Initiating coverage of Chinese AI labs: constructive on Z.ai, cautious on MiniMax

Overview

The report initiates coverage of Z.ai and MiniMax and compares the two companies across model capability, accumulated R&D, commercialization path, inference economics, and valuation. Bernstein is overall constructive on leading Chinese AI labs continuing to iterate near the global technology frontier, believing that domestic policy support, domestic chip capacity expansion, and overseas adoption of cost-effective open-source models will drive long-term growth, but sees clear differences in investment certainty between the two companies.

Core views

The commercial success of AI labs first depends on whether they can establish differentiated capability peaks and occupy a position on the Pareto frontier of reasoning capability and price; annual recurring revenue and inference margins are downstream results of model competitiveness. Z.ai has a team of basic scientists, Tsinghua University connections, broad internet platform customers, and stronger coding model positioning, while GLM-5.2 demonstrates good parameter efficiency, so it is viewed as a purer exposure to frontier R&D capability. MiniMax’s M3 lags the frontier, making M3 Pro the key proof point; its multimodal and video generation businesses face relatively weaker competition, but their commercialization ceiling may be lower than that of agentic coding and productivity scenarios.

Analysis framework

The report uses a combination of top-down and bottom-up approaches: top-down to assess China’s AI capital expenditure, revenue TAM, policy, and compute supply; bottom-up to compare model parameter scale, benchmark performance, accumulated R&D team capability, developer adoption, distribution partners, annual recurring revenue, inference margins, and financial forecasts, and to determine target prices using relative valuation and discounted cash flow methodology.

Methodology notes

• Competitiveness analysisReasoning capability-price Pareto frontier

  Models must provide a hard-to-substitute combination of capability and cost and form differentiated capability peaks.

  The report believes that entering the Pareto frontier is only a basic threshold; only the ability to solve difficult tasks, create higher willingness to pay, and scale services at lower cost can support the next round of training investment.

• Technical capability assessmentAccumulated R&D and proprietary model benchmarking

  Measure continuous iteration capability based on the team’s scientific research background, model architecture innovation, parameter scale, task completion capability, and developer adoption.

  Model iteration is viewed as a capability-driven process, and short-term market sentiment cannot replace analysis of the R&D team, model progress, and actual developer usage.

• Market size estimationCapex-to-revenue conversion and knowledge worker spending method

  Estimate China’s AI revenue TAM respectively from AI capex conversion and per-capita spending by knowledge workers.

  The report estimates China’s 2030 AI revenue TAM using an approximately 3:1 relationship between capex and incremental revenue, and cross-checks it using the number of knowledge workers and wage differences.

• Valuation methodsRelative valuation and discounted cash flow (DCF)

  Determine target prices by combining forward P/E or P/S multiples with discounted cash flow results.

  Z.ai uses 25x expected 2030 P/E discounted at 14% per year and combines it with DCF valuation; MiniMax uses 4.5x expected 2029 P/S discounted at 14% per year and combines it with DCF valuation.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Z.ai (2513.HK)

  Core bullish exposure

  Strengths

  The basic scientist team has strong accumulated R&D capability; GLM-5.2 demonstrates good intelligence and parameter efficiency; the coding direction aligns with enterprise AI adoption; it has internet platform customers, talent and data channels related to Tsinghua University, and relatively broad overseas distribution support.

  Weaknesses

  Kimi K3 has raised the bar for frontier models, and the company’s next-generation pre-training model must prove it remains at China’s frontier; R&D spending is high, and the company is still loss-making in the near term.

  Comparison

  Compared with MiniMax, Z.ai is viewed as a more repeatable and purer exposure to frontier reasoning R&D capability, with a higher commercialization ceiling in coding and agent markets.

  Risks

  Next-generation GLM model performance below expectations, rapid iteration by competitors, delays in compute expansion, price competition pressuring inference margins, and high R&D investment weighing on profitability.

• MiniMax (100.HK)

  Neutral allocation exposure

  Strengths

  The share price has fallen significantly from its high; competition in the multimodal market is relatively weaker; H3 is expected to improve positioning and support annual recurring revenue growth; video generation aligns with its accumulated expertise in computer vision.

  Weaknesses

  M3 lags the frontier, and M3 Pro needs to complete an unusually significant parameter-scale leap; the R&D roadmap has gone through multiple strategic adjustments; global distribution partners have decreased; the revenue TAM and commercialization ceiling of multimodal and video generation may be lower than those of agentic coding.

  Comparison

  Compared with Z.ai, MiniMax has lower visibility in its model pipeline and frontier R&D capability, so a lower forward P/S valuation is applied.

  Risks

  M3 Pro fails to restore frontier credibility, price competition intensifies in the low-complexity agent market, commercialization potential for video generation is limited, annual recurring revenue growth is weaker than expected, and valuation multiples are downgraded.

Key data

• Z.ai rating and target priceOutperform; HK$1,350Current price HK$1,014.00 as of 2026-08-03, implying approximately 33.1% upside.
• MiniMax rating and target priceMarket Perform; HK$275Current price HK$230.20 as of 2026-08-03, implying approximately 19.5% upside.
• 2030 revenue TAM for leading Chinese AI labsUS$100 billion to US$200 billionExcludes a large number of consumer use cases; the report believes leading labs may capture a disproportionate share of value.
• China’s total 2030 AI revenue TAMapproximately US$200 billion to US$300 billionEstimated based on cumulative AI capex of over US$1 trillion and the conversion relationship from capex to revenue.
• Capability gap between Chinese and U.S. frontier modelsapproximately 3 to 4 monthsThe report believes GLM-5.2 and Kimi K3 may have further narrowed the previous gap of about 6 to 9 months.
• Z.ai compute and annual recurring revenueapproximately 1 gigawatt of available compute by year-end; annual recurring revenue is expected to exceed US$2 billion around year-endCompute expansion is viewed as an important support for revenue outperformance.
• Z.ai profitability inflection pointaround 2028The report expects the company to reach non-GAAP operating profit breakeven under the assumption of rapid growth in R&D investment.
• Z.ai model scaleGLM-5.2 pre-training total parameters of approximately 744 billionThe report believes its intelligence level and cost efficiency are outstanding.
• MiniMax model leap requirementincrease from 428 billion to 2.7 trillion total parametersThe required leap for M3 Pro is more than twice the level commonly seen among Chinese peers in recent years.

Impact & implications

If leading Chinese labs continue to narrow the gap with U.S. frontier models and reduce inference costs after domestic compute expansion, their revenue growth could be significantly higher than current market expectations. From an investment perspective, investors should prioritize repeatable model R&D capability, coding and agent scenarios, developer adoption, and inference unit economics, rather than trading solely based on the most recent model release. The report therefore prefers Z.ai; while MiniMax may benefit from H3 and multimodal demand, its valuation and market confidence still depend heavily on whether M3 Pro can restore its frontier position.

Risks

• Frontier models iterate extremely quickly, and a single model release may cause market expectations and share prices to reset sharply.
• U.S. labs have significantly stronger access to compute, and the gap between Chinese models and the global frontier may widen again.
• If domestic chip and compute supply expansion falls short of expectations, it will constrain training scale, inference supply, and revenue growth.
• Convergence in model capabilities and price competition in low-complexity tasks may compress inference margins.
• Sustained high R&D and training investment may delay profitability and the free cash flow inflection point.
• Overseas market access, geopolitics, and technology restrictions may constrain the international commercialization of Chinese models.
• The commoditization potential and revenue ceiling of multimodal and video generation markets may be lower than those of the agentic coding market.
• Valuation depends on forward revenue, margin, and discount-rate assumptions, and forecast deviations may significantly affect target prices.

What to watch

• Z.ai’s GLM-5.3 update planned for August 2026 and the next-generation pre-training model expected to launch in the fourth quarter of 2026.
• Whether Z.ai’s approximately 1 gigawatt of available compute by year-end can be delivered on schedule, and whether annual recurring revenue can exceed US$2 billion around year-end.
• Whether GLM models’ coding, long-horizon task, and agent capabilities remain on the capability-price Pareto frontier.
• Whether MiniMax M3 Pro can achieve a leap from 428 billion to 2.7 trillion total parameters and restore frontier credibility.
• The actual boost from MiniMax H3 to annual recurring revenue, user adoption, and market positioning.
• High-frequency indicators such as web traffic, API revenue, developer usage, and the number of global model service platform partners.
• Capacity expansion progress for China’s domestic chips, inference infrastructure, and local compute factories.
• Inference pricing, gross margins, and the intensity of price competition in low-complexity agent tasks.
• Adoption and commercial licensing models of Chinese open-source models in non-U.S. overseas markets.