On the evening of October 18, Runjian Co. disclosed a US$200 million (about RMB 1.338 billion) cross-border AI services contract. Unlike compute exports built on selling servers or building data centers abroad, this contract sells neither hardware nor chips: an overseas client calls large-model inference services via API and pays for the tokens consumed — while all the compute behind it stays in smart computing centers inside China.

A Contract That Reads Like No Traditional Story

The term runs a full year, from October 19, 2026 to October 18, 2027, with usage-based postpaid billing settled monthly. Notably, it carries a minimum-purchase floor: if the client's annual token consumption falls short of US$200 million, the difference must be made up per agreement.

Contract Essentials

Value: US$200M (~RMB 1.338B), roughly 35.6% of Runjian's H1 2026 revenue (RMB 3.756B)

Model: API-based, usage-postpaid, monthly settlement, US$200M annual floor

Infrastructure: compute stays in domestic centers (Wuxiang Cloud Valley), accessed remotely

Seller: Wuxiang Cloud Valley delivers LLM inference services to the overseas client

Why Tokens Can Go Global

Debates about China's AI going global used to center on whether its models could catch up, or whether domestic chips could enter international markets. Now, model API services, inference compute and token-based billing are becoming a new export direction. Chinese AI companies such as DeepSeek, Moonshot, Zhipu and MiniMax have built real traction among overseas developers — where capability meets competitive pricing, applications and enterprise workflows pay.

At least two kinds of players populate this chain: model vendors selling APIs of their own models, and compute service providers orchestrating resources, deployment and delivery. Runjian belongs to the second — even without a frontier model of its own, its computing-center operations carve out a slice of the global market.

The 6.97% Gross-Margin Reality

Contract value is not profit. Runjian's H1 2026 report shows compute-network revenue of about RMB 519 million, up 50.31% — but costs grew faster (+55.08%), leaving a gross margin of just 6.97%, down 2.86 percentage points, weighed down by depreciation of newly commissioned fixed assets. Operating cash flow for the half-year was about -RMB 2.876 billion: heavy upfront investment, recovery later — a classic compute-business profile.

Token volume measures scale, not value. Client retention, per-inference cost, gross margin and cash collection decide whether this business is real.

An Inflection from Selling Hardware to Selling Service

The deeper significance is the template it offers: as rules tighten around AI hardware exports, converting domestic compute into remotely deliverable inference service opens a “service-based export” channel for China's AI industry. Whether it matures comes down to three indicators — overseas renewal and repurchase, improving compute gross margins, and contract revenue converting into cash flow that covers depreciation and operations.

For teams like Jiutian building industry deployments, the signal is just as clear: the token is becoming the base unit of compute commerce — the same unit its Tianxing platform uses to meter industry-model calls. The inference economy's ledger is being rewritten from “buyout” to “pay per token.”

(Compiled from 36Kr / Alpha Factory Institute, Runjian Co. announcements and other public reports)