Goldman Sachs: Commercialization of Chinese Large Language Models Accelerates: Z.AI and MiniMax Revenue Targets, Cost Pressures and Computing Constraints

What has changed in the conditions for Chinese models to catch up?

Experts surveyed by Goldman Sachs believe the performance gap between Chinese and US models has temporarily widened since mid-2026, but real-world feedback from increased use of Chinese models, investment in post-training and recursive self-improvement could help narrow the gap by mid-2027.
The experts believe anti-distillation measures primarily raise the cost of acquiring data rather than preventing models from scaling: Chinese labs are obtaining high-quality data through model application frameworks and synthetic data pipelines. Z.AI believes rising funding, data and engineering barriers will concentrate resources further among leading labs.

The experts observed that large state-owned enterprises and listed companies tend to deploy models locally and have an incentive to train their own models to protect industry knowledge. Excluding multimodal and short-video generation, coding accounts for 60%–70% of text API demand, while office tasks account for 20%–30%.
Enterprises use more than one procurement model. Z.AI says rapid iteration makes it easier for customers to use the latest models through APIs, and enterprise customers are increasing API purchases; enterprise coding agents contributed the vast majority of its new annualized recurring revenue.
Cost continues to limit the scope of high-end models: the experts estimate that about 80% of enterprise and individual tasks are simple enough for small models with around ten billion parameters or on-device models, while the remaining approximately 20% require higher-value frontier models.

Z.AI maintained its guidance for year-end 2026 annualized recurring revenue (ARR) of 3 billion dollars, a target management considers conservative; the metric is calculated monthly using a rolling four-week average. Goldman Sachs expects year-end ARR of 3.2 billion dollars.
The company plans to launch a new model with tens of trillions of parameters in the fourth quarter of 2026. As training progresses, the company expects R&D spending in the second half to accelerate significantly from the first half, rising to approximately twice that level; the new model's parameter count will increase severalfold.
On commercialization, paying users and the ZCode coding product are growing, and revenue sharing from cloud providers has begun to contribute revenue. The company says its gross margin improved quarter over quarter and is now roughly sufficient to cover general and administrative expenses and sales and marketing expenses; R&D, however, remains a significant expense affecting near-term profitability.

MiniMax says its ARR had already reached its original year-end 2026 target of 1 billion dollars before the recent launch of M3.1 Flash; Goldman Sachs expects it to reach 1.2 billion dollars by year-end. Management hopes to grow ARR severalfold in 2027 and plans to launch M3 Pro, with approximately 3 trillion parameters, in early to mid-November.
The company says M3.1 Flash is priced 50% below M3, but model architecture and infrastructure optimization keep its margin comparable to M3. Management acknowledges gaps in gross and operating margins versus global peers and expects price competition among Flash-tier models may intensify.
MiniMax expects less competitive pressure in the more capable Pro tier; approximately sixty percent of revenue comes from overseas, and international customers also help support margins.

The experts say China's computing utilization has risen to nearly 100% since November 2025, up from 60%–70% in October 2025; rental prices for advanced overseas servers have risen by approximately 2–3 times from a year earlier. However, higher rental prices for Nvidia chips have narrowed their cost-performance advantage over domestic chips; together with increased domestic computing supply, this could place a ceiling on further price increases.
Both companies are increasing their use of domestic chips. Z.AI says domestic chips already handle a substantial share of inference tasks, and infrastructure optimization and higher utilization have helped improve inference margins despite rising memory costs; MiniMax says deliveries of the relevant domestic chips began in the second half of 2026, and it plans to expand their use significantly the following year.
Goldman Sachs rates Z.AI and MiniMax Buy, with respective 12-month price targets of 1,560 Hong Kong dollars and 760 Hong Kong dollars.