← Back to stories A scientist conducting an experiment in a lab, focusing on precision and safety.
Photo by Mikhail Nilov on Pexels
虎嗅 2026-04-03

Zhipu AI (智谱AI) Posts First Post‑IPO Annual Report: Rapid Revenue, Massive R&D Losses, Pivot to Token‑based MaaS

Results in brief

Zhipu AI (智谱AI) released its first annual report since listing, showing sharp top‑line growth alongside heavy investment-driven losses. The company reported 2025 revenue of RMB 724 million, up 131.9% year‑on‑year, with cost of sales rising 213.3% to RMB 427 million. Gross profit reached RMB 297 million (up 68.7%) and gross margin improved to 41% — well above MiniMax’s reported 25.4% margin. But adjusted net loss widened to RMB 3.182 billion, driven primarily by RMB 3.184 billion in R&D spend; that loss is roughly 4.4 times full‑year revenue and more than ten times gross profit.

The filing clarifies an important accounting treatment: short‑term GPU usage for model iteration is booked into R&D, while long‑term reserved GPU contracts are treated as capital expenditure (2025 capex was RMB 74.7 million). Analysts note that export controls on cutting‑edge accelerators have made long‑term compute commitments strategically and financially material for Chinese AI firms.

Growth engine: AutoClaw, tokens and MaaS

Zhipu says its growth is shifting from projectized local deployments to a cloud‑centric MaaS (Model‑as‑a‑Service) strategy that sells tokenized model calls. The company attributes a large part of its Q1 momentum to AutoClaw — a one‑click deployment of its “Longxia” agent — and a reported 83% API price increase in March that coincided with a demand surge. Zhipu reported a 400% jump in GLM model call volume; subscription users reportedly topped 100,000 within two days and 400,000 within 20 days after launch. The MaaS API platform reached ARR of RMB 1.7 billion (about $250m), a roughly 60‑fold year‑on‑year reported increase.

Zhipu has formalized a new internal KPI, TAC (Token Architecture Capability), combining token volume, effective quality of calls and economic conversion efficiency — reflecting an industry shift from raw parameter counts and price wars to agent‑driven, multi‑call token economics. The company also says nine of ten major Chinese internet firms have integrated Zhipu models in some scenarios, and it is expanding token sales to partners in the Middle East and Southeast Asia.

Risks and the road ahead

The report signals a clear strategic pivot, but risks are concentrated. Cloud/MaaS now accounts for 26.3% of revenue (up from 15.5% in 2024), yet Zhipu’s MааS income is heavily dependent on a small number of large customers whose call volumes represent a significant share of ARR — losing one could materially damage revenue. And although the margin profile for cloud deployments improved (from 3.3% to 18.9%), local deployments still carry higher margins (48.8% in 2025).

Can selling tokens and APIs scale to cover sustained R&D investment? Zhipu is betting yes, and the numbers show both promise and peril: stronger unit economics and explosive API demand on one hand; multi‑billion RMB annual cash burn on the other. Investors and rivals alike will watch whether the TAC framework and a cluster of large customers can turn rapid usage growth into durable, profitable platform revenue.

AI
View original source →