West Tide — Overseas AI News Column under Physix Frontier Daily curated global AI industry updates, bringing you the latest firsthand insights from abroad.
🔥 Today's Big Events
Figure AI Valuation Soars to $150 Billion; F.04 Next-Gen Robot Enters Pre-Production Phase
Figure AI founder Brett Adcock announced that the company's latest valuation has reached $150 billion (approx. 1.1 trillion RMB), nearly quadrupling from its previous round of $40 billion. Concurrently, the next-generation F.04 humanoid robot has completed design lock-in and entered the component delivery and pre-production phase, which Adcock called "the company's biggest technological leap." The Figure 03 has already completed pilots at BMW auto factories, with the BotQ factory achieving a monthly capacity of 240 units. The "overtake" of 750 robots vs. 180-250 human employees has become a global talking point. Morgan Stanley raised China's full-year humanoid robot shipment forecast from 14,000 units to 50,000 units. Figure AI may launch an IPO in 2027.
Source: Figure AI Official / AIbase / CSDN
🧠 AI Large Models
1. Apple Sues OpenAI for Trade Secrets: 41-Page Complaint Details "Mass Exodus" of 40 Former Employees
On July 12, Apple filed a 41-page complaint with the U.S. District Court for the Northern District of California, accusing OpenAI and its hardware subsidiary io Products of stealing trade secrets through systematic poaching of Apple employees, accessing internal documents, and contacting suppliers. Bloomberg's Mark Gurman noted that the complaint did not name former Apple Chief Designer Jony Ive (now a partner with io Products), likely due to multiple legal strategy considerations. Over 40 former Apple employees have already joined the OpenAI ecosystem. Apple is requesting a court injunction and damages.
Source: PitchBook / Phoenix Tech / Bloomberg
2. Elon Musk's Lawsuit Against OpenAI Dismissed by Court
Several lawsuits previously filed by Elon Musk against OpenAI have been dismissed by the court. The court determined that Musk's claims lacked sufficient legal basis, and since OpenAI had transitioned from a non-profit to a hybrid public-benefit/profit structure, the original contractual obligations had been redefined. Musk's side stated they are considering an appeal. This marks the latest twist in the 18-month legal battle between Musk and OpenAI.
Source: TechCrunch / Reuters
3. Grok 4.5: xAI Challenges Claude Opus at 1/4 the Cost
xAI released the Grok 4.5 model, which approaches the performance level of Claude Opus 4.8 in multiple benchmark tests but with inference costs only 1/4 as high. Grok 4.5 is specifically optimized for Cursor IDE programming scenarios—reportedly trained using Cursor programming tracking data. It scored 1558 on the Arena leaderboard, ranking sixth globally. API pricing is set at $21/M input + $105/M output, directly challenging Anthropic's enterprise customer base.
Source: AIbase / FourWeekMBA / xAI
4. Anthropic Discovers Claude's "Hidden Thinking Space" J-space
Anthropic's research team discovered a previously unidentified "hidden thinking space" (J-space) within Claude models—the model performs additional internal reasoning steps before generating output, which do not appear in the final result. This discovery is significant for model safety and interpretability research. Anthropic stated it will enhance monitoring capabilities for J-space in future versions to ensure model behavior aligns with stated intentions.
Source: Anthropic Blog / AI D-A-M-N
5. GPT-Red Adversarial Training System: OpenAI Releases New Safety Alignment Solution
OpenAI released the GPT-Red (Generative Pre-trained Red-teaming) adversarial training system, designed to improve model safety through automated red-teaming. The system can automatically generate thousands of adversarial input scenarios, covering attack vectors such as jailbreaking, data leakage, and harmful content generation. GPT-Red has been applied to pre-release safety assessments for the GPT-5.6 series models and will gradually be opened to enterprise API users.
Source: OpenAI Blog / TechCrunch
6. Top 5 on Arena Leaderboard Swept by Anthropic
The latest comprehensive leaderboard from Arena.ai shows that Anthropic's Claude series models occupy all top 5 positions: Claude Fable 5 (#1, 1631 points on coding leaderboard), Claude Mythos 5 (#2), Claude Opus 4.8 (#3), Claude Sonnet 5 (#4), and Claude Fable 5 Lite (#5). Anthropic's annualized revenue has exceeded $47 billion, surpassing OpenAI's $25-33 billion. This dominant performance solidifies Anthropic's status as the strongest current model lab.
Source: Arena.ai / FourWeekMBA
7. Kimi K3 Open Source Model Breaks into Coding Leaderboard: Chinese Model Takes Global #1 for First Time
Moonshot AI's Kimi K3 (2.8 trillion parameter MoE), released on 7/16, topped the Arena Frontend Code Arena leaderboard with 1679 points—surpassing Claude Fable 5 (1631) and GPT-5.6 Sol (1618). This is the first time a Chinese open-source model has taken first place in a major global coding arena. It won first place in 6 out of 7 frontend sub-domains, with an average win rate of 76%. Full weights will be released on 7/27. Musk commented "Impressive." API pricing is only 30% of Fable 5.
Source: National Business Daily / FourWeekMBA / Arena.ai
💻 AI Software & Applications
8. AI Agents Impact SaaS Model: Gartner Predicts 20% of Spending Affected
Gartner released a report predicting that by 2027, 20% of enterprise SaaS spending will be impacted by AI Agents. Traditional seat-based SaaS models face challenges from "Service-as-Software" (outcome-based billing). A significant portion of Anthropic's $47 billion annualized revenue comes from enterprise-grade Agent products like Claude Cowork. PitchBook analysis indicates that enterprise software is transitioning from "seat-based SaaS" to "outcome-based Service-as-Software."
Source: Gartner / PitchBook / Axios
9. Codex Message Encryption Raises Enterprise Audit Concerns
The message encryption mechanism of OpenAI's Codex programming tool has raised audit concerns among enterprise IT departments. Corporate security teams pointed out that Codex's end-to-end encryption makes code auditing and compliance checks difficult, especially in heavily regulated industries like finance and healthcare. OpenAI responded that it will launch an enterprise-level audit API, allowing administrators to view Codex operation logs without breaking encryption.
Source: TechCrunch / AI D-A-M-N
10. Microsoft Sales Team Pushes Custom Copilot Versions to AI Labs
Microsoft's sales team is targeting major AI labs with custom versions of Microsoft 365 Copilot, including exclusive workflow solutions for labs like Anthropic, OpenAI, and Google DeepMind. The July update for Copilot added Vision features (real-time screen sharing analysis) and Outlook AI email editing. Teams Channel Agent has also entered public preview.
Source: Windows Forum / Microsoft / AI D-A-M-N
11. EU Orders ChatGPT Return to WhatsApp
The European Data Protection Board required OpenAI to restore ChatGPT services on WhatsApp, after OpenAI temporarily removed the feature due to GDPR compliance issues. The EU requires OpenAI to submit a remediation plan within 30 days, ensuring user data processing in the WhatsApp integration complies with the legal basis under Article 6 of GDPR. This decision will affect millions of European users.
Source: EU Data Protection Board / TechCrunch
12. Codex Weekly Active Users Exceed 8 Million
Weekly active users of OpenAI's Codex programming tool surpassed 8 million, a 60% increase from 5 million in May. Codex has become the world's largest AI programming assistant platform, surpassing GitHub Copilot's 6.5 million weekly actives. After OpenAI acquired Ona (formerly Gitpod), Codex gained persistent cloud agent capabilities—tasks continue executing in the cloud even after users close their laptops.
Source: AIToolsRecap / OpenAI
13. Gemini Managed Agents Officially Launched
Google launched Gemini Managed Agents hosting service, allowing enterprises to deploy and manage AI Agents on Google Cloud without building their own infrastructure. It supports multi-step workflow automation, tool calling, database access, etc. Pricing is based on Agent runtime, competing with Anthropic Claude Cowork and OpenAI ChatGPT Work. Alphabet stock rebounded 2% following the announcement.
Source: Google Cloud / Bloomberg
🤖 Humanoid Robots
14. Figure AI Valued at $150 Billion: F.04 Pre-Production Starts
(See Today's Big Events) Figure AI's valuation soared from $40 billion to $150 billion. The next-gen F.04 enters pre-production, BotQ produces 240 units/month, and 750 robots are already deployed. Adcock calls F.04 "the company's biggest technological leap."
Source: Figure AI Official / AIbase
15. Agility Robotics Advances IPO: Strong Secondary Market Performance Post-SPAC Merger
After going public via SPAC at a $2.5 billion valuation (ticker AGLT), Agility Robotics has seen strong secondary market performance. Digit robots have accumulated over 65,000 operating hours and are deployed at Amazon, GXO, and Toyota. The company is evaluating the possibility of a direct IPO to achieve a higher valuation. NVIDIA Halos safety framework selected Digit as its first commercial deployment platform.
Source: TechFastForward / Yahoo Finance
16. Tesla Optimus Target of 1,000 Units/Month: Fremont Factory Renovation Complete
Tesla's Fremont factory in California has removed the Model S/X production lines and renovated them into dedicated Optimus humanoid robot lines in just 46 days. Third-generation Optimus production starts in July-August, aiming for 2,000 units/week (nearly 10,000/month) by year-end. Long-term planning targets 1 million units/year. Tesla Chips (AI5) + FSD + Optimus + Dojo constitute a complete physical AI closed loop.
Source: EEVblog / IDN Times
17. Boston Dynamics Atlas Does Continuous Backflips: Deepest Motion Control IP in Industry
Boston Dynamics released the latest demo video of the Atlas robot, showcasing continuous backflips, front flips, and side rolls. Atlas possesses the deepest accumulation of motion control IP in the industry. Hyundai Motor has completed the full acquisition of Boston Dynamics and joined the NVIDIA Cosmos alliance on 7/15. The Atlas IPO window is expected in 2027-2028.
Source: Boston Dynamics / Andrew.ooo
18. 1X NEO Home Version Priced at $20,000
Norwegian company 1X Technologies released the NEO home humanoid robot, priced at $20,000 (approx. 145,000 RMB), targeting household scenarios. NEO uses a wheeled chassis + dual-arm manipulation, capable of completing chores like dishwashing, folding clothes, and carrying items. IDTechEx predicts the average price of humanoid robots will drop from $35,000 in 2025 to a range of $17,000-$37,000 by 2030.
Source: 1X Technologies / IDTechEx
19. Global Humanoid Robot Shipments Increase 5x Year-over-Year
Morgan Stanley significantly raised China's 2026 humanoid robot shipment forecast from 14,000 units to 50,000 units (3x+). Globally, H1 2026 humanoid robot shipments increased approximately 5x year-over-year. Drivers include cost reduction (some platforms as low as $28,000), IDTechEx assessment that some high-load logistics scenarios have achieved a 6-month ROI, and dense deployment in Chinese manufacturing.
Source: Morgan Stanley / IDTechEx / IFR
🚗 Autonomous Driving
20. Waymo Expands to Four Cities: Las Vegas Achieves Fully Driverless Operation
In addition to Los Angeles, San Francisco, and Phoenix, Waymo added Las Vegas operations, with the Las Vegas route already achieving fully driverless operation (no safety drivers). Waymo's sixth-generation Driver system hardware has been significantly streamlined (cameras 29→13, LiDAR 5→4), relying on DeepMind Genie3 world models to shift from rule-driven to generative intelligence.
Source: Waymo Official / Carscoops
21. California Approves Waymo Bay Area Expansion
The California Public Utilities Commission (CPUC) approved Waymo's expansion of Robotaxi services to the entire San Francisco Bay Area, including suburbs like Oakland, Berkeley, and Palo Alto. This is the largest geographic operational permit Waymo has received to date. Waymo has cumulatively completed over 50 million paid rides.
Source: CPUC / TechCrunch
22. Waymo London Test Fleet Expanded to 100 Vehicles
Waymo expanded its London test fleet from 50 to 100 Jaguar I-PACE vehicles, covering more London urban roads. The UK Department for Transport issued an extended testing license to Waymo. Waymo plans to launch commercial operations in London in 2027, which would be its first commercial market outside the US.
Source: Waymo UK / Reuters
23. Dallas + Avis Partnership: Rental Giant Operates Waymo Fleet
Waymo partnered with Avis Budget Group, where Avis will handle daily operations for the Waymo Robotaxi fleet in the Dallas area—including charging, cleaning, and basic maintenance. This is the first time the autonomous driving industry has established an operational partnership with a major rental car company. Dallas becomes Waymo's fifth US operational city.
Source: Avis Budget Group / Waymo
24. Charlotte Ojai Unmanned Vehicle Road Testing Begins
Waymo started road testing of unmanned Ojai vehicles in Charlotte (with safety drivers) to prepare for future commercial operations. The Ojai is a custom pure-electric autonomous vehicle for Waymo, manufactured by Hyundai/Kia. Charlotte will become Waymo's sixth US operational city.
Source: Waymo / Charlotte Observer
25. Tesla FSD v14.3.5: AI Compiler Rewrite Speeds Up by 20%
Tesla released FSD v14.3.5, with core changes involving a complete rewrite of the underlying inference engine using self-developed AI compilers, improving inference speed by 20%. The new version also optimized perception in adverse weather and construction zone identification. Cumulative FSD mileage exceeded 9 billion miles. The Cybercab production version has started public road testing in Austin, completely removing steering wheels and pedals.
Source: EEVblog / Electrek
🌐 World Models & Physical AI
26. DeepMind × A24 Film Collaboration: AI-Assisted Movie Production
Google DeepMind partnered with renowned independent film production company A24 to apply AI technology to the movie production process—including scene generation, visual effects previews, and script assistance. This is not DeepMind's first venture into creative fields, but the collaboration with A24 marks the shift of AI+Film from experimentation to commercialization.
Source: Google DeepMind / Variety
27. ATL Saathi Educational Tool: AI-Assisted STEM Learning
Indian battery giant ATL launched the Saathi AI educational tool, providing personalized learning assistance for STEM students. Based on customized large language models, Saathi transforms complex physics and chemistry concepts into interactive learning experiences. ATL plans to deploy Saathi in 10,000 schools across India.
Source: ATL Official / TechCrunch India
28. Da Xiao Kairos 3.1: World Model Achieving SOTA in 12 Benchmarks
Chinese world model company Da Xiao Robot released Kairos 3.1, achieving SOTA (State-of-the-Art) results in 12 benchmark tests. Kairos 3.1 supports high-fidelity physical simulation, long-sequence prediction, and multimodal interaction. The product is positioned as "the brain of robots"—providing a unified cognitive and simulation foundation for physical AI.
Source: Da Xiao Robot Official / WAIC
29. Sora Upgraded to 4K: Leap in OpenAI Video Generation Quality
OpenAI upgraded the Sora video generation model to 4K resolution output, while supporting up to 60 seconds of continuous generation. The new version significantly improved physical consistency, lighting effects, and character expressions. Sora 4K is now open to ChatGPT Plus and Enterprise users. Competitors Kuaishou Kling AI and Runway Gen-4 are also iterating rapidly.
Source: OpenAI / TechCrunch
30. NVIDIA Cosmos Alliance Continues to Expand
The NVIDIA Cosmos world foundation model alliance continues to expand its membership. Hyundai/Kia joined on 7/15, with Boston Dynamics indirectly connecting via Hyundai Motor. Cosmos provides a unified simulation and training platform for physical AI, with Hitachi integrating IWIM domain knowledge with Cosmos. The goal is to enable robots to learn real-world physical laws in virtual environments.
Source: Andrew.ooo / NVIDIA
31. Meta Muse Image: Open Source Image Generation Tool
Meta released the Muse Image open-source image generation tool, built on Llama vision models. It supports text-to-image, image editing, and style transfer. Muse Image weights are fully open source and can run on local GPUs. This is Meta's latest move to compete with DALL·E and Midjourney in the image generation field.
Source: Meta AI / Hugging Face
🌊 West Tide — Overseas AI News Column under Physix Frontier
Bringing Western tides eastward, firsthand frontier news.
© 2026 Physix Frontier. All rights reserved.