Physix FrontierWestTide · AI Industry Morning Brief | July 22, 2026
WestTide · AI Industry Morning Brief | Tuesday, July 22, 2026
Physix Frontier · WestTide Column — Daily highlights from the global Physical AI and Large Model race
🔥 Headlines
1. Kimi secretly preparing for HK IPO, pre-money valuation at $31.5 billion
Moonshot AI has sent listing proposals to investors, expecting to land on the HKEX as early as within 6 months. It is currently completing a funding round with a pre-money valuation of approximately $31.5 billion, after which it will immediately launch a Pre-IPO round (targeting a $50 billion valuation). ARR has surpassed $300 million, with API revenue accounting for over 70%. The company is dismantling its VIE structure to transition to H-shares, with CICC and Goldman Sachs serving as joint sponsors.
Source: 21st Century Business Herald / Bloomberg
2. AMD invests $5 billion in Anthropic; latter purchases 2GW of compute
AMD and Anthropic have signed a strategic agreement: Anthropic will purchase up to 2GW of AMD Instinct MI450 GPUs (Helios rack-level solution) starting in H1 2027, while AMD will make an equity investment of up to $5 billion in Anthropic. Both parties will also initiate multi-year engineering collaboration, with AMD fully adopting Claude to optimize its own chip and software development. The total transaction scale reaches tens of billions of dollars.
Source: AMD Newsroom / WSJ / CoinDesk
3. OpenAI agents hit 10M weekly active users, ChatGPT Work ignites enterprise sector
OpenAI's agent products (Codex + ChatGPT Work) have surpassed 10 million weekly active users, less than two weeks after ChatGPT Work launched. Usage grew by approximately 2.5x following the release of GPT-5.6, with over 1 million users applying Codex to non-software development scenarios. Zapier, RingCentral, Virgin Atlantic, Nvidia, etc., have become the first batch of enterprise clients.
Source: Bloomberg / TNW
🧠 Large Models
4. GPT-5.6 fully open, Sol/Terra/Luna tiers benchmark against Anthropic
The three tiers of OpenAI's GPT-5.6 series (Sol flagship / Terra balanced / Luna lightweight) are fully online, directly benchmarking against Anthropic's Fable 5 / Sonnet 5 / Haiku 4. Sol ranks second only to Kimi K3 in AI Arena Code, while Terra has become the default driving model for ChatGPT Work.
Source: OpenAI / Artificial Analysis
5. Google releases three Gemini efficiency models, flagship 3.5 Pro continues to delay
Google released three models in one go: Gemini 3.6 Flash (coding improvement + 17% reduction in token consumption), 3.5 Flash-Lite (ultra-fast at 350 tokens/s), and 3.5 Flash Cyber (dedicated to cybersecurity). However, there is still no release date for the market-anticipated flagship Gemini 3.5 Pro—originally planned for June launch, it has been delayed by several weeks due to subpar coding capabilities. Google revealed for the first time that pre-training for Gemini 4 has begun.
Source: Google Blog / 36Kr / Bloomberg
6. Huawei Xiaoyi and Xiaohongshu LLMs win gold with perfect scores at IMO 2026
Huawei's Xiaoyi AI won gold with a perfect score of 42 points at the IMO 2026 (International Mathematical Olympiad), becoming the first non-human contestant to achieve a perfect score at the IMO. The system is based on the openPangu 2.0 Pro model, utilizing a "Star Orbit Architecture" hybrid reasoning system, completing structured proofs across 17 logical anchors within 217 seconds. Related capabilities will be opened to consumers by the end of July.
Source: IT Home / Phoenix Net
7. Moonshot AI's ARR triples in 3 months, Kimi K3 compute strained, new users paused
Within 48 hours of Kimi K3's release, user request volume approached the cluster's capacity limit, causing Moonshot AI to pause new C-end subscriber sign-ups. ARR surged from $100 million in March to $300 million by mid-June. K3 features 2.8 trillion parameters and a 1 million Token context window, topping the Code Arena with a score of 1679.
Source: STAR Market Daily / Cailian Press
8. Microsoft evaluating integrating Kimi K3 into Copilot, potentially cutting inference costs by $600M annually
Microsoft is evaluating the possibility of introducing Kimi K3 for certain inference tasks in Copilot. K3 is priced at $15 per million output tokens, matching Sonnet 5 but with higher inference efficiency. If implemented, it is expected to save Microsoft approximately $600 million in annual inference costs. This marks the first time a Chinese open-source model enters the supply chain evaluation for Microsoft's core products.
Source: Reuters / Goldman Sachs Research
9. First year of large-scale edge AI deployment, Minitron Intelligence valuation breaks 20B RMB
Minitron Intelligence's MiniCPM series of edge-side models have exceeded 50 million deployments across smartphones, car systems, and IoT devices. Following its latest funding round, the company's valuation broke through 20 billion RMB. Edge inference is moving from "usable" to "easy-to-use," with 2026 defined as the inaugural year for large-scale edge AI implementation.
Source: 36Kr / Investment Community
10. AI safety firm RealAI completes hundreds of millions RMB Series B
RealAI completed a Series B funding round of hundreds of millions of RMB, focusing on AI safety, deepfake detection, and large model security assessment. As AI applications accelerate penetration, total financing in the AI safety sector in H1 2026 has already exceeded 5 billion RMB.
Source: Investment Community
💻 Software & Applications
11. ChatGPT Work redefines "AI Office": Not just answering, but completing
Since ChatGPT Work launched, users are no longer just asking questions but letting AI break down tasks, execute across applications, and generate reports/slides/documents. It connects Slack and Google Drive, can schedule recurring tasks, and controls desktop software via Computer Use. Pro/Enterprise/Edu subscribers can already use it.
Source: TNW / OpenAI Blog
12. 75% of Google's code is generated by AI
Google revealed that 75% of its internal code is now generated by AI, entering production environments only after engineer review. This proportion has significantly increased from 50% at the start of the year, making it one of the highest cases of AI coding penetration among major global tech companies.
Source: Google / eWeek
13. Kimi K3 tops Frontend Code Arena with 1679 points
In the Arena Frontend Code competition, Kimi K3 scored 1679 points, surpassing Claude Fable 5 and GPT-5.6 Sol, marking a milestone for open-source models in the coding domain. Musk commented "Impressive" under the evaluation, subsequently revealing that xAI is training Grok 4.6 with 2 trillion parameters.
Source: Artificial Analysis / X
14. DeepSWE Test: Gemini 3.6 Flash scores 49%, leading 3.5 Flash
Google's newly released Gemini 3.6 Flash scored 49% in the DeepSWE test (compared to 37% for 3.5 Flash), but still lags behind models like Meta Spark 1.1 and GLM-5.2 in the Artificial Analysis comprehensive ranking. The developer community gave substantial negative feedback regarding its frontend generation capabilities.
Source: 36Kr / Artificial Analysis
15. Zhipu AI's Tang Jie responds to "Is GLM still viable?": Epic level plus
Facing netizen doubts like "Qwen and Kimi have evolved epically, is GLM still viable?", Zhipu founder Tang Jie responded "Epic level plus." Since its IPO, Zhipu's stock price has soared by over 1000%, with market cap briefly breaking 1 trillion HKD, as it accelerates next-generation model R&D.
Source: Weibo / Sina Finance
🤖 Robotics
16. Humanoid robot financing exceeds 104.1 billion RMB in H1 2026, body manufacturer IPO wave accelerates
According to incomplete statistics, global humanoid robot sector financing in H1 2026 broke through 104.1 billion RMB, with body manufacturers accounting for over 60%. Leading firms like Unitree and Agibot have started IPO preparations, indicating the industry is shifting from "financing-driven" to "mass production verification" stages.
Source: GGII Robotics / Investment Community
17. Shihe Robot completes hundreds of millions RMB Series C
High-altitude work robot firm Shihe Robot completed a Series C round of hundreds of millions of RMB, with funds allocated for product iteration and overseas market expansion. Its wall-climbing robots have been applied in wind turbine blade inspection, ship rust removal, etc., with cumulative shipments exceeding 1000 units.
Source: Investment Community
18. AI materials discovery firm CuspAI secures $450 million funding
UK-based AI new material discovery company CuspAI completed a $450 million funding round, using deep learning to accelerate new material R&D, having achieved commercialization in areas like battery separators and catalysts. AI + Materials Science is becoming a new investment hotspot.
Source: TechCrunch
19. LimX Dynamics completes $200 million Pre-IPO, valuation at 15 billion RMB
Humanoid robot company LimX Dynamics completed a $200 million Pre-IPO round, reaching a valuation of 15 billion RMB. Its full-size bipedal robots have completed hundred-unit deployments in industrial inspection scenarios, making it one of the largest single financings in the domestic robotics sector in 2026.
Source: 36Kr / Investment Community
20. Europe's first humanoid robot unicorn: Humanoid completes $152 million Series A
Dutch humanoid robot startup Humanoid completed a $152 million Series A round, becoming Europe's first humanoid robot unicorn. Its robots have entered trial use in logistics sorting scenarios, with per-unit costs controlled within $50,000.
Source: TechCrunch / Sifted
🚗 Autonomous Driving
21. Tesla FSD China entry in regulatory approval sprint, expected Q3 approval
Tesla's FSD (Full Self-Driving) entry into China has entered the final regulatory approval stage, with approval expected in Q3 2026. Initially compatible only with HW4.0 hardware vehicles, the buyout price is approx. 64,000 RMB or a subscription of 499-699 RMB/month. If approved, it will become the first foreign intelligent driving system of L2+ level implemented in China.
Source: Sina Finance / Autohome
22. Apollo Go globalization accelerates: Swiss road tests begin, Dubai already commercially operating
Baidu's Apollo Go globalization is speeding up: Road tests in Switzerland officially began, while Dubai has achieved fully unmanned commercial operation (in partnership with AutoGo), with cumulative operating mileage exceeding 330 million km. Additionally, Apollo Go has entered Kazakhstan, marking the first time Chinese driverless cars have entered Central Asia.
Source: Baidu / Smart Car Tech
23. Horizon Robotics deepens AI cooperation with VW, L3 delivery as early as H2 2027
Horizon Robotics and Volkswagen Group are deepening their AI chip cooperation. Based on Journey 6 series chips, the L3 autonomous driving solution could be delivered as early as H2 2027. VW has become Horizon's largest single customer in Europe, with cumulative cooperation value exceeding 10 billion RMB.
Source: 36Kr / Autohome
24. Pony.ai × JD Auto Service reach national strategic partnership
Pony.ai and JD Auto Service signed a national strategic cooperation agreement, leveraging JD Auto Service's high-density offline maintenance network to provide scaled maintenance support for Robotaxi fleets in Beijing, Shanghai, Guangzhou, Shenzhen, and multiple new first-tier cities. The 7th generation Robotaxi has taken the lead in achieving positive unit economics in first-tier cities.
Source: Pony.ai / JD
25. Lyft CEO personally tests Apollo Go: Exclaims "Amazing"
During a business trip to China, US mobility giant Lyft's CEO personally experienced Apollo Go's driverless taxi, posting "Amazing" repeatedly on social media. Both parties are exploring possibilities for cooperation in overseas markets.
Source: Weibo / 36Kr
🌐 World Models & Infrastructure
26. Anthropic annualized revenue exceeds $47 billion, secretly filed for IPO in June
Anthropic's annualized revenue has broken through $47 billion (a massive increase from $10 billion in all of 2025), and it secretly filed an IPO prospectus with the SEC in June. Valued at $965 billion, infrastructure partners already include Google, Amazon, SpaceX, and AMD.
Source: CoinDesk / MarketWatch
27. SpaceX rents out Colossus 1 data center compute, monthly rent $1.25 billion
Elon Musk's SpaceX has opened a new business line, renting out idle compute power from the Colossus 1 data center externally. Anthropic has signed a contract for full rental, with a monthly fee of $1.25 billion and a contract term until May 2029.
Source: WSJ / Bloomberg
28. Google starts "Gemini 4" pre-training, calls it "most ambitious yet"
While releasing the Flash series, Google formally revealed for the first time that pre-training for the next-generation flagship Gemini 4 has begun, calling it the "most ambitious pre-training project to date." This move is seen as a strategic response to the delay of 3.5 Pro.
Source: Google DeepMind / CNBC
29. HK IPO fundraising in H1 2026 hits 210.2 billion HKD, up 92% YoY
HKEX data shows 87 new listings in H1 2026, with IPO fundraising reaching 210.2 billion HKD (+92% YoY). Chapter 18C has opened listing channels for unprofitable specialized technology companies, with AI large model enterprises being the biggest beneficiaries.
Source: HKEX / Sina Finance
30. Zhipu AI stock soars 1000% post-IPO, market cap briefly breaks 1 trillion HKD
As the "world's first large model stock," Zhipu AI's share price surged over 1000% after listing, with market cap briefly breaking 1 trillion HKD. Meanwhile, MiniMax retraced about 75% from a peak of approx. 410 billion HKD, showing clear divergence between the two "pioneers."
Source: HK Stocks / Sina Finance
31. xAI training 2 trillion parameter Grok 4.6, Musk says "may surpass Kimi"
After commenting "Impressive" under the Kimi K3 evaluation, Musk revealed that xAI is training the Grok 4.6 model with 2 trillion parameters, stating it "has the potential to surpass Kimi K3." The AI arms race is accelerating on a monthly basis.
Source: X / Bloomberg
Today's Industry Judgments
1. Chinese large models enter a new phase of "capital validation." The valuation jump from $31.5B to $50B for Moonshot AI, Zhipu's 1000% post-IPO surge, and DeepSeek stunning the world with cost efficiency—Chinese AI is moving from a "technology race" to "secondary market pricing." However, valuation sustainability remains to be verified: The structural contradiction between C-end user churn vs. B-end API growth needs answers from the prospectus.
2. Compute arms race upgrades to "infrastructure alliances." AMD-Anthropic's $5B investment + 2GW compute purchase, combined with SpaceX-Colossus, Google-Broadcom, and other multi-party collaborations, indicates AI infrastructure is evolving from a "buyer-seller relationship" to a deep alliance model of "equity binding + engineering collaboration."
3. Open-source models surpass closed-source for the first time, battle for pricing power begins. Kimi K3 topped the Code Arena as an open-source model, priced at $15/million tokens matching Sonnet 5—Goldman Sachs states this marks Chinese AI moving from "price wars" to contending for "pricing power."
4. Google trapped in "version number anxiety." Two delays for 3.5 Pro, Flash series criticized as "worst ever," Gemini 4 pre-training starting—Google is caught between efficiency models and flagship models, with developer reputation eroding.
5. Humanoid robot sector financing breaks 100 billion, mass production verification period arrives. H1 financing of 104.1 billion RMB, LimX Dynamics' $200M Pre-IPO, birth of Europe's first unicorn—H2 2026 will be the critical turning point from "PPT → Production Line."
Physix Frontier · WestTide | Daily AI industry signals, driven by passion.
🌊 West Tide — Overseas AI news column under Physix Frontier
West tide guides East, frontier firsthand.
© 2026 Physix Frontier