A telecom budget becomes an AI-infrastructure signal
China Mobile has turned a procurement target into a market signal. At the 2026 Mobile Network Conference in Guangzhou, the operator set out plans for more than RMB500 billion in total procurement and investment in 2027, a figure Digitimes converted to roughly US$75.1 billion . The headline number is not a pure AI capex figure: the plan also includes more than RMB100 billion for computing power and more than RMB100 billion for terminal procurement and sales . But the direction is unmistakable. China’s largest mobile carrier is treating AI computing capacity not as a side business, but as a new layer of national-scale telecom infrastructure.
That matters because telecom operators buy differently from cloud start-ups. They purchase on long cycles, standardize across regions, and influence entire vendor ecosystems. A RMB500 billion shopping list can set expectations for accelerators, optical networking, liquid cooling, storage, edge nodes, power systems and integration services. In this case, China Mobile is also linking compute to the network itself: the company’s management framed future infrastructure around three coordinated networks — communications, computing power and intelligence .
What was actually announced
The latest reported figures came from China Mobile chairman Chen Zhongyue’s keynote at the Guangzhou conference, held from November 5 to 7, 2026 . Chen said the company plans to raise 2027 procurement volume above RMB500 billion, including more than RMB100 billion for computing-power investment, more than RMB100 billion for terminal procurement and sales, and RMB10 billion in channel resources . BofA Securities, in a note summarized by AASTOCKS, also described the conference as showing a clear pivot toward integrated communications, computing-power and intelligence infrastructure .
This distinction is important. The US$75 billion figure describes the full 2027 procurement and investment envelope, while the directly stated computing-power component is above RMB100 billion. BofA also warned that the RMB100 billion AI-infrastructure plan is a broader investment and procurement concept, and should not be compared mechanically with listed-company capex . In other words, investors should not read every yuan as immediate equipment spending, and suppliers should not assume a straight-line order book. But for the industrial chain, the direction of travel is still clear: computing has become the largest new investment category for China Mobile .
Digitimes framed the plan as a major shift from conventional telecom services toward large-scale AI infrastructure . That reading fits the Chinese-language reports from the event: Chen said the “new intelligent economy” needs new infrastructure, and described communications networks as routing traffic, computing networks as routing model workloads, and intelligent networks as routing tasks and intent .
From base stations to token factories
The most interesting part of the announcement is not simply that China Mobile will spend more. It is how the company is defining the infrastructure it wants to buy. The plan points to a future in which the telecom network is not only carrying data generated by AI applications, but also coordinating where AI workloads run, how models reach users, and how inference capacity is distributed across cloud and edge locations.
China Securities Journal reported Chen’s description of the AI-driven traffic shift: mobile internet traffic has risen 18% this year, twice last year’s growth rate, while AI-generated content has driven a fivefold increase in related traffic . He also said AI is changing the upstream-downstream balance: the overall network’s average uplink-to-downlink ratio is 1:9, while AI-application traffic has already reached 1:6 and some future applications may approach 1:1 . That is a classic carrier problem: if applications change traffic symmetry, latency and bandwidth planning must change too.
China Mobile’s proposed answer is an integrated computing network. The company plans to build multiple ultra-large intelligent-computing clusters after already constructing several 10,000-card clusters, and expects its intelligent-computing capacity to double in each of the next two years . It has started construction of three gigawatt-class intelligent-computing parks this year and plans two more next year, bringing the total to seven . China Securities Journal added that these parks will use standardized modular designs and high-density liquid-cooling technology .
That is why the procurement plan could ripple well beyond server vendors. Gigawatt-class campuses require power procurement, grid coordination, cooling systems, land, fiber routes, optical transport, storage, security and software orchestration. China Mobile also said it would expand 400G OTN capacity, connect national computing hubs and computing resources, increase use of hollow-core fiber, and upgrade CDN capacity . The objective is not just more chips; it is smoother access to compute, more efficient interconnection between compute resources, and faster edge distribution .
Domestic supply chains get a demand map
For Chinese AI hardware and infrastructure suppliers, the China Mobile plan is a demand map. The company’s ambitions touch almost every layer of the AI stack: accelerators, GPU-like domestic chips, high-speed interconnects, RoCE switching, storage, model-service platforms, optical systems, liquid cooling and data-center construction. Financial news service CLS reported that Huawei, ZTE, H3C, Moore Threads, xFusion and other compute-ecosystem vendors appeared at the conference exhibition, with attention on supernodes, GPUs and optical interconnect schemes .
China Mobile is also pushing software and protocol layers that could shape procurement requirements. China Securities Journal reported that the company has built a “Xinhé” heterogeneous super-fusion platform to mask differences among underlying hardware and support migration and deployment optimization of mainstream models across more than ten domestic chip types . It also introduced OISA, described as China’s first open protocol for GPU card-to-card interconnection, and GSE, an inter-machine interconnection protocol intended to improve cluster efficiency . These details matter because large buyers do not merely select products; they define compatibility rules.
The company’s “MoMA” model-service platform, inference-computing network and computing-power pools are designed to improve token output per accelerator card, according to the same report . That phrase — token output per card — is a useful signpost. China Mobile is thinking less like a traditional bandwidth utility and more like an operator of AI production capacity, where performance is measured not only in FLOPS, but in usable inference throughput, latency and utilization.
The strategic message
The announcement lands at a moment when AI infrastructure is becoming a national industrial priority. But China Mobile’s role is different from that of a hyperscaler. A telecom operator controls access networks, transport networks, edge sites, billing relationships, enterprise channels and government-linked infrastructure mandates. If it defines AI capacity as part of the network, then compute may be procured, routed and sold with the discipline of telecom infrastructure.
That does not remove execution risk. The RMB500 billion plan is a target, not completed spending. The RMB100 billion-plus computing line is a broad category, not a detailed supplier-by-supplier award . Gigawatt-class parks also require electricity, cooling and utilization at scale. Buying capacity is easier than filling it with profitable workloads.
Still, the signal is powerful. China Mobile is telling equipment makers, chip suppliers, data-center builders and software partners that AI compute is moving into the core procurement calendar. The carrier is also telling the market that future telecom competition may be fought not only over subscribers and spectrum, but over who can deliver compute, models and intelligent services with carrier-grade reliability. That shopping cart will definitely require more than 64 bits.
Comments
Be the first to comment.