Alibaba Raises 80 Billion HKD to Fund China's AI Infrastructure Push

Trend One: China's AI Infrastructure — From Chips to Cloud, One Unified Push

On August 24, Xiaomi unveiled three self-developed chips in a single event. The Xuanjie O3 is a 3nm AI flagship SoC that scored 5.22 million on AnTuTu — the first mobile chip to break the 5 million mark. The Xuanjie O100 is a 6nm 3D wafer-level stacked AI accelerator with 1.22TB/s bandwidth. The Xuanjie D100 is China's first 3nm autonomous driving chip. Lei Jun disclosed that Xiaomi's large-chip R&D has consumed over 21 billion yuan across five-plus years, with a team of nearly 3,000. The Xiaomi 18 Fold will be the first device powered by the O3, launching in September. CCTV News covered the announcement with the headline "Three Chips Launched — Another Breakthrough for China's Chip Industry."

Also on August 24, Alibaba completed pricing for an 80 billion HKD share placement, drawing nearly 3x oversubscription — over 200 billion HKD in orders. Sovereign and long-only funds accounted for over 40% of allocations, with sovereign wealth funds from the Middle East, Europe, and Asia participating. The proceeds will be entirely directed toward full-stack AI capabilities and infrastructure. Following the placement, Chairman Joe Tsai and CEO Eddie Wu together purchased 120 million HKD in Alibaba stock.

On the same day, ByteDance completed a reorganization of its office AI products. The TRAE and Coze teams have been merged into the Doubao ecosystem. TRAE IDE and CLI will continue as coding product lines under the Doubao brand. A standalone AI office product called "Doubao Work" is expected to launch within the week. ByteDance stated the restructuring aims to better coordinate product and technology resources.

In the open-source ecosystem, Hugging Face is exploring a sale at a valuation of 13 billion dollars or higher, nearly triple its 4.5 billion dollar valuation from 2023. The platform now hosts over 2 million models, 1.5 million datasets, and 1.5 million AI applications.

These four moves, unfolding on a single day, map the full chain of China's AI infrastructure buildout: foundational silicon (Xiaomi), cloud and compute capital (Alibaba), application-layer consolidation (ByteDance), and the global open-source backbone (Hugging Face). The scale of Alibaba's placement — 100% earmarked for AI — and the sovereign fund participation signal that AI infrastructure is now viewed as a national strategic asset class.

Trend Two: The Model Race Enters a Multimodal, Multi-Front Phase

On August 21, DeepSeek quietly launched V4-Flash-Vision-Exp, its first multimodal vision model. The model matches V4-Flash on text tasks while approaching Claude Opus 4.8 on multimodal agent benchmarks. Images are capped at 384 tokens regardless of resolution, and off-peak pricing is 0.22 dollars per million tokens — identical to the text-only V4-Flash. The model supports JPEG, PNG, GIF, and WebP, with a Files API for multi-request image reuse.

On August 23, a mysterious model called Ox Alpha appeared on OpenRouter, described as "a reasoning model designed for coding, sustained agentic work, and production workload." Stripe CEO Patrick Collison called it "very impressive." The provider has chosen to remain anonymous during the preview period. Speculation has ranged from Zhipu's unreleased GLM to a stealth Microsoft MAI variant, but no source has been confirmed. The model is currently free to use.

Arena's Week 34 blind rankings show Zhipu's newly released glm-5.3-max debuting at 13th overall with an ELO of 1487, while also cracking the top 10 in the coding leaderboard. Moonshot's kimi-k3-max rose to 10th overall, its first entry into the top 10. In the agent leaderboard, Kimi K3 Max reached 4th place with a net improvement rate of 10.41% and a task confirmation success rate of 17.97%, entering the first tier alongside Anthropic models.

Alibaba Cloud also released Wan 3.0, a video generation model supporting native 30-second video and multimodal reference inputs including text, images, audio, video, and documents. It has launched on Leonardo AI, Segmind, SeaArt, Scenario, and FlovaAI, with a 50% discount during the launch period. On the Arena video leaderboard, ByteDance's dreamina-seedance-2.5-720p debuted at 4th place.

The model landscape is fragmenting along capability axes: DeepSeek is competing on multimodal price-performance, Ox Alpha is a wildcard of unknown origin, domestic Chinese models are steadily climbing Arena rankings across text, code, frontend, image, video, and agent categories, and Wan 3.0 pushes into native long-form video generation. No single model or company dominates all categories.

Trend Three: The Price-Performance Reckoning

Anthropic's annualized revenue reached 65 billion dollars in July, up from 47 billion in May. The company told investors it expects Q3 profitability and has 6,000 customers spending 100,000 dollars or more annually. Yet the Ramp AI Index — based on billing data from 70,000 companies — shows that Fable 5, Anthropic's flagship model, accounted for only 3.5% of Anthropic model spending in July. Its paid token usage was just 6%. Opus 4.8 led at 28%, followed by Sonnet 4.6 at 8.3% and Fable 5 at 8%. The 10 dollars per million token input price appears to be pushing enterprises toward cheaper alternatives within Anthropic's own lineup.

On the open-source side, open-weight models' token share on Vercel's AI Gateway surged from 28.4% in June to 62% in August — a doubling in two months — steadily taking share from OpenAI and Anthropic. The cost of open-source inference is approaching frontier-model levels.

Meanwhile, NVIDIA is reportedly in talks to invest in Perplexity at a valuation exceeding 30 billion dollars, over 50% higher than its previous round. Perplexity's annualized revenue has grown from 250 million to over 750 million dollars, partly driven by its AI agent "Perplexity Computer."

The tension is clear: Anthropic is generating massive revenue but its most expensive model is losing ground to cheaper alternatives — both its own and open-source. NVIDIA is betting on the application layer. The AI industry's center of gravity is shifting from "who has the smartest model" to "who delivers the best performance per dollar."

Trend Four: Agents Move from Demo to Default

Shopify quietly enabled WebMCP tools by default for every Liquid storefront, giving approximately 5 million stores an agent-callable interface for catalog search, cart management, and checkout — with no installation required. The tools run on the shopper's live session and trigger the store's real UI. Custom store actions still require manual registration, but the standard commerce flow is now agent-ready out of the box. Combined with Cloudflare's similar edge-level move weeks earlier, "agent-ready" has shifted from a research topic to infrastructure.

a16z data on Codex enterprise weekly active growth shows extreme sector divergence: lawyers at 108x, sales and recruiting at 41x each, marketing at 26x, healthcare at 24x, and engineers at just 5x — the lowest. Codex has evolved from code completion into what a16z describes as a "digital colleague" capable of autonomous task decomposition.

Xiaopeng's robotics unit completed its first funding round of over 900 million dollars at a post-money valuation exceeding 6.3 billion dollars — a new record for China's embodied AI private equity financing. IDG Capital led the round, with Tencent and Alibaba as strategic investors. The IRON humanoid robot is scheduled to enter mass production by end of 2026, starting with Xiaopeng's own stores and campuses, with market launch in China and overseas in 2027. Separately, at the second World Humanoid Robot Games, the Tiangong team won the 400-meter large-format final in 38.15 seconds, with the top three all breaking the human world record. In the 100-meter preliminaries, a robot clocked 9.39 seconds, surpassing Usain Bolt's record — though most robots crashed into the wall after the finish line.

On the safety front, OpenAI Chief Global Affairs Officer Chris Lehane warned on August 23 that frontier AI models are now capable of planning and executing sophisticated cyberattacks. OpenAI has suspended development of some of its most advanced internal models. Separately, South Korea's National Assembly passed an amendment to the Personal Information Protection Act, allowing AI developers to use personal data within defined limits after review by the Personal Information Protection Commission. Twitch faces a class-action lawsuit in California for using streamer content — including broadcasts, clips, chat logs, and photos — to train Amazon's AI models without consent. The platform admitted to the practice on August 12 and made the opt-out non-retroactive.