润建股份签下2亿美元海外大模型推理服务合同

（本文作者为 Chelsea_Sun，钛媒体经授权发布）

NextFin News — Overseas developers do not need a rack in Shenzhen to burn Chinese compute. They need an API key, a network path and a price per token. That is the business Runjian Shares put on the calendar in mid-October: a one-year, $200 million commitment from a foreign customer to buy large-model inference produced in domestic intelligent computing centers and delivered over the wire.

The deal is structured as post-paid, metered service with a twist. If the client’s real consumption falls short of the annual minimum, the shortfall is still owed under the contract’s terms. Runjian is not shipping servers or chips. It is renting thought-cycles—input tokens in, output tokens out—while the hardware never leaves the country.

For China’s AI stack, that is a different export story from the ones that usually dominate. Debate has long fixed on whether local models can match Western peers and whether domestic accelerators can clear foreign procurement. Token billing opens a third lane. Keep the capital stock at home. Sell the output as a utility. Firms that train models can meter their own weights. Firms that only operate halls of GPUs can still participate if they can schedule jobs, keep uptime high and invoice cleanly. Runjian’s background is telecom operations and, more recently, computing networks. This contract fits the operator mold more than the foundation-model mold.

Headline value is easy to misread. Two hundred million dollars is about a third of the company’s first-half 2026 sales. Accounting will recognize revenue as service is performed over a year that runs into October 2027, not when the press release hits. The floor protects volume; it does not guarantee margin or prompt payment.

Existing numbers already hint at the squeeze. Runjian’s computing-network line grew quickly in the first half of 2026, yet costs grew faster. Gross margin sat in the high single digits and slipped from a year earlier. Management blamed heavy depreciation while new assets ramped—the classic lag between energizing a hall and filling it with paid work. Apply a similar margin to the full notional of the new contract and the implied gross profit is modest relative to the splash. Power, maintenance and financing sit underneath that.

Competition will press the same numbers. Once many parties can expose interchangeable inference endpoints, buyers shop latency, reliability, support and price. Model companies fight on quality and efficiency per call. Capacity companies fight on utilization. Either way, tokens processed is a vanity metric until cost per token and cash collection are known.

Cross-border delivery adds friction that a domestic SaaS contract can ignore: knowing the customer, constraining data and content, and living with rules in the buyer’s jurisdiction even when the GPUs stay onshore. Capital intensity shows up in the cash-flow statement as well. Runjian’s operating cash was deeply negative in the first half of 2026 as computing and energy investments ran ahead of collections weighted toward later months. That pattern is normal for infrastructure businesses. It is also unforgiving if racks sit half idle.

What will decide whether “tokens abroad” is a real export category is boring and decisive. Does the customer renew. Does segment margin recover as utilization rises. Do invoices turn into cash fast enough to fund the next tranche of hardware. Until those answers arrive, the contract is proof of demand at a point in time—not yet proof of a durable business.

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App