China Mobile Launches MoMA: 300+ AI Models, 30% Lower Costs

China Mobile, the world's largest mobile carrier by subscribers, has officially launched its MoMA (Mobile Model-as-a-Service) platform, aggregating more than 300 mainstream AI models under a single unified API gateway while slashing per-token costs by approximately 30%. The platform, announced on May 8, represents one of the most ambitious attempts by a telecom giant to position itself as a central hub in the rapidly expanding AI model ecosystem.

Built on domestically produced computing infrastructure and powered by China Mobile's proprietary inference engine, MoMA introduces what the company calls a 'token-intensive operation model' — a first-of-its-kind approach designed to optimize how AI workloads are distributed, cached, and billed across hundreds of competing models.

Key Takeaways at a Glance

• 300+ AI models integrated through a single API gateway, including DeepSeek, Tongyi Qianwen (Alibaba), Doubao (ByteDance), Kimi (Moonshot AI), and GLM (Zhipu AI)
• 30% reduction in per-unit token costs through intelligent routing and resource scheduling
• 50%+ decrease in resource utilization overhead for long-tail model deployments
• 3 routing strategies: cost-priority, performance-priority, and balanced modes
• Sub-second failover when models experience timeouts, throttling, or outages
• Confidential model service using hardware-isolated secure containers for sensitive deployments

A Unified Gateway for China's Fragmented AI Landscape

China's AI model market has exploded over the past 18 months. Dozens of large language models now compete for enterprise adoption, from Alibaba's Qwen series to ByteDance's Doubao and rising stars like DeepSeek and Moonshot AI's Kimi. For enterprise customers, this fragmentation creates a real headache: evaluating, integrating, and managing multiple model APIs requires significant engineering effort.

MoMA addresses this pain point directly. The platform provides a unified API gateway that abstracts away the differences between individual model providers. Developers can access text generation, speech processing, and multimodal understanding capabilities through a single integration point, rather than maintaining separate connections to each provider.

The platform also includes China Mobile's own foundation model, Jiutian ('Nine Heavens'), alongside the third-party offerings. This positions China Mobile not just as an aggregator but as a participant in the model ecosystem — a strategy reminiscent of how cloud providers like AWS and Microsoft Azure offer both their own AI services and third-party model marketplaces.

Smart Routing Engine Cuts Costs by 30%

Perhaps the most technically interesting feature of MoMA is its intelligent routing engine, which dynamically selects the optimal model for each request based on user-defined priorities. The system offers 3 distinct strategies:

• Cost-priority mode: Routes requests to the most affordable model capable of handling the task
• Performance-priority mode: Selects the highest-performing model regardless of cost
• Balanced mode: Optimizes for the best cost-performance ratio

This approach mirrors a trend gaining traction in the Western AI ecosystem as well. Startups like Martian and Not Diamond in the U.S. have built similar model-routing solutions, and OpenAI itself has explored routing between different model tiers. However, MoMA operates at a significantly larger scale, with 300+ models in its catalog compared to the handful typically available through Western routing services.

The cost savings come from multiple technical optimizations working in concert. China Mobile's proprietary inference engine, deployed on domestic GPU and accelerator hardware, handles resource scheduling for 'long-tail' models — those that see intermittent rather than constant demand. By intelligently consolidating these workloads, the platform achieves the reported 30% reduction in per-token costs and over 50% improvement in resource utilization.

Advanced Caching and Token Compression Drive Further Savings

Beyond routing optimization, MoMA employs several additional techniques to reduce the total cost of AI inference:

• Intelligent caching: Frequently requested responses are cached and served without invoking the underlying model, reducing both latency and cost
• Context reuse: For multi-turn conversations, the platform reuses previously computed context rather than reprocessing entire conversation histories
• Token compression: Input and output tokens are compressed where possible, reducing the billable volume without sacrificing output quality

These techniques are particularly valuable for enterprise deployments where the same types of queries recur frequently. In sectors like government services, financial services, industrial automation, and healthcare — all explicitly targeted by MoMA — repetitive query patterns are common, making caching and compression especially effective.

The combination of smart routing and these optimization techniques could make MoMA particularly attractive to cost-conscious enterprise buyers who want access to premium models without premium pricing. It effectively creates a 'wholesale' layer for AI inference, leveraging China Mobile's massive infrastructure to negotiate better economics than individual enterprises could achieve on their own.

Confidential Computing Addresses Enterprise Security Concerns

Enterprise adoption of AI models has been hampered globally by data security concerns, particularly in regulated industries. MoMA tackles this challenge head-on with its 'Confidential Model' service, which deploys AI models inside secure, hardware-isolated containers.

This approach uses confidential computing technology — a hardware-based security paradigm that protects data during processing, not just at rest or in transit. By running model inference inside trusted execution environments, MoMA ensures that even the platform operator cannot access the data being processed. This is a critical assurance for government agencies and financial institutions handling sensitive information.

The confidential computing approach also addresses a unique concern in the model-aggregation business model: when an enterprise sends data to a third-party model through a routing platform, there are effectively 2 trust boundaries to cross. MoMA's hardware isolation reduces this risk by ensuring data remains encrypted throughout the entire pipeline, from the chip level upward.

How MoMA Fits Into the Global AI Platform Wars

China Mobile's launch comes at a pivotal moment in the global AI infrastructure landscape. In the West, the model-as-a-service market is dominated by cloud hyperscalers: Amazon Bedrock offers access to models from Anthropic, Meta, and Mistral; Microsoft Azure AI provides OpenAI models alongside open-source alternatives; and Google Vertex AI aggregates both proprietary and third-party models.

MoMA represents the telecom industry's most significant bid to claim a piece of this market. While Western telecoms have largely remained on the sidelines of the AI platform wars, China Mobile — with its 990+ million mobile subscribers and extensive enterprise relationships — has the distribution advantage to make this play credible.

The emphasis on domestic computing infrastructure is also notable in the context of ongoing U.S. export controls on advanced AI chips. By building its inference engine on domestically produced hardware, China Mobile reduces its exposure to supply chain disruptions and positions MoMA as a sovereign AI infrastructure option — a selling point for government and defense-adjacent customers.

What This Means for Developers and Enterprises

For developers and enterprise technology leaders, MoMA's launch signals several important trends:

• Model commoditization is accelerating: When 300+ models are available through a single API with intelligent routing, individual model differentiation matters less than platform-level optimization
• Cost optimization is becoming a platform feature: Rather than leaving cost management to end users, platforms are building it into the infrastructure layer
• Telecom companies are serious AI contenders: China Mobile's entry suggests that telecoms' infrastructure advantages — data centers, network connectivity, enterprise relationships — translate meaningfully to AI service delivery
• Confidential computing is going mainstream: Hardware-based security for AI inference is moving from niche to expected feature for enterprise platforms

For Western companies operating in or competing with Chinese markets, MoMA represents a new type of competitor: not a model developer or a traditional cloud provider, but a telecom-scale aggregator that can leverage its existing infrastructure and customer base to rapidly capture market share in the AI services layer.

Looking Ahead: The Rise of AI Model Aggregators

MoMA's launch accelerates a broader industry trend toward model aggregation and orchestration. As the number of capable AI models grows — with new entrants appearing almost weekly — the value proposition shifts from 'which model is best' to 'which platform helps me use the right model at the right time for the right price.'

China Mobile has not yet disclosed specific pricing details or international availability plans for MoMA. However, the platform's scale — 300+ models with unified billing, intelligent routing, and enterprise-grade security — sets a high bar for competitors in both the Chinese and global markets.

The coming months will reveal whether other major telecoms follow China Mobile's lead. If MoMA proves successful, it could inspire similar moves from carriers like Deutsche Telekom, AT&T, or NTT — companies that share China Mobile's infrastructure advantages but have yet to make a comparable play in the AI platform space. For now, China Mobile has fired a significant opening shot in the telecom industry's AI ambitions.