OpenRouter just closed another funding round at a valuation approaching $1.3 billion. Its model is straightforward: one API entry point, 300+ models behind it, a small markup on token usage, and annual revenue growing fivefold.
Three days later, China Mobile unveiled MoMA at its Mobile Cloud Conference in Suzhou.
Also 300+ models.
Not a coincidence. Two very different answers to the same question.
300 models, two completely different business math
OpenRouter's math is clear: developers pay to call models, the platform takes a cut, and the more people use it, the more it makes. Its moat is embedded in developer workflows — switch models with a single parameter, near-zero migration cost.
MoMA's math lives somewhere else entirely.
China Mobile isn't trying to make money off developers' token usage. It did three things OpenRouter never bothered with:
First, centralized token lifecycle management. Streaming real-time billing with end-to-end latency under one minute, pay-as-you-go. OpenRouter has similar mechanisms, but MoMA added full-chain risk control — every token consumption is traceable, auditable, and closed-loop. What does that mean? Government and financial clients don't fear high costs — they fear not knowing where the money went. MoMA solves that.
Second, three-tier smart routing. "Cost-first," "quality-first," "balanced" strategies, auto-switching. Second-level failover when models time out or get throttled. OpenRouter has fallback too, but MoMA frames it as a "neutral routing principle" — no vendor lock-in, balancing business outcomes against compute costs.
Third, confidential model service. Models deployed in confidential containers with hardware isolation — "usable but invisible." This is a capability OpenRouter doesn't have at all, and it's China Mobile's killer pitch for government and financial compliance.
See the pattern? MoMA isn't China's OpenRouter. It's AI infrastructure built for enterprise compliance scenarios.
Why China Mobile, not Unicom or Telecom?
China Mobile has structural advantages the other two telcos and even internet cloud companies can't match:
Compute network. 43 EFLOPS of intelligent compute centers, nationwide compute调度 capability. MoMA runs on domestic compute with a self-developed inference engine, cutting per-token cost by ~30%. This isn't subsidized pricing — it's scale economics.
Enterprise relationships. How many ministries, provinces, and municipalities does China Mobile already serve? That client list is the moat. Enterprise clients don't want the "best model" — they want the most compliant, most controllable platform where someone answers the phone when things break.
Terminal reach. The conference also announced AI-eSIM, embedding LLM capabilities directly into eSIM chips with real-time cloud model scheduling. AI toys, smart wearables — these terminals live natively on China Mobile's network.
But the gaps are real
Pretending China Mobile has no weaknesses is dishonest.
Developer experience is a genuine problem. OpenRouter's rise came from developer口碑 — clean docs, friendly SDKs, active community. In everything disclosed about MoMA so far, there's no evidence of deep developer workflow optimization. Enterprise deals don't close on GitHub stars, but you can't build an ecosystem without developers.
Innovation velocity. The decision-making chain at a state-owned enterprise and the iteration speed of an internet company operate on different timescales. Smart routing and token metering — OpenRouter and SiliconFlow already solved these. China Mobile is catching up, not pioneering.
Model ecosystem openness. MoMA接入 DeepSeek, Qwen, Doubao, Kimi, GLM — but these are "partnership" relationships. If any model vendor decides to go direct and bypass aggregation platforms, will MoMA's 300+ shrink? The risk is there.
The real bet
Back to the opening question: why is China Mobile building an "OpenRouter"?
The answer isn't in developers. It's in the "last mile" of enterprise AI adoption.
In 2026 China, models aren't the bottleneck — knowing how to use them is. Government, finance, healthcare, industrial sectors all have data, scenarios, and budgets. But they don't know how to pick models, how to schedule compute, and they're terrified of handing sensitive data to third parties.
What MoMA is trying to do is package "pick a model, allocate compute, control cost, ensure security" into a standardized service. Enterprise clients don't need to understand AI — they just call one API.
That's orders of magnitude heavier than OpenRouter's "give developers more choice." And ten times harder to pull off.
What to watch next
MoMA just launched. Three metrics actually matter:
- Enterprise client signings within six months — not "strategic partnerships" on a slide, but actual clients generating token volume
- The ratio of Jiutian (self-developed) model calls vs. external models — if 90% of platform traffic is Jiutian, MoMA is just a shell. If external models get real share, the smart routing is actually working
- AI-eSIM adoption velocity — novel concept, but it lives or dies on how fast terminal ecosystems adopt
China Mobile has compute, clients, and terminal reach. What it lacks is developer ecosystem and technical iteration speed. Whether MoMA can match OpenRouter's growth curve depends not on model count, but on whether it can find a balance between enterprise compliance and developer openness.
Worth watching.