"Token New Era" -- Ten Questions and Answers about China's AI Industry

"Token New Era" -- Ten Questions and Answers about China's AI Industry

China's foundational AI model industry is entering a critical phase, shifting from "expectation-driven" to "demand-driven." Morgan Stanley systematically answered investors' ten core questions about the industry in a recent research report, suggesting that model quality is now the primary variable determining market structure, and industry differentiation will accelerate.

According to Morgan Stanley's report released on March 27, China's AI market is at a distinct turning point, with rapidly accelerating demand in coding and agent scenarios. Domestic model capabilities have reached or even surpassed the level of top US models from a year ago, and local pricing is more aligned with economic benefits, jointly improving return on implementation.

2026 is the critical year for whether Chinese enterprise AI demand can replicate the US growth curve seen in 2025. Taking Anthropic as reference, its annual recurring revenue (ARR) grew from $1 billion in December 2024 to $19 billion in March 2026, a nearly 19-fold increase in 15 months.

The Chinese market has similar conditions for following such a path, especially in coding. Internet giants like Tencent, Alibaba, and ByteDance have already integrated related tools into existing ecosystems, driving demand from demonstration toward full deployment. The bank maintains an "overweight" rating for Zhipu and MiniMax, with target prices at 800 HKD and 1100 HKD, respectively.

Question 1: Is AI demand experiencing linear growth, or an inflection point explosion?

Demand is inflection point-driven, not linear growth.

As long as model quality is sufficient to unlock real application scenarios, usage will shift from linear growth to "convex curve" explosive growth. The strongest evidence comes from the US market: Anthropic's annual recurring revenue (ARR) soared from $1 billion in December 2024 to $19 billion in March 2026, a nearly 19-fold increase in just 15 months.

China currently has similar conditions for this explosive growth: domestic model capabilities have surpassed those of the leading US models from a year ago, and local pricing better fits China's labor economics—together significantly improving expected AI implementation returns.

For agents, OpenClaw is a key catalyst, pushing usage scenarios from single-turn interactions to multi-step task execution, greatly increasing the amount of tokens consumed per task. Internet giants like Tencent, Alibaba, ByteDance have integrated OpenClaw tools into their ecosystems, marking a shift from "developer experimentation" to "full ecosystem deployment."

Question 2: Will API pricing rise, fall, or diverge?

Pricing will not move in one direction; differentiation is the main theme.

On one hand, strong models establish pricing power. If a model uniquely unlocks high-value tasks (agent coding, long-horizon workflows, enterprise-grade reliability), customers are willing to pay premiums as returns are quantifiable. On the other hand, as hardware and algorithm efficiency continues improving, unit inference costs will fall, putting pricing pressure on models whose capabilities stagnate.

The final result is a differentiated pricing structure: models maintaining frontier capabilities can enjoy both price and volume growth; models that fail to iterate continually will see prices decline—even as usage grows, margins become uncertain.

Question 3: If pricing is not the main battleground, where is the focus of competition?

The main battleground has shifted from token pricing to model capabilities.

This is a key change compared to last year—2025 saw an all-out price war in China; now, quality is more important than price, especially in the fastest growing agent and coding scenarios.

In multi-step workflows, what customers buy is not "cheap tokens," but "successful task completion." The report provides a clear mathematical example: if single-step success rate rises from 85% to 98%, the final completion rate for a 20-step task jumps from 4% to 67%. Under this logic, the model with the lowest per-token price may have the highest actual comprehensive cost per task.

The report also notes that companies with leading models can easily extend into the low-end market, but companies based solely on low price are unlikely to penetrate the high-end.

Question 4: Why is the foundational large model industry still a "life-or-death" battle?

Small technology gaps, endless iteration cycles, and converging monetization modes make the industry highly brutal.

The gap in capabilities among Chinese large model companies is often smaller than investors imagine, making the market highly unstable. In this industry, "standing still" is not a neutral outcome, but means loss of position—companies must keep investing and iterating to avoid falling behind.

The convergence of business models further intensifies elimination pressure. Revenue growth and profit margin are mainly determined by product strength, switching costs remain low, which means companies losing technological momentum will quickly lose business and financial defenses, eventually reducing the number of truly reliable companies in the industry.

Question 5: What determines profitability?

The core issue is whether gross profit growth can continuously outpace R&D spend growth.

The basic token business economic model is clear: revenue = token usage × price; main cost is inference computation, and main operating expense is training-related R&D. With model and inference chip efficiency improving, leading models' gross margins should rise.

However, operating profit prospects are complex. Anthropic is a cautionary example: even with monthly revenue at $14 billion in February 2026, the company announced a new $30 billion financing round at the same time, emphasizing continued frontier development—high revenue does not mean normalized training intensity.

The baseline scenario is that Zhipu and MiniMax are both expected to become profitable from 2029. The report stresses that more important than the exact year of profitability are two tracking indicators: ongoing usage growth and continuing improvement in unit economics.

Question 6: How should investors track model capability?

Combine token pricing, usage, and third-party evaluations—single indicators are insufficient.

  • Token price: the most important indicator, as it is the company's real-time expression of market positioning. The gap to the best models' price is becoming a proxy for actual competitiveness.
  • Token usage: actual consumption reflects real user and developer choices. Third-party API aggregators like OpenRouter can be referenced, especially focusing on growth in agent task loads, which consume much more tokens per task than simple workflows.
  • Third-party evaluations: Artificial Analysis offers structured assessments, LMArena reflects real users’ blind preference choices; together, they provide a more complete external perspective.

Question 7: As internet giants aggressively enter the B-side, what is the future for independent model companies?

Competitive boundaries are converging, but ultimately still boil down to model capability competition.

Alibaba has made cloud and AI its strategic focus, deeply linking model development with enterprise workflows. Tencent's agent products now cover all scenarios—personal, developer, and enterprise. OpenAI is also shifting its commercialization focus toward enterprise products and coding deployment. Top companies are aligned: AI is evolving from "consumer-end feature" to "a tool for directly generating enterprise revenue."

In this context, independent model companies can no longer rely solely on the "cloud neutrality" label as a moat, and internet giants cannot fully compensate for model capability gaps with ecosystem traffic alone. For enterprise clients deploying AI, the core is still model quality—better coding reasoning and more reliable workflow completion rates.

Question 8: What factors determine a company's survival?

Talent first, compute second, organization third—all three are necessary.

  • Top research talent: This remains a research-driven industry. Technical judgement at the top level is a competitive factor, and whether management can make the right research direction decisions directly determines the company's technological trajectory.
  • Compute and capital: Frontier training is costly, inference profitability depends on infrastructure quality. Weak compute acquisition is a structural disadvantage—not only affecting training efficiency but also lessening the ability to respond to demand at reasonable cost.
  • Organizational execution: In a fast-iterating market, the ability to turn research into product, product into usage, usage into monetization is almost as important as the model itself.

Question 9: If everyone is improving, will models ultimately converge?

Overall capabilities converge, but not fully; the market will not be winner-takes-all.

Different companies have variation in architecture choice, training data, product focus, and technical path, generating ongoing differentiated strengths. The report believes that in a still rapidly expanding market, multiple companies can grow simultaneously, even with some overlapping capabilities—the significance of overall market expansion is much greater than prematurely worrying about commoditization.

In the long run, the more realistic market outcome is not "one winner, everyone else eliminated," but several truly strong companies with their own advantaged fields, competing in a market big enough to support multiple winners. As AI expands from productivity tool to consumer scenarios, differences in personal taste, style, and preference will further reinforce this diversified pattern.

Question 10: How to understand open-source/closed-source, model iteration, and global expansion risk collectively?

Iteration is a must, open-source/closed-source is a strategic choice, global expansion risk centers on compute and compliance.

Model iteration is expected to be at a pace of one flagship model per year (e.g., GLM 4.7 to GLM 5, MiniMax M2 series to M3 series), with minor upgrades driven by reinforcement learning in between. Ceasing iteration means loss of competitive position.

On open-source/closed-source, the report argues that the answer is not binary. Closed-source models have stronger commercial defensibility, lowering risk of disintermediation; open-source helps ecosystem-building, adoption, and accelerates technical feedback. Thus most Chinese model companies are likely to adopt a mixed strategy: close the newest and strongest models, open-source some others.

For global expansion, compute acquisition remains the greatest risk. Both training and inference are highly dependent on high-performance chips, and tightening export controls will slow model progress and harm cost competitiveness. The next risk is data and safety compliance: if model deployment, user service, and data storage can be localized overseas, cross-border data transfer issues are relatively manageable; but local privacy regulations and the recognition of Chinese entity data access rights remain sources of uncertainty.

Risk warning and disclaimerThe market has risks, and investment needs caution. This article does not constitute personal investment advice, nor does it take into account individual users' unique investment objectives, financial situations, or needs. Users should consider whether any opinions, views, or conclusions herein fit their specific circumstances. Invest accordingly, at your own risk.