Z.AI ARR reaches $1.6B as API cloud revenue surges 2,736%

Z.AI's August annualized recurring revenue reached $1.6 billion, already surpassing the $123 million in API revenue the company booked across the entire first half, as management tripled its year-end ARR target to $2.4 billion from a prior $1 billion.

"The shift in revenue structure is essentially a commercial outcome that naturally unfolded after model capability improved," Chairman Liu Debing said in the earnings release. "The core breakthrough of the first half was proving for the first time that models are no longer tools but possess the ability to independently complete a task."

Cloud and API revenue surged 2,736 percent year-over-year to RMB825 million ($122.8 million), accounting for 86.5 percent of total revenue compared with 15.2 percent a year earlier. Total first-half revenue rose 400 percent to RMB954 million ($142 million), while on-premises deployments fell 20.5 percent to RMB128.7 million. Gross margin contracted to 26.4 percent from 50 percent as inference compute costs climbed, though the open platform and API segment's gross margin swung from negative 0.4 percent to positive 24.6 percent.

JPMorgan raised its price target on Z.AI (02513.HK) from HKD1,800 to HKD2,000 with an Overweight rating, lifting FY2026 and FY2027 revenue forecasts by 14 percent and 18 percent. The bank cited the company's rapid model iterations and effective conversion of capability improvements into paid usage across China's large language model sector. Z.AI shares closed up 9.63 percent at HK$1,195 on earnings day before pulling back 5.6 percent as investors digested the scale of the cloud transition.

The company has released four model iterations this year — GLM-5.0 through GLM-5.3 — each focused on converting performance gains into billable tokens. GLM-5.3 Flash, with 320 billion total parameters and 18 billion active, cuts inference costs to one-tenth of its predecessor GLM-5.2. The model was deployed anonymously on OpenCode and OpenRouter under the account "Ox-Alpha," ranking first in OpenRouter call volume on launch day with cumulative traffic exceeding 62 trillion tokens over six days.

Z.AI achieved this scale using 100,000-card-class Chinese-made AI chips, marking the first time a hyperscale inference workload ran entirely on domestic silicon. Per-token inference costs fell 80 percent from the start of the year through proprietary engine optimization, and end-to-end service performance on the same domestic hardware improved threefold. Co-founder Tang Jie said the company's "virtuous cycle where models optimize systems and systems support models has been completed for the first time."

The ARR trajectory puts Z.AI well ahead of rival MiniMax, which disclosed $800 million in ARR in August — half of Z.AI's figure. Platform metrics reinforce the monetization story: paid daily active users on the MaaS platform grew 603 percent from the start of the year, total token call volume rose more than 40-fold, and average API unit pricing increased 101 percent, indicating the company raised prices while scaling volume rather than competing on cost alone.

R&D spending rose 33.6 percent to RMB2.131 billion ($317.1 million) in the first half, while adjusted net loss narrowed 12.1 percent to RMB2.07 billion. The company said gross profit from business segments exceeded administrative and selling expenses for the first time, beginning to partially offset R&D investment.

Liu categorized large model evolution into five stages — Chat, Coding, Agent, Co-work, and Autonomous AI — with Z.AI currently operating between stages two and four. The company is conducting scenario validation in cybersecurity, legal, financial, and education sectors, with cybersecurity advancing fastest. Overseas expansion has shifted from sovereign model localization to API service exports, with partnerships under discussion with overseas cloud service providers.

With management guiding to $2.4 billion in year-end ARR — implying roughly $200 million in monthly revenue by December — the company's ability to sustain the current growth rate will determine whether the stock's post-earnings valuation holds. JPMorgan's HKD2,000 target implies roughly 67 percent upside from the HK$1,195 post-earnings close, and the bank's Overweight rating reflects confidence that China's LLM sector can convert model capability into durable paid usage.

This article is for informational purposes only and does not constitute investment advice.