Understanding the rapidly evolving landscape of Chinese artificial intelligence models and their global implications

Based on Goldman Sachs equity research report "Navigating China Internet/AI Models" published May 4, 2026

What Is Happening in China's AI Industry?

China's artificial intelligence sector is experiencing explosive growth, with Chinese AI models narrowing the performance gap with leading US models while offering dramatically lower pricing. Daily token consumption in China exceeded 140 trillion in March 2026—a thousandfold increase from early 2024. This surge reflects both aggressive innovation in model efficiency and expanding enterprise adoption across sectors.

Major Chinese tech companies and independent AI startups are launching increasingly competitive foundation models, with particular strength in multi-modal capabilities (video, image, audio) and cost-efficient architectures designed to work around computing constraints.

Why Does This Matter Now?

The AI model landscape is fragmenting, with implications for:

• Global competitiveness: Chinese models are challenging US dominance in specific use cases
• Enterprise adoption: Businesses worldwide are evaluating more cost-effective alternatives
• Cloud infrastructure: Hyperscalers are racing to build capacity amid unprecedented token demand
• Geopolitical dynamics: Technology restrictions are accelerating domestic chip development and alternative architectures

The investment community is closely watching whether Chinese players can sustain innovation momentum, achieve profitability, and defend competitive positions in an increasingly crowded market.

Key Debate #1: Is the US-China AI Gap Narrowing or Widening?

The Current State

Recent benchmarks show Chinese models performing comparably to leading US models on many tasks, particularly excelling in:

• Pricing competitiveness (often 5-10x cheaper than US alternatives)
• Inference speed
• Specific task completion (especially agentic applications)
• Multi-modal generation (video, image, audio)

Notable achievements include ByteDance's Seedance 2.0 and Alibaba's Happy Horse delivering state-of-the-art performances in multi-modal generation.

The Uncertainty

Arguments for narrowing gap:

• Chinese models are achieving comparable benchmark scores on multiple evaluation frameworks
• Innovations in training efficiency are offsetting compute limitations
• Strong performance in multi-modal applications, with Chinese models ranking among global leaders
• Growing adoption signals real-world utility

Arguments for potential widening:

• US models have larger training budgets and significantly more advanced chip access
• Complex coding scenarios still dominated by US state-of-the-art models
• Recent assessments suggest faster US innovation velocity
• High-value feedback loops for coding remain stronger in the US ecosystem
• New anti-distillation tactics from leading US models could limit knowledge transfer

What It Means

Chinese AI companies are pursuing a distinctive path focused on architectural efficiency rather than raw computing power. The exceptionally tight computing capacity in China continues to drive a unique approach focusing on training/inference efficiencies, data quality and post-training, with innovative architectures that utilize less chips and memory.

This approach has produced impressive results but faces ongoing challenges in the most complex, high-value scenarios like sophisticated code generation, where the quality of feedback data remains a limiting factor.

Key Debate #2: Where Are the Competitive Moats in Chinese AI Models?

The Fragmentation Challenge

The rapid launch of new models has raised questions about market structure. Recent launches include:

• DeepSeek V4 (April 23, 2026): 1M context window requiring only 7-10% of predecessor's memory
• Tencent Hy3.0: Trained in under three months
• Xiaomi MiMo V2.5: Entry of hardware manufacturer into foundation models
• Alibaba Qwen3.6-Max: Continued iteration from cloud hyperscaler

This raises critical questions:

• Can hundreds of similar-sized parameter models coexist?
• What differentiates one 200-300 billion parameter model from another?
• Will pricing competition erode profitability?

Emerging Differentiation Factors

Technical capabilities:

• Coding excellence: Some models rank highest on third-party coding benchmarks
• Multi-modal integration: ByteDance, Alibaba and MiniMax are most focused on native multi-modal fusion
• Task completion rates: Success in real-world agent deployments (OpenClaw, Hermes AI agents)
• Long-context understanding: 1M+ token windows becoming standard

Business model advantages:

• Independent players: High organizational efficiency and faster decision-making in identifying next key AI model developments
• Internet giants: Best positioned to capture AI infrastructure/cloud opportunity given strong underlying operating cash flows from core businesses

However, mega-caps will likely need separate standalone incentive schemes for AI chip/model teams to incentivize and retain top AI talent versus independent AI native players.

The Pricing Power Question

Contrary to competitive concerns, pricing power for Chinese AI models has actually increased in 2026. Key evidence:

• Knowledge Atlas/Zhipu: 100% pricing increase year-to-date as of March 2026
• MiniMax: Implemented KV Cache price increases before DeepSeek V4 launch
• DeepSeek V4: Price discounts described as temporary (through end-May 2026)

This is attributed to strong demand and tight computing capacity globally. With rising token prices driven by computing supply-demand tightness, GPU/memory cost push, and improving Chinese AI model performances, coding and multi-modal scenarios show the highest room for pricing gap narrowing versus US state-of-the-art models.

Key Debate #3: Is Token Growth Sustainable?

The Explosive Growth

Daily token consumption on third-party API platforms peaked in late March 2026, with subsequent weeks averaging 79% of that peak. Additional metrics include:

• ByteDance's Doubao model: Over 120 trillion tokens daily (doubled in three months)
• Enterprise clients: 140 companies with daily token usage exceeding 1 trillion (up from 100 at end of 2025)
• Overall China growth: 140 trillion average daily tokens in March vs. 100 trillion at end of 2025

China's enterprise token consumption increased 263% from first half to second half of 2025.

The Cash Flow Question

For Chinese hyperscalers (2026 estimates):

• Aggregate capex: ~$70+ billion
• Capex represents ~60% of operating cash flow

For US hyperscalers (for comparison):

• Aggregate capex: $700+ billion
• Capex represents ~90% of operating cash flow

The outlook suggests potential for further capex increases over second half 2026-2028, driven by sustained token demand.

Sustainability Factors

Multiple drivers support continued token growth:

• AI agents taking over 24/7 tasks
• Enterprises rewarding workers who embrace AI/utilize tokens
• Further enterprise adoption of agent platforms (Claw, Hermes, Co-worker)
• Multi-modal applications expanding token consumption
• Consumer AI assistants adding new demand sources

Cloud revenue acceleration context: US hyperscalers reported significant acceleration in Q1 2026—Google Cloud at 63% year-over-year growth, Microsoft Azure at 40%, and AWS at 29%.

Similar trends are expected for Chinese hyperscalers, with Alibaba Cloud growth estimated at +40% for the March quarter (up from +36% in December quarter).

What Changes Ahead

Token pricing likely has upside potential driven by:

• Improving model performance justifying higher rates
• Cloud providers passing through infrastructure cost increases
• Enterprise willingness-to-pay increases with demonstrated ROI
• Potential shift toward outcome-based pricing (per task vs. per token)

Alibaba has publicly commented that the to-B co-worker/agent market potential "will be larger than all other industries combined (given the US$50tn global white-collar job market)" and estimates its cloud + AI revenues will grow at above 40% CAGR over the next five years.

Key Debate #4: How Will the Pivot to Domestic Chips Impact the Industry?

The Constraint and Response

US restrictions on advanced chips have accelerated China's pivot toward domestic alternatives:

Primary domestic options:

• Huawei Ascend 910C and 950 series: Scaling production from second half 2026
• Internet giants' self-developed chips: Including Alibaba's T-Head
• Optimized architectures: Requiring less memory and compute

Near-Term Challenges

Key obstacles include:

• Near-term supply bottlenecks for domestic chips
• Memory cost push driving up capital expenditures
• Hardware performance gaps requiring offsetting through highly optimized, compute-efficient architectures

Memory cost pressures are expected to drive up near-term capital expenditures for China hyperscalers, following similar trends referenced by US hyperscalers.

Model Adaptation Progress

Chinese models are accelerating Day-0 adaptation to domestic inference chips, with multiple leading models already demonstrating compatibility with Huawei's Ascend chips and other domestic alternatives.

Long-Term Trajectory

The pivot to domestic chips will accelerate over 2026-28, though near-term supply bottlenecks persist. To offset hardware performance gaps, Chinese AI model training is expected to increasingly rely on highly optimized, compute-efficient architectures.

The exceptionally tight computing capacity in China is driving a unique focus on efficiency and optimized architectures—forced innovation that may ultimately create competitive advantages in efficiency.

Key Debate #5: OS-Level vs. In-App AI Agents—What's at Stake?

The Battleground

The consumer AI assistant landscape shows significant momentum:

Doubao (ByteDance):

• ~150 million daily active users (March 2026)
• 63% of overall time spent share among AI apps
• +78% month-over-month growth in March 2026

Other players:

• Alibaba's Qwen app: Growing users from transaction capabilities
• Various AI-native applications: Total engagement up 36% month-over-month

A fundamental question is emerging: Will AI agents operate primarily at the operating system level or within individual applications?

OS-level agents (e.g., Doubao Phone Assistant):

• Control primary user interface
• Can operate across all apps
• Potential to disrupt traditional app engagement
• Raise data privacy concerns

In-app agents (e.g., upcoming WeChat AI agent):

• Integrated with existing ecosystems
• Maintain platform control and data access
• Leverage existing user relationships
• Defend established moats

Why This Matters

OS-level Agentic AI represents a profound paradigm shift that could impact traditional apps by taking over the primary traffic entry point medium-term.

Key risks identified:

• If OS-level agents become the default user interface, standalone apps risk being reduced to backend utility providers, losing valuable user engagement and data
• Potential disruption to China's "super apps" (like Tencent's WeChat) which operate as walled gardens
• Questions about blocking OS agents citing data privacy and security concerns

The Likely Outcome

A fierce strategic battle over interoperability, data permissions, and ecosystem control is expected.

Mega-caps with deep integration across payments, logistics, and social graphs (like Tencent and Alibaba) will likely double down on their own in-app agentic commerce capabilities to defend their moats, while hardware manufacturers and independent AI players will likely push for broader OS-level agentic opportunities.

Recent developments supporting this view:

• April 24, 2026: Doubao embedded "Doubao Help You Choose" feature into navigation bar, formally entering eCommerce
• Upcoming WeChat AI agent launch expected to counter OS-level threats

What Companies and Investors Should Watch

Key Indicators of Market Evolution

1. Model performance in high-value coding scenarios (justifies premium pricing)
   • MiniMax's upcoming M3 model with trillion+ parameters
   • Knowledge Atlas/Zhipu's GLM series leadership
2. Multi-modal capability expansion (video generation quality, audio-visual sync)
   • MiniMax's Hailuo 3 launch expected May 2026
   • ByteDance and Alibaba multi-modal developments
3. Task completion rates vs. benchmark scores (real-world utility)
   • Agent adoption on platforms like OpenClaw, Hermes
4. Enterprise adoption metrics (beyond free trial periods)
   • ARR (Annualized Recurring Revenue) trends
   • Knowledge Atlas/Zhipu disclosed $250M ARR by end-March 2026
   • MiniMax February average ARR of $150M
5. Token pricing trends (indicator of supply-demand balance)
   • Recent 100% price increases at Zhipu
   • Post-promotional pricing sustainability
6. Domestic chip availability (production ramps in second half 2026)
   • Huawei Ascend 910C and 950 series scaling
7. OS vs. in-app agent adoption patterns (determines future app landscape)
   • DAU/MAU metrics for competing approaches

Investment Implications

The sector presents:

• High growth potential in a massive addressable market
• Significant uncertainty around competitive dynamics and profitability
• Structural differences from US AI development paths
• Multiple phases of evolution (current phase: proving commercial viability)

Near-term volatility is likely as new models launch frequently, pricing strategies evolve, enterprise adoption patterns become clearer, and regulatory frameworks develop.

What Happens Next?

Short-Term (2026)

• Continued rapid model launches and performance improvements
• MiniMax M3 model launch expected May 2026
• Pricing competition followed by selective premium pricing for high-value use cases
• Expansion of multi-modal capabilities
• Growing enterprise AI agent adoption
• Domestic chip production scaling (Huawei Ascend second half 2026)

Medium-Term (2027-2028)

• Market consolidation around differentiated players
• Shift toward outcome-based pricing models (charge per successful task vs. per token)
• OS-level vs. in-app agent battle intensifying
• Cloud hyperscaler margin improvement from pricing power
• Profitability inflection for leading independent AI companies

Long-Term Questions

• Can Chinese models match US capabilities in the most complex scenarios?
• Will efficiency innovations create sustainable competitive advantages?
• How will geopolitical factors reshape global AI market access?
• What business models prove most durable in the AI layer?

The Bottom Line

China's AI model industry is at a critical juncture. Companies have demonstrated impressive technical capabilities and cost efficiency while facing legitimate questions about sustainable differentiation and profitability. Analysis across five key debates—US-China gap, competitive moats, token growth sustainability, domestic chip transition, and agent architecture—provides a framework for understanding this rapidly evolving sector.

For enterprises globally, the proliferation of capable, cost-effective Chinese AI models represents both opportunity (better economics, diverse options) and complexity (integration challenges, geopolitical considerations, vendor evaluation).

The industry is moving beyond the "prove the technology works" phase into "prove sustainable business models" phase—a transition that will separate long-term winners from transitional leaders.