Morgan Stanley Private Meeting: MiniMax and Zhipu ARR Growing Rapidly, Model Competition Enters Tiered Elimination Phase
nashnova research
Minutes from a Morgan Stanley closed-door session show MiniMax and Zhipu both beating ARR expectations — MiniMax's year-end forecast raised to $1.3 billion, Zhipu's August ARR already past $2 billion — as analysts call the start of a tiered shakeout in China's foundation-model race.
Where is MiniMax's revenue coming from, and how fast is it growing?
As of late August, MiniMax's ARR — annualized recurring revenue, the current run-rate projected over a full year — reached $800 million, ahead of market expectations.
Official guidance targets $1 billion-plus by year-end, but analysts have raised their estimate to $1.3 billion. This means → the Street thinks management is sand-bagging; the real upside depends on new model launches.
Three new models — M3.1, H3.1, and M3 Pro — are planned for the second half. Compute reserves are sufficient for all three and already being stockpiled for a 10-trillion-parameter model next year.
Why is MiniMax's gross margin slipping?
Three factors drove H1 margin compression: deliberate price cuts and user subsidies at M3's launch, an inference-efficiency ramp still underway, and a rising share of text-model revenue pulling down the blended rate.
In plain terms = multimodal models — those handling images and video — carry margins above 50%, but pure-text models are thinner. Selling more text actually dragged the average down.
Management calls these mostly one-off H1 effects and expects margins to improve sequentially in H2.
How steep is Zhipu's growth curve?
Zhipu's ARR trajectory reads like a monthly step-function: March $250 million → June $530 million → July $1 billion → August $1.6 billion. On a weekly basis, August already exceeded $2 billion.
Year-end guidance was raised to $2.4 billion, though the company itself calls that conservative — the actual figure hinges on domestic-chip delivery schedules.
This reflects a growth bottleneck that has shifted from demand to compute supply: the faster chips arrive, the higher the revenue ceiling rises.
What is supporting Zhipu at this scale?
Zhipu operates a 100,000-card domestic-chip cluster with enough advanced compute reserved to train trillion-parameter models.
Its top 10 clients contribute over 40% of revenue. Nine of China's ten largest internet companies are Zhipu customers, and four of them have designated GLM as their company-wide primary model.
GLM-5.3 Flash, released in August, runs on a new architecture and infers entirely on the domestic-chip cluster. That architecture carries forward into GLM-6, due in October.
Where does Tencent's Hunyuan stand?
Hunyuan 4 Preview shipped in late August, ahead of schedule. Internal blind tests put its performance slightly above GLM-5.3 and Kimi K3.
WorkBody is evolving from an office assistant into an Agent-layer platform. This means → Tencent is betting that even if its model is not the outright leader, the ecosystem platform can monetize independently.
Key milestone ahead: October's Global Digital Ecosystem Conference, where new products and strategic direction will be unveiled.
What does "tiered shakeout" actually mean?
Analysts argue that companies with SOTA capability — state-of-the-art, the current performance frontier — can use distillation (compressing a large model's knowledge into a smaller, cheaper version) to offer high-value-for-money products.
In plain terms = the strongest players can "punch down" — selling both a flagship and a budget line. Companies whose only edge is low price cannot break through the technology barrier, and their margins are being squeezed. A pure price war is not sustainable.
Buy-side investors currently prioritize model capability first, then ARR growth and margins. As long as the market recognizes a model's strength, valuation tolerance stays high. October's wave of new model launches will be the critical test of whether this tiered structure accelerates toward consolidation.
市场有风险，内容仅供研究参考，不构成投资建议。