $865.8 million in the first three quarters of 2025.
2. OpenAI reportedly shares 20% of its revenue with Microsoft, based on a prior $13 billion investment deal.
3. Microsoft also returns about 20% of Bing and Azure OpenAI Service revenues to OpenAI, but these are deducted from Microsoft's net revenue share figures.
4. OpenAI’s revenue was at least $2.5 billion in 2024 and $4.33 billion in the first three quarters of 2025, with previous reports estimating $4 billion for 2024 and $4.3 billion for the first half of 2025.
5. Sam Altman stated OpenAI’s annualized revenue run rate will exceed $20 billion by year-end 2025, with potential to reach $100 billion by 2027.
6. OpenAI’s inference spend was approximately $3.8 billion in 2024 and $8.65 billion in the first nine months of 2025.
7. OpenAI’s total compute spend was about $5.6 billion in 2024, with a “cost of revenue” of $2.5 billion for the first half of 2025.
8. Inference spend is largely cash-based, while training spend is mostly non-cash, covered by Microsoft-awarded credits.
9. OpenAI may be spending more on inference costs than it is earning in revenue, raising concerns about profitability amid high valuations in the AI sector.
10. OpenAI relies primarily on Microsoft Azure for compute, with additional deals involving CoreWeave, Oracle, AWS, and Google Cloud.

1. Google announced Gemini 3 on November 18, 2025, positioning it as a major step toward AGI with advanced reasoning and nuanced understanding.
2. Gemini 3 is being deployed across Google Search (AI Mode), the Gemini app, AI Studio, Vertex AI, and the new agent-focused platform, Google Antigravity.
3. Gemini 3 Pro is available in preview and is integrated into Search, Workspace, the Gemini app, and developer platforms.
4. Gemini 3 Pro surpasses Gemini 2.5 Pro, OpenAI GPT-5.1, and Claude Sonnet 4.5 on benchmarks such as LMArena, Humanity’s Last Exam, GPQA Diamond, and MathArena Apex.
5. Multimodal capability scores include 81% on MMMU-Pro, 87.6% on Video-MMMU, and 72.1% on SimpleQA Verified for factual accuracy.
6. Gemini 3 Deep Think, a new reasoning mode, achieves 41% on Humanity’s Last Exam, 93.8% on GPQA Diamond, and 45.1% on ARC-AGI-2 with code execution.
7. Gemini 3 features broader multimodal input, longer context windows, and enhanced planning, enabling applications such as research analysis, handwritten translation, visualisation generation, and sports evaluation.
8. AI Mode in Search now includes generative UI elements and interactive simulations.
9. Google Antigravity, launched for developers, enables agentic workflows with autonomous planning and execution of complex software tasks, integrating with tools like AI Studio, Vertex AI, Gemini CLI, Cursor, GitHub, JetBrains, and Replit.
10. Gemini 3 Pro leads the Vending-Bench 2 leaderboard for long-horizon planning, maintaining consistent decision-making over a simulated year.
11. Gemini Agent, offering agentic capabilities, is available to Google AI Ultra subscribers via the Gemini app.
12. Gemini 3 underwent the most extensive safety testing to date, including external evaluations by Apollo, Vaultis, and Dreadnode.
13. Security enhancements include reduced sycophancy, improved prompt-injection resistance, and stronger misuse protection.

1. Perplexity launched its Comet AI browser on Android today, following its July desktop debut.
2. The Android version includes most desktop features: default AI search engine, tab-based assistant queries, voice mode for tab questions, and cross-tab search summarization.
3. Comet browser can autonomously research and shop, provides transparency on assistant actions, and includes a built-in ad blocker.
4. Upcoming Android features in the next weeks include a cross-site conversational agent, assistant shortcuts for quick actions, and a full password manager.
5. Earlier this month, Comet Assistant on desktop was updated for more complex, long-running tasks such as transferring website data to spreadsheets.
6. iOS launch is planned, but Android was prioritized due to high demand from carriers and OEMs, though no new partnerships were announced.
7. Perplexity previously partnered with Motorola to preload the app, but no confirmation on pre-installing the Comet browser.
8. Competing AI browsers from OpenAI, Opera, and The Browser Company (now Atlassian) have mainly focused on desktop, with limited mobile presence.
9. Security concerns have been raised about vulnerabilities in AI agents, with Perplexity acknowledging in October that new AI-driven attack paradigms require fundamental security rethinking.

1. NVIDIA Apollo, a family of open AI physics models for industrial and computational engineering, was introduced at SC25 in St. Louis.
2. Industry leaders including Applied Materials, Cadence, LAM Research, Luminary Cloud, KLA, PhysicsX, Rescale, Siemens, and Synopsys are adopting Apollo models for simulation and design acceleration in automotive, aerospace, and other sectors.
3. Apollo models are optimized for scalability, performance, and accuracy across domains such as electronic device automation, semiconductors, structural mechanics, weather and climate, computational fluid dynamics, electromagnetics, and multiphysics.
4. The models leverage advanced machine learning architectures including neural operators, transformers, and diffusion methods, combined with domain-specific knowledge.
5. Pretrained checkpoints and reference workflows for training, inference, and benchmarking will be provided, enabling customization and integration.
6. Applied Materials achieved up to 35x acceleration in ACE+ multi-physics software modules using NVIDIA GPUs and CUDA, enabling near-real-time modeling of semiconductor process chambers with AI surrogates and digital twins.
7. Cadence used the NVIDIA-powered Millennium M2000 Supercomputer and Fidelity Charles Solver to train an AI physics model for real-time digital twins of full aircraft, showcased in October 2025 at NVIDIA GTC Washington, D.C.
8. LAM Research is collaborating with NVIDIA to accelerate plasma reactor simulation for semiconductor manufacturing using AI physics.
9. KLA will use Apollo models to accelerate simulations and enhance semiconductor process control solutions.
10. Northrop Grumman and Luminary Cloud use NVIDIA AI physics models and CUDA-X libraries to accelerate spacecraft thruster nozzle design, enabling rapid exploration of thousands of designs.
11. PhysicsX’s AI-native platform integrates with NVIDIA AI physics infrastructure and Siemens Simcenter X, reducing product development cycles for automotive, aerospace, and energy sectors.
12. Rescale is integrating Apollo models into its AI physics operating system, enabling real-time inference and exploration of vast design spaces with high-fidelity and high-speed AI surrogates.
13. Siemens is embedding NVIDIA AI physics into Simcenter STAR-CCM+, allowing designers to explore design options orders of magnitude faster by blending first-principles simulations with AI surrogates.
14. Synopsys is using NVIDIA AI physics to achieve up to 500x speedups in computational engineering, with GPU-accelerated fluid simulation tools like Ansys Fluent benefiting from AI surrogate initialization.
15. NVIDIA Apollo models will be available soon on build.nvidia.com, HuggingFace, and as NVIDIA NIM microservices.

1. Grok 4.1, released by xAI, is positioned as the most capable Grok model to date, emphasizing significant improvements in creative, emotional, and collaborative interactions, as well as reliability.
2. Internal and external testing shows Grok 4.1 delivers substantial accuracy gains in math, coding, and advanced reasoning, with improved handling of longer, complex prompts and better step-by-step reasoning consistency.
3. Grok 4.1 demonstrates a dramatic reduction in hallucination rates and improved latency, resulting in more grounded and faster responses.
4. In blind, real-traffic comparisons, Grok 4.1 achieves a 64.78% win-rate over previous Grok versions.
5. Grok 4.1 ranks #1 on the LMArena Text Leaderboard (Thinking Mode) with an Elo of 1483, outperforming Gemini 2.5 Pro, Claude Sonnet/Opus, and GPT-4.5 preview models.
6. Both Grok 4.1 variants lead the EQ-Bench for emotional intelligence, surpassing Kimi K2 Instruct, Horizon Alpha, Gemini 2.5 Pro, GPT-5 Chat, and Claude Opus 4.
7. Grok 4.1 is among the top models in Creative Writing v3, outperforming o3, Claude Sonnet 4.5, and Grok 3.
8. FactScore benchmarks indicate major improvements in factual precision, especially for biography-oriented tasks.
9. Grok 4.1 is immediately available to all users via grok.com, X, iOS, Android, and Tesla Auto mode, with explicit selection as “Grok 4.1” in the model picker.
10. Pricing includes a Basic plan (limited access), SuperGrok (Rs 700/month) with extended features and 128K token memory, and SuperGrok Heavy ($300/month) with 256K token memory, unlimited Grok 3 usage, and early feature access.
11. System-level upgrades in Grok 4.1 include enhanced training efficiency, longer-context handling, improved tool-use reliability, cleaner data pipelines, and stronger training signals.
12. Hands-on tests show Grok 4.1 excels in text generation and coding, provides logical reasoning, but underperforms in image and video generation compared to leading tools.
13. Grok 4.1 marks a shift for xAI from a personality-driven model to a stable, consistent, and competitive LLM suitable for real-world applications.
14. As of November 19, 2025, Grok 4.1 positions xAI as a serious contender against OpenAI and Anthropic in the AI race.

1. In the early 2010s, Elon Musk identified the risk of AI being dominated by profit-driven entities and co-founded OpenAI in 2015 to prioritize human benefit over shareholder profit.
2. As of November 2025, OpenAI is valued between $500 billion and $750 billion, with its for-profit arm structured as a public benefit corporation.
3. Musk now operates his own for-profit AI company, xAI, while OpenAI and other major AI players are interconnected through complex, profit-oriented partnerships.
4. The current AI ecosystem, termed "the Blob," consists of interlocking partnerships, mergers, funding arrangements, and government initiatives, involving both US and overseas interests.
5. The Stargate initiative exemplifies these arrangements, linking OpenAI, Oracle, Nvidia, Softbank, an Abu Dhabi investment firm, and the US government.
6. A recent deal involves Microsoft investing at least $5 billion in Anthropic, Anthropic committing to purchase $30 billion in compute from Microsoft Azure, and Nvidia investing in Anthropic, which will use Nvidia chips for development.
7. Anthropic's valuation surged from $183 billion two months ago to $350 billion following these deals.
8. Anthropic, lacking its own cloud or non-AI revenue, now has stock-for-compute deals with Microsoft, Amazon, and Google.
9. Nvidia, Microsoft, and Anthropic describe their relationship as mutually reinforcing, with each company acting as both customer and partner to the others, aiming to proliferate AI solutions globally.

1. Gemini AI assistant has officially replaced Google Assistant on Android Auto as of November 2025, with global rollout beginning today.
2. Gemini enables natural language conversations and can be activated via "Hey Google," the car screen mic button, or long-pressing the steering wheel voice control.
3. Users can add local stops using Gemini, leveraging real-time business insights, reviews, and FAQs, with follow-up queries supported.
4. Messaging capabilities include summarizing recent texts, dictating new messages, chaining requests, and translating messages.
5. Gemini's live mode allows users to learn, brainstorm, and rehearse meetings while driving, though it cannot be interrupted mid-sentence unlike the mobile version.
6. Music requests are enhanced, allowing users to generate custom playlists based on mood, duration, or specific themes.
7. Productivity features include hands-free access to email summaries, Google Calendar, and navigation to reservations.
8. Gemini's integration with Android Auto is considered a major upgrade over Google Assistant, with future potential for proactive safety features.

1. Xiaomi launched and fully open-sourced its embodied foundation model MiMo-Embodied on November 23, 2025.
2. MiMo-Embodied unifies autonomous driving and embodied AI under a single modeling framework, enabling cross-domain intelligence.
3. The model supports affordance reasoning, task planning, spatial understanding (embodied AI), and environment perception, state prediction, driving planning (autonomous driving) for full-scenario applications.
4. MiMo-Embodied demonstrates two-way knowledge transfer, with capabilities learned in indoor robotics enhancing outdoor driving intelligence and vice versa.
5. The model employs a multi-stage training pipeline: skill learning, chain-of-thought inference enhancement, and fine-grained RL optimization for end-to-end reliability.
6. MiMo-Embodied achieved state-of-the-art results in 17 embodied AI benchmarks and breakthroughs in 12 autonomous driving benchmarks, surpassing both open and closed systems.
7. The model shows stronger generalization and major gains in vision-language multimodal benchmarks.
8. MiMo-Embodied and its source code are available on Hugging Face and ArXiv.

1. Tesla published detailed performance and safety data for its Full Self-Driving (Supervised) software in North America.
2. FSD users reportedly drive 5 million miles per major collision and 1.5 million miles per minor collision, compared to NHTSA averages of 699,000 miles and 229,000 miles, respectively.
3. Tesla’s new data claims FSD users travel 2.9 million miles between major collisions and 986,000 miles between minor collisions, versus NHTSA’s 505,000 and 178,000 miles, respectively.
4. Tesla defines major collisions as those involving airbag or non-reversible pyrotechnic restraint deployment, per 49 C.F.R. § 563.5.
5. Collisions are counted if FSD was active within five seconds before impact, including disengagement or system aborts.
6. Tesla will update this data quarterly, using a rolling twelve-month aggregation.
7. Tesla will not release injury rates, focusing instead on collision frequency and airbag deployment as proxies for severity.
8. Previous Tesla safety reports were criticized for focusing on Autopilot rather than FSD and lacking detail.
9. Waymo claims its vehicles are 5x safer than human drivers and 12x safer regarding pedestrians, with more transparent data reporting.
10. Waymo’s co-CEO Tekedra Mawakana emphasized the industry’s responsibility for transparency in safety data at TechCrunch Disrupt in October 2025.

1. "Qwen Panic" describes Silicon Valley's anxiety over Alibaba's rapid AI progress, notably the open-source Qwen models and Qwen App, which directly compete with ChatGPT.
2. Alibaba's Qwen models, released within a year, are now widely adopted by U.S. developers and major tech firms due to their power and low cost.
3. On November 14, 2025, the Financial Times reported a White House memorandum alleging Alibaba's technology supported PLA-linked operations targeting U.S. interests, causing Alibaba's U.S.-listed shares to drop 4.2% within hours.
4. Qwen's open-source strategy has eroded Silicon Valley's traditional AI advantages, with companies like Amazon and Apple integrating Qwen into key products and services.
5. Stanford researchers built the S1 reasoning model on Qwen2.5-32B for under $50, and the Allen Institute for AI and Mira Murati's lab use Qwen for multimodal and fine-tuning tasks.
6. Alibaba's Qwen App anchors a full-stack AI ecosystem, integrating cloud, foundation models, consumer apps, and real-world services across e-commerce, logistics, finance, and more.
7. Alibaba's ability to embed AI across its vast digital ecosystem contrasts with ChatGPT's stand-alone nature, intensifying U.S. concerns.
8. The ATOM Project, launched by U.S. industry leaders, aims to restore American open-source AI leadership through a dedicated lab, funding, and a roadmap to match or surpass Qwen and DeepSeek.
9. Nvidia CEO Jensen Huang acknowledged Chinese open models' lead and emphasized the strategic importance of openness, prompting OpenAI to prepare its first open-source release in years.
10. Qwen's low cost exposes the fragility of the U.S. AI economic model, as developers can train and deploy high-performance models for a fraction of U.S. costs.
11. Airbnb's CEO confirmed heavy reliance on Qwen for its new AI agent, citing speed and cost advantages over OpenAI, while AWS now offers Qwen models on Amazon Bedrock.
12. The AI race is shifting from model intelligence to ecosystem integration and adoption speed, with Alibaba positioned to deploy AI at scale across commerce, finance, and logistics.
13. U.S. policymakers and CEOs now face the prospect of Chinese open, cheap, scalable, and ecosystem-integrated AI becoming the global standard.

1. China’s first large-scale flying car manufacturing facility has opened in Guangzhou, initiating domestic mass production of low-altitude economy products.
2. The factory will increase output after completing verification and airworthiness certification, moving from small-batch testing to scaled manufacturing.
3. XPENG AEROHT in Huangpu District has produced the first “Land Aircraft Carrier” flying vehicle from the world’s first flying car mass-production line.
4. The facility can manufacture up to 10,000 flying cars annually, releasing one aircraft every 30 minutes at full capacity.
5. EHang’s EH216-S autonomous passenger aircraft has secured the world’s first eVTOL type certificate, production certificate, and standard airworthiness certificate.
6. EHang recently conducted its first commercial demonstration flight of the EH216-S in Guangzhou Development Zone.

1. A consortium of six German research institutions and two startups launched the Soofi (Sovereign Open Source Foundation Models) project to develop an open-source AI language model with 100 billion parameters for European enterprises.
2. The German Federal Ministry for Economic Affairs and Energy is funding the project with €20 million until July 2026.
3. The project addresses the dependency of German companies on US and Chinese AI models and the lack of a sufficiently large European LLM for industry-specific customization.
4. Participants include Fraunhofer IAIS, Fraunhofer IIS, DFKI, the universities of Würzburg, Hannover, TU Darmstadt, the Berlin University of Applied Sciences, and startups Ellamind and Merantix Momentum, led by the KI Bundesverband.
5. In addition to the base LLM, a reasoning model for structured, multi-step problem solving and AI agent technologies for practical use cases will be developed.
6. Model training will occur on Deutsche Telekom's Industrial AI Cloud, provided by T-Systems as a sovereign AI infrastructure.
7. The final model will comply with the EU AI Act and align with European values, with funding via the 8ra initiative established by twelve EU member states.
8. The project aims to advance LLM research, strengthen European collaboration, retain AI talent in Germany, and contribute to a robust European AI ecosystem alongside initiatives like the Gigafactory and EU AI Action Plan.

1. Schwarz Gruppe is investing €11 billion to build Germany's largest data center in Lübbenau, Spreewald, with completion targeted for end of 2027.
2. The data center will cover 13 hectares and house over 100,000 next-generation GPUs for cloud, AI, and data applications.
3. The facility will be located on a brownfield site, repurposing an existing coal power plant, using 20,000 m³ of concrete and 110,000 tons of steel, and integrating waste heat into the district heating network to supply 75,000 households.
4. The project aims to enhance Germany's digital sovereignty and security, addressing hybrid threats such as sabotage and cybersecurity.
5. Schwarz Digits, the group's digital division, employs 10,000 IT staff supporting 600,000 employees and 14,200 stores across 32 countries.
6. The data center will primarily serve internal needs but also offers services to external clients, leveraging the group's proprietary STACKIT cloud, developed since 2017.
7. STACKIT was initiated after a foreign authority confiscated servers during a trademark dispute, highlighting the need for data localization and legal control.
8. The cloud stores the entire process chain, including product, pricing, and loyalty program data, ensuring compliance with German and European data protection regulations.
9. Waste heat from the data center will be recycled into district heating, with contracts already in place for ecological integration.
10. The managers emphasize the strategic importance of data self-determination, digital independence, and resilience in the face of geopolitical risks.
11. Germany's digital sovereignty is rated as "inadequate," with the need for pragmatic implementation of data protection as a competitive advantage.
12. The group cites daily cyberattacks and industrial espionage as key motivators for secure, local data storage.
13. The future vision includes leveraging quantum computing, with optimism about German startups leading in this field, and the synergy between AI and quantum computing as transformative for industry.

1. Generative AI adoption in corporate law departments in the US more than doubled from 23% in 2024 to 52% in 2025.
2. The proportion of in-house professionals experimenting with AI via beta products remained steady at 15% for both years.
3. Those in the research phase declined from 22% in 2024 to 17% in 2025.
4. Passive planners dropped from 30% in 2024 to 14% in 2025.
5. Respondents neither using nor planning to use AI fell sharply from 10% in 2024 to 2% in 2025.
6. 91% of in-house professionals cited efficiency as the primary benefit of AI adoption.
7. Corporate counsel are insourcing more legal work and reducing dependency on outside counsel due to AI.
8. Cost savings from outside counsel have not yet materialized despite increased AI use.
9. 24% of in-house legal teams are “very likely” to push for changes to the billable hour structure, considering value billing.
10. Many in-house legal teams lack budget for AI tools, training, and talent for complex tasks.
11. Reduced reliance on billable hours and alternative billing models such as fixed fees, value billing, or subscriptions are anticipated.
12. AI adoption introduces new compliance and risk considerations, necessitating governance frameworks for ethical and regulatory standards.
13. Survey results indicate change is occurring more rapidly than many legal experts expected.

1. President Donald Trump is considering an executive order to challenge state AI regulations via lawsuits and withholding federal funding.
2. The draft order directs Attorney General Pam Bondi to create an “AI Litigation Task Force” to sue states for AI regulations allegedly violating federal laws on free speech and interstate commerce.
3. The order, titled “Eliminating State Law Obstruction of National AI Policy,” could be signed as early as this week.
4. The AI Litigation Task Force will collaborate with White House technology advisors, including special adviser for AI and crypto David Sacks, to identify state law violations.
5. The order targets state laws, specifically in California and Colorado, that mandate AI developers to publish transparency reports and other requirements.
6. Big Tech trade groups, such as Chamber of Progress (backed by Andreessen Horowitz, Google, and OpenAI), oppose state-level AI regulation and advocate for federal, light-touch regulation.
7. A super PAC funded by Andreessen Horowitz, OpenAI cofounder Greg Brockman, and Palantir cofounder Joe Lonsdale is campaigning against New York Assembly member Alex Bores, who authored a state AI safety bill.
8. House Republicans are renewing efforts to pass a federal moratorium on state AI regulations after a previous attempt failed.