BofA: China AI Model Upgrade Wave Accelerates as Cloud Giants Ramp Up Spending
A sweeping upgrade cycle across China's artificial intelligence model landscape is accelerating alongside a dramatic surge in cloud infrastructure spending by the country's leading internet platforms, with independent data pointing to Chinese models dominating global developer usage metrics, according to a newly published Wall Street research report.
According to BofA Global Research's "China AI Model Monitor" report published on August 31, 2026, authored by analysts Alex Liu, Miranda Zhuang, Joyce Ju, and Joanna Du at Merrill Lynch (Hong Kong), the month of August was defined by three converging forces: rapid model upgrades from both cloud incumbents and independent AI labs, selective token price increases driven by compute scarcity, and a sharp acceleration in capital expenditure by China's top cloud platforms.
Model Upgrades Intensify Across the Board
The report documents a broad-based wave of model releases throughout August. Tencent launched its Hunyuan 4 preview on August 28, while Alibaba released Qwen 3.8 Max and Qwen 3.8 Flash on August 3. Independent AI lab Zhipu AI introduced GLM-5.3 Flash on August 19. These releases reflect an ongoing competitive dynamic in which both established internet platforms and startup AI laboratories are racing to push model capabilities forward on a near-monthly cadence.
In terms of raw intelligence benchmarks, Anthropic's Claude Opus 5 continues to lead the Artificial Analysis Intelligence Index with a score of 63, followed by Claude Fable 5 at 62 and OpenAI's GPT-5.6 Sol at 61. Among Chinese models, Moonshot AI's Kimi K3 ranks fifth globally with a score of 60 — representing approximately 95% of Claude Opus 5's score — followed by Zhipu's GLM-5.3 and Alibaba's Qwen 3.8 Max, both at 60. This places Chinese open-source models firmly within striking distance of the leading proprietary Western systems on standardized capability assessments.
Chinese Models Dominate Developer Usage Platforms
Usage data from third-party platforms paints an even more striking picture of Chinese model adoption. On OpenRouter, an AI model aggregation platform widely tracked as a proxy for developer sentiment, DeepSeek V4 Flash led all models with 31.6 trillion tokens consumed month-to-date as of August 17, followed by Tencent's Hy3 at 26.2 trillion tokens and Xiaomi's MiMo-V2.5 at 19.1 trillion tokens. By provider share, DeepSeek accounted for 22% of weekly token volume among the top 50 models on OpenRouter in the week of August 17, gaining 4.7 percentage points compared to the same week in July.
On the Vercel AI platform, DeepSeek led average month-to-date token usage share at approximately 30%, up from 26% in July, while Anthropic's share declined to roughly 25% from 30.1%. However, the spend share picture diverges sharply: Anthropic retained dominant monetization leverage, capturing approximately 65% of total spend on the Vercel platform, underscoring the gap between volume-driven adoption of low-cost Chinese open models and the premium pricing commanded by leading Western proprietary systems.
On the coding-focused OpenCode platform, DeepSeek continued to dominate token consumption, while Zhipu's GLM-5.3-Flash — reportedly operating under the alias "Ox-alpha" — recorded a sharp surge in usage following its late-August launch.
Token Pricing: Costs Fall but DeepSeek Bucks the Trend
The BofA report notes that the Silicon Data LLM Token Expenditure Index declined 26% month-on-month in August to $1.07, down from $1.45 in July, continuing a broader deflationary trend in AI inference costs. However, DeepSeek raised its token prices during the month — a move the analysts attribute not to a strategic pivot away from its high-performance, low-cost positioning, but rather to near-term compute tightness. US frontier models continue to command a substantial pricing premium: Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens, compared to DeepSeek V4.0 at $0.70 input and $2.00 output.
Cloud Capex Surges as AI Infrastructure Buildout Accelerates
Perhaps the most consequential data point in the report concerns capital expenditure. In the second quarter of 2026, Tencent's capex reached RMB53 billion (approximately US$7.3 billion) and Alibaba's reached RMB68 billion (approximately US$9.4 billion). BofA now forecasts combined capex for China's top four cloud platforms to rise 98% year-on-year in full-year 2026, followed by a further 44% increase in 2027. Forward 12-month consensus capex estimates for Tencent and Alibaba were revised upward by 25% and 29%, respectively, over the 30 days prior to the report's publication.
Cloud revenue is accelerating in tandem: Tencent's cloud revenue grew in the low-20% range year-on-year in the second quarter, while Alibaba's cloud revenue expanded 45% year-on-year in the same period.
BofA projects the total China cloud market will grow from RMB389 billion in 2025 to RMB1,729 billion by 2030, driven predominantly by AI-related workloads. Within that, the AI Model-as-a-Service segment is forecast to be the fastest-growing component, expanding from RMB21 billion in 2025 to RMB676 billion in 2030, a compound annual growth rate of 81%.
Chatbot Engagement: ByteDance Widens Lead
In China's consumer AI chatbot market, ByteDance's Doubao maintained a commanding lead with weekly daily active users reaching 171.3 million and total weekly time spent of 11.5 billion minutes in the week of August 3. DeepSeek ranked second with 30.7 million weekly DAUs, followed by Alibaba's Qwen at 28 million.
Outlook: Key Catalysts on the Horizon
The report identifies a series of near-term events that could materially impact the competitive landscape. Alibaba is expected to unveil a potential model upgrade at its Cloud Summit scheduled for September 22–24. MiniMax may launch its M3.1 and M3 Pro models in the September–October timeframe. DeepSeek's next-generation V5 model and Tencent's Hy4 official release or Hy5 preview are both anticipated in the fourth quarter of 2026. Additionally, lock-up expiry events for Zhipu AI and MiniMax in early January 2027 — unlocking 40% and 90% of shares, respectively — are flagged as potential market-moving catalysts for investors tracking the independent AI lab segment.
The convergence of accelerating model capabilities, surging developer adoption of cost-efficient Chinese open models, and an unprecedented infrastructure investment cycle suggests that the competitive dynamics of the global AI industry are entering a new phase — one in which Chinese platforms are playing an increasingly central role both in capability benchmarks and real-world usage.
Related Coverage:
Moonshot AI’s Kimi K3: How China’s AI Startups Can Scale Globally Without Building Alone
Alibaba’s Qwen3.8-Max Challenge: How China’s AI Stack Is Closing the Gap With Silicon Valley