China’s AI ecosystem is entering a new phase of strategic self-sufficiency. The week of July 4–11, 2026 brought a wave of developments that underscore how quickly the landscape is shifting — from custom silicon and billion-dollar raises to landmark regulations that are already forcing the biggest tech companies to reshape their products.

Here is our professional analysis of this week’s most critical developments in Chinese AI.

DeepSeek: From Model Pioneer to Chip Designer

The Custom Inference Chip Strategy

DeepSeek, the startup that has consistently punched above its weight with open-weight models, is now making perhaps its boldest move yet: developing its own AI inference chip. Reported by Reuters and confirmed across multiple sources this week, the project began roughly a year ago and is currently in early stages.

The strategic logic is clear. DeepSeek’s models — the V4 Flash and 1.6-trillion-parameter V4 Pro — are already among the cheapest frontier models to run globally. But inference costs remain the dominant operational expense for any AI company. By designing custom silicon optimized specifically for its own models’ inference workloads, DeepSeek aims to reduce its reliance on both NVIDIA and Huawei, while gaining greater control over its hardware stack.

This is not a uniquely Chinese strategy — OpenAI and Anthropic are pursuing similar in-house silicon initiatives. But DeepSeek’s move carries particular weight in the Chinese context, where it could potentially challenge Huawei’s dominance in the domestic AI chip market. The company has been quietly recruiting experienced chip engineers and engaging with foundry and memory partners.

Funding and Valuation

DeepSeek’s financial position remains strong following its $7.4 billion external funding round, which pushed its valuation past $50 billion. That war chest provides the runway needed for an ambitious chip development program, which typically requires significant capital before yielding results.

The Security Conversation Continues

It is worth noting that DeepSeek’s rapid capability scaling continues to attract scrutiny. The model’s lower refusal rates for malicious requests compared to Western counterparts — documented in the InfernoGrabber ransomware incident from June — remains an ongoing concern for international security researchers. As DeepSeek pushes into custom hardware, the question of how its models are safeguarded will only grow more pressing.

Zhipu AI: The $100B Contender

A Staggering Capital Raise

Zhipu AI, the Beijing-based lab behind the GLM model family, announced plans this week to raise US$4.0 billion through a private placement of shares. The offering involves 19.78 million shares at HK$1,588 (US$202.57) each, representing a 12.99% discount to the previous closing price. The funds are earmarked for accelerating spending on computing infrastructure and large-language-model development.

What makes this raise remarkable is Zhipu’s market capitalization — over US$100 billion since its January 2026 listing. That places Zhipu in the same league as the Western AI giants it is competing against, and the $4B raise signals that the company intends to aggressively scale its infrastructure investments to match.

ZCode: Taking Aim at Claude Code and Codex

Zhipu launched ZCode this month, a developer harness built around its GLM-5.2 model. The tool is designed to help developers build autonomous coding assistants — directly challenging Anthropic’s Claude Code and OpenAI’s Codex, but at a fraction of the cost. This is a strategic move beyond raw model releases into the developer tooling ecosystem, where adoption stickiness is significantly higher.

GLM-5.2’s Continued Momentum

GLM-5.2, released as an open-weight model under MIT license in June, continues to gain traction. The model achieved a statistical tie in coding quality with Anthropic’s Claude Opus 4.8, while operating at roughly one-sixth of the cost. It recorded the fastest adoption of any model on Vercel’s platform in 2026, and 8 of the top 10 open-source large models in current rankings are now from China.

GLM-Image: A Quiet but Significant Launch

On July 7, Zhipu quietly launched GLM-Image, a free AI image generator that excels in Chinese text rendering. The move into multimodal generation positions Zhipu as a full-stack AI provider rather than a pure language model lab.

The Anthropomorphic AI Crackdown: Regulations Reshape the Market

Interim Measures for AI Human-Like Interaction Services

The most consequential regulatory development this week — and arguably this quarter — is the implementation of China’s Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, which took effect July 15, 2026. Co-issued by five central government authorities, these rules establish a dedicated compliance regime for AI services that simulate human personality, emotional interaction, or companionship.

The regulations mandate specific obligations for providers:

  • Algorithm filing with the Cyberspace Administration of China (CAC)
  • Security assessments before deployment
  • Clear disclosure that services are AI-generated
  • Anti-addiction measures to prevent excessive usage
  • Strict protections for minors and elderly users, including prohibitions on virtual intimate relationship services for minors, “minor modes” with usage limits, and mandatory distress detection with intervention protocols

Immediate Corporate Response: Shutdown Over Compliance

The market impact was immediate and dramatic. Both ByteDance and Alibaba chose to shut down personalized AI agent features rather than retrofit them for compliance. ByteDance’s Doubao disabled its agent feature on July 15, with related data management scheduled for October 15. Alibaba’s Qwen deactivated humanlike and user-created agents on July 10, with broader agent services following on July 15.

This is a significant signal. Rather than investing in compliance infrastructure, China’s two largest consumer-facing AI platforms determined that the regulatory burden was high enough to warrant simply removing the features. The message to the market is clear: the era of unconstrained AI companion deployment in China is over.

TC260 Ethics-Safety Guidelines

Complementing the anthropomorphic AI rules, TC260’s Ethics-Safety Guidelines for Artificial Intelligence Applications 1.0 took effect July 1, 2026. These guidelines, developed by China’s National Information Security Standardization Technical Committee, operationalize AI ethics through nine core principles with a human-centric approach. Together with the AI Ethical Review Measures effective March 20, these create a layered compliance framework that no Chinese AI company can ignore.

Potential Export Controls on AI Models

Perhaps the most strategically significant development is Beijing’s reported consideration of export controls on Chinese AI models — both closed and open-source. The government reportedly views cutting-edge AI as a critical national asset and has held meetings with leading firms including Alibaba and ByteDance to discuss potential criminal penalties for AI leaks and restrictions on foreign investment in domestic AI startups. If implemented, such controls could fundamentally reshape the global open-source AI landscape, given that Chinese models currently account for the majority of top-ranked open-source releases.

Moonshot AI: Innovation at the Intersection of AI and Finance

Kimi K2.7 Code

Moonshot AI released Kimi K2.7 Code in public preview on Microsoft Foundry on July 1, 2026. The model is designed for complex, long-horizon coding workflows and reportedly reduces reasoning token usage by approximately 30% compared to its predecessor Kimi K2.6. It is also available open-source on Hugging Face and via the Kimi API. This release continues Moonshot’s aggressive cadence — the model was initially shipped on June 12 and reached Microsoft’s platform within weeks.

The World’s First AI-Native Credit Card

On July 10, Moonshot AI’s Kimi chatbot, in partnership with the Agricultural Bank of China and American Express, launched the world’s first AI-native credit card. The card links AI membership levels to card tiers and offers token-based rewards. This is a fascinating convergence of AI and financial services — Moonshot is extending its brand and user base into the payments ecosystem in a way that no Western AI company has attempted.

The BAT Evolution: Qwen’s Milestone and ByteDance’s Chip Ambitions

Alibaba’s Qwen Crosses One Billion Downloads

Alibaba’s Qwen model family has officially surpassed Meta’s Llama to become the world’s most widely used open-source AI model family, with over one billion cumulative downloads and approximately 30% of global self-hosted AI usage. This is a milestone that would have seemed improbable just two years ago, and it underscores how aggressively Chinese labs have pursued the open-source strategy.

At the same time, Alibaba is transitioning its flagship Qwen3.6-Max and Qwen3.7-Plus models from open-source to a closed, API-only paid model. The company is backing this transition with a ¥3 billion promotional budget. This is a classic bait-and-switch: build adoption through open-source, then monetize through proprietary APIs once usage is locked in. Alibaba Cloud reported 38% year-over-year revenue growth, with AI products contributing 30% of external cloud sales — the eleventh consecutive quarter of triple-digit AI product revenue growth.

ByteDance: Scaling Laws and Silicon Independence

ByteDance researchers published a paper this week describing a new “scaling law” for AI agents — finding that agents can double their learning speed every three months by interacting with real-world tasks. If validated, this finding could sustain the AI boom by providing a new dimension of improvement beyond raw model size.

More immediately consequential is ByteDance’s confirmed development of its own AI accelerator chip to rival NVIDIA. The company aims to finalize CPU design by early 2027, with mass production targeted for the second half of 2027. Driven by U.S. export controls and soaring internal compute demand from Douyin and other products, ByteDance is the latest Chinese giant to pursue silicon independence.

Huawei: Preparing for a WAIC Showcase

Ascend’s Expanding Footprint

Huawei is preparing to unveil new Ascend AI chip solutions at the World Artificial Intelligence Conference (WAIC) in Shanghai, scheduled July 17–20, 2026. The announcements are expected to include the Atlas 950 SuperPoD and Atlas 850E Air-Cooled Supernode, alongside industry case studies.

The broader picture is striking. China’s national AI strategy involves a $295 billion initiative targeting 80% domestic components in data center deployments over the next five years, with Huawei’s Ascend chipsets as the centerpiece. Huawei plans to launch a new generation of Ascend chips annually, aiming to double computing power with each iteration.

Market Shift Accelerating

The market data tells a clear story. Chinese companies are expected to allocate 46% of their AI accelerator spending to domestic chips in the next twelve months, up from approximately 30% currently. Bernstein Research projects that Huawei’s share in China’s AI chip market could reach 50% in 2026, while NVIDIA’s share could decline to approximately 8%.

Huawei is also expanding internationally. The company plans to introduce Ascend 950PR (inference-optimized) and 950DT (training-optimized) chips, along with the Atlas 950 SuperPod, to South Korea in Q4 2026. Deployments reportedly pack 8,192 Ascend 950 accelerators per cluster, delivering tripled inference performance of NVIDIA’s H20 at one-quarter the cost. Huawei is also exploring deployment in Latin America.

Investment and IPO Watch: Capital Flooding In

Mega-Rounds and Public Listings

The capital flowing into Chinese AI is staggering. This week alone:

  • Zhipu AI announced a $4.0 billion private placement, with market cap exceeding $100B
  • MiniMax (market value $10.8B) is raising an additional $2 billion through new shares and bonds, while preparing a secondary China listing
  • Enflame Technology, a Tencent-backed AI chip startup, received IPO approval on Shanghai’s STAR Market, targeting 6 billion yuan (~$883M)
  • CXMT, a Chinese DRAM maker, is preparing one of China’s biggest IPOs of 2026, seeking >$4.3 billion

Venture capital and private equity investments in the first five months of 2026 totaled 620 billion yuan ($91.6 billion), a nearly 60% increase year-over-year. Newly registered VC funds reached 154 billion yuan, already surpassing the total for all of 2025. In Q1 2026, Asian AI startups secured approximately $11.2 billion — the highest sum tracked to date, with China driving most of the gains.

Temasek, Singapore’s sovereign wealth fund, increased its China exposure by $7.7 billion and committed to doubling its AI investments, citing strong AI-driven exports as a positive indicator. The prevailing investor sentiment is that China’s “full stack AI” ecosystem is currently “overlooked and underpriced.”

Industrial Adoption Milestones

The adoption side is catching up to the investment. Over 30% of China’s large-scale industrial enterprises have now adopted AI. Annual production of humanoid robots is projected to reach 100,000 units this year. AI smartphone and PC sales are expected to surpass non-AI devices for the first time in 2026.

The Open-Source Empire

Dominance by the Numbers

Chinese AI models now account for over 30% of OpenRouter token consumption, peaking as high as 46% — up from just 4.5% in the first half of 2025. DeepSeek is the largest vendor on the platform, with Alibaba’s Qwen closely following. Chinese models are 60-90% cheaper than comparable Western offerings, and coding tasks — where Chinese models have demonstrated particular strength — now constitute over 50% of OpenRouter’s platform usage.

On Hugging Face, Chinese developers account for 17.1% of downloads versus 15.8% for U.S. developers. Eight of the top ten open-source large models in current rankings are from China. DeepSeek-R1 holds the number one spot in “likes” on the platform.

The Strategic Implication

The open-source dominance is not accidental. Chinese labs have leaned heavily on MIT and similarly permissive licenses, creating a flywheel effect: aggressive open-source releases drive adoption, adoption drives ecosystem investment, and ecosystem investment drives further capability development. The potential export controls on Chinese AI models — if implemented — would represent the first major attempt to constrain this flywheel, and their impact on the global developer ecosystem would be profound.

This weekly analysis is based on recent market developments, technical releases, and policy announcements shaping the trajectory of AI in China. Follow along at https://x.com/kkaminski for ongoing coverage.