PHYSIX FRONTIER EASTVOICE
2026.07.23 · Thursday
Issue 002
FIVE MINUTES A DAY, THE GLOBAL AI SHIFT
$50B
Moonshot AI targets pre-IPO valuation for Hong Kong listing
$5B
AMD invests in Anthropic with 2GW chip deal reshaping infrastructure
~$200B
Alphabet market cap evaporated amid Google Gemini delays and damage
42/42
Huawei XiaoYi AI achieves perfect score at IMO 2026
Top Stories 3 · dual commentary ｜ Section Picks 29 · one-line takes
Moonshot AI targets $50B valuation, AMD invests $5B in Anthropic, and Google faces $200B market cap loss.
Moonshot AI (Kimi) plans to launch its final pre-IPO funding round in August, targeting a $50 billion pre-money valuation (~338.6 billion yuan), with a Hong Kong listing expected within 6 months. A current round at ~$31.5B pre-money is closing imminently. The company has raised over $3.9B in 2026 alone, with total funding exceeding 37.6 billion yuan. ARR has reached $300M, with API revenue comprising over 70%. VIE structure dismantling for H-share listing is underway.
Source: 21st Century Business Herald / China Economic Net
AMD and Anthropic formally signed their strategic partnership: Anthropic will deploy up to 2GW of AMD Instinct MI450 GPUs (Helios rack-scale), with the first gigawatt deploying in H1 2027. AMD commits up to $5B in equity — its first investment in Anthropic. A multi-year engineering collaboration will see AMD adopt Claude across its chip optimization and software development. AMD CEO Lisa Su: "We will establish Helios as a major platform for the next generation of AI infrastructure."
Source: AMD Newsroom / WSJ / MarketWatch
Google's anticipated Gemini 3.5 Pro, originally scheduled for June, has been delayed again due to coding benchmark shortfalls. Meanwhile, the released Gemini 3.6 Flash was lambasted by developers as "the worst model ever" — poor frontend generation and spatial reasoning. Google confirmed Gemini 4 pre-training has begun. Alphabet has lost approximately $200 billion in market cap over the past two weeks amid model delays and reputational damage.
Source: Bloomberg / 36Kr / eWeek
29 AI stories worth your time, each with a one-line take.
Huawei's XiaoYi AI agent scored a perfect 42 points at the 67th International Mathematical Olympiad, becoming the first non-human contestant to achieve full marks. The system uses a "StarTrack Architecture" — symbolic logic engine + openPangu 2.0 Pro deep cognition network + hierarchical abstraction verification hub. IMO President Prof. Gregor Dolinar confirmed the perfect score. Capabilities will be available to consumers via XiaoYi Brain and XiaoYi Claw by end of July.
What it means for you｜Capabilities will be available to consumers via XiaoYi Brain and XiaoYi Claw by end of July.
One-line take｜—
Source: IT Home / Phoenix News / Toutiao
OpenAI's Codex + ChatGPT Work combined weekly active users surpassed 10 million, nearly doubling from ~5 million in early July. Usage grew ~2.5x after the GPT-5.6 release. ChatGPT Work achieved this milestone in under two weeks. Enterprise clients include Zapier, RingCentral, Nvidia, with finance teams using agents for month-end reporting and forecasting automation.
What it means for you｜Enterprise clients include Zapier, RingCentral, Nvidia, with finance teams using agents for month-end reporting and forecasting automation.
One-line take｜—
Source: Bloomberg / TNW / NetEase
Microsoft is assessing Kimi K3 integration into select Copilot inference pipelines. At $15 per million output tokens (matching Sonnet 5), K3 offers superior inference efficiency. Full deployment could save Microsoft an estimated $600M annually. This marks the first Chinese open-source model to enter Microsoft's core product supply chain evaluation.
One-line take｜—
Source: Reuters / Goldman Sachs Research
Moonshot AI's funding timeline: 2023 angel round at $300M → 2024 A+ at $2.5B → 2025 C at $4.3B → May 2026 D at $20B → current Pre-IPO target $50B. A 166x valuation increase in 3 years and 3 months, making it China's most-funded LLM startup at 37.6 billion yuan total. National Social Security Fund and China Mobile have joined as investors.
One-line take｜—
Source: 21st Century Business Herald / Huafeng Capital
After commenting "Impressive" on Kimi K3 benchmarks, Elon Musk disclosed xAI is training Grok 4.6 with 2 trillion parameters. If successful, it would be the second-largest model after K3's 2.8T parameters. The AI arms race continues accelerating on a monthly cadence.
One-line take｜—
Source: X / Bloomberg
Anthropic's annualized revenue has surpassed $47 billion (up from ~$10B in all of 2025). The company confidentially filed its IPO prospectus with the SEC in June at a $965B valuation. Infrastructure partners include Google, Amazon, SpaceX, TeraWulf, and AMD, with preliminary talks reportedly underway with Meta.
One-line take｜—
Source: CoinDesk / MarketWatch / WSJ
Google disclosed that 75% of internal code is now AI-generated and engineer-reviewed. It also officially confirmed Gemini 4 pre-training has begun, calling it "the most ambitious pre-training effort to date" — widely interpreted as a strategic response to two consecutive 3.5 Pro delays.
One-line take｜—
Source: Google / eWeek / CNBC
ChatGPT Work is now available to Pro/Enterprise/Edu subscribers, enabling multi-step task decomposition across applications (Slack, Google Drive), desktop software control via Computer Use, and recurring task scheduling. Enterprise admins can set spending caps, tool permissions, and compliance review.
What it means for you｜ChatGPT Work enables multi-step task decomposition across applications, desktop software control, and recurring task scheduling for subscribers.
One-line take｜—
Source: TNW / OpenAI Blog
After going live on Vertex AI, Gemini 3.6 Flash was slammed by developers — losing to Claude Fable 5, GPT-5.6 Sol, and even Meta Spark 1.1 in frontend code arenas. Output showed illogical component nesting and broken interaction patterns. Spatial reasoning accuracy regressed significantly compared to 3.5 Flash.
What it means for you｜Output showed illogical component nesting and broken interaction patterns, with spatial reasoning accuracy regressing significantly compared to previous versions.
One-line take｜—
Source: 36Kr / Artificial Analysis
Flash-Lite delivers 350 output tokens/sec (~2x previous generation) at $0.30/$2.50 per million tokens (input/output). SWE-Bench Pro score of 54.2% exceeds standard 3.5 Flash's 49.6%, but developers note "fast wrong answers = efficiently wasting time."
What it means for you｜Developers note that fast wrong answers efficiently waste time despite the speed increase and higher benchmark scores.
One-line take｜—
Source: Google Blog / 36Kr
Google launched Gemini 3.5 Flash Cyber for vulnerability discovery, verification, and remediation. It found 55 issues in the V8 JavaScript Engine (10 undiscovered by other models). Due to dual-use concerns, it's available only via CodeMender invite-only program for government and trusted partners.
What it means for you｜Due to dual-use concerns, it is available only via CodeMender invite-only program for government and trusted partners.
One-line take｜—
Source: Google / 36Kr
Post-K3 launch demand exceeded cluster capacity, forcing Moonshot AI to pause C-end new subscriptions. ARR surged from $100M (March) to $300M (mid-June). K3 topped Code Arena at 1,679 points with 2.8 trillion parameters — the world's largest open-source model. Future membership will split Kimi core benefits and Kimi Code benefits.
What it means for you｜Post-K3 launch demand exceeded cluster capacity, forcing Moonshot AI to pause C-end new subscriptions temporarily.
One-line take｜—
Source: Sci-Tech Board Daily / Cailian Press
Responding to doubts about GLM's competitiveness vs. Kimi and Qwen, Zhipu AI founder Tang Jie replied "Epic Plus." Zhipu's stock has surged over 1,000% since listing, with market cap briefly exceeding HK$1 trillion.
One-line take｜—
Source: Weibo / Sina Finance
Dutch startup Humanoid closed a $152M Series A, becoming Europe's first humanoid robot unicorn. Its robots are piloting in logistics sorting with per-unit costs under $50,000. The company targets 1,000+ unit production by 2027.
One-line take｜—
Source: TechCrunch / Sifted
Humanoid robot company LimX Dynamics completed a $200M Pre-IPO round at ~15 billion yuan (~$2.1B) valuation. Its full-size bipedal robots have achieved 100+ unit deployment in industrial inspection, with funds earmarked for production scaling and overseas expansion.
One-line take｜—
Source: 36Kr / China Venture
Global humanoid robot funding exceeded 104.1 billion yuan in H1 2026, with body companies accounting for 60%+. Unitree and Zhiyuan Robotics have initiated IPO preparations, signaling the shift from "funding-driven" to "production-validation" phase.
One-line take｜—
Source: Gaogong Robot / China Venture
Shihe Robotics raised several hundred million yuan in Series C for wall-climbing robots used in wind turbine inspection and ship hull maintenance, with 1,000+ cumulative shipments.
One-line take｜—
Source: China Venture
UK-based CuspAI raised $450M for deep learning-accelerated materials R&D (battery separators, catalysts). AI + materials science is a rising investment theme.
One-line take｜—
Source: TechCrunch
AI safety firm RealAI completed a Series B round, focused on deepfake detection and LLM security evaluation. The AI safety sector attracted over 5 billion yuan in H1 2026 funding.
One-line take｜—
Source: China Venture
Horizon Robotics and Volkswagen deepened their collaboration, with L3 autonomous driving solutions based on Journey 6 series chips expected for H2 2027 delivery. VW is Horizon's largest European client, with cumulative cooperation exceeding 10 billion yuan.
One-line take｜—
Source: 36Kr / AutoHome
Pony.AI signed a national strategic partnership with JD Auto, leveraging JD's high-density maintenance network for Robotaxi fleet scaling across tier-1 and new tier-1 cities. Pony.AI's 7th-gen Robotaxi has achieved positive unit economics in tier-1 cities.
One-line take｜—
Source: Pony.AI / JD.com
Tesla FSD enters final regulatory approval for China launch, expected Q3 2026. Initial HW4.0 support at ~64,000 yuan buyout. Will directly compete with Huawei ADS and XPeng XNGP.
What it means for you｜Initial HW4.0 support at ~64,000 yuan buyout will directly compete with Huawei ADS and XPeng XNGP.
One-line take｜—
Source: Sina Finance / AutoHome
Baidu's Apollo Go is expanding on four fronts: Switzerland road tests launched, Dubai full driverless commercial operations (with AutoGo), Kazakhstan entry (first Chinese robotaxi in Central Asia), and cumulative operating mileage exceeding 330 million km. Lyft CEO tested it and called it "Amazing."
One-line take｜—
Source: Baidu / AutoTech / Weibo
Apollo Go partnered with Uber to deploy fully driverless taxis in Dubai. Dubai's Crown Prince tested the service and declared "the future is here." Apollo Go holds Dubai's sole full-driverless test permit — a global first.
One-line take｜—
Source: Baidu / UAE News Agency
Following the AMD-Anthropic deal, CEO Lisa Su stated: "Together, we will accelerate AI adoption at scale and establish Helios as a major platform for the next generation of AI infrastructure." AMD stock is up over 156% YTD in 2026.
One-line take｜—
Source: AMD Newsroom / MarketWatch
Anthropic's compute partners now span Google (TPU + $35B lease guarantees), Amazon (Trainium), SpaceX (Colossus 1 at $1.25B/month), AMD (MI450 + $5B investment), and TeraWulf, with preliminary Meta talks reportedly underway.
One-line take｜—
Source: WSJ / Bloomberg / CoinDesk
HKEX H1 data: 87 new listings raising HK$210.2B (+92% YoY). Chapter 18C enables pre-profit tech companies to list. Zhipu (+1,000%) and MiniMax (-75%) show stark divergence, providing lessons for upcoming AI IPOs.
One-line take｜—
Source: HKEX / Sina Finance
Mianbizhineng's MiniCPM on-device model surpassed 50 million device deployments, reaching a valuation exceeding 20 billion yuan (~$2.8B). 2026 is being defined as the "year of on-device AI at scale."
One-line take｜—
Source: 36Kr / China Venture
Moonshot AI paused C-end subscriptions due to K3 demand explosion; Anthropic previously capped user access due to surging demand. Both events reveal: efficient models don't reduce compute demand — they stimulate more applications through lower inference costs, creating a Jevons Paradox-style compute demand spiral. Physix Frontier · EastVoice | Daily AI industry signals. Because we love it. 🔔 East Voice — Physix Frontier's China AI column for global readers East Voice, West Reach. Chinese Perspective. © 2026 Physix Frontier
What it means for you｜Efficient models do not reduce compute demand but stimulate more applications through lower inference costs, creating a spiral.
One-line take｜—
Source: Sci-Tech Board Daily / CoinDesk
📖 The full EastVoice issue on the Chinese forum
📈 Physix Frontier · Alpha — the investor's read
Headline teardowns · curated signals · market moves. Five minutes to find the alpha of the AI era →