The plan released September 7. One line deserves a circle.
The Ministry of Industry and Information Technology issued the 15th Five-Year Plan for the information and communications industry. Thirteen main indicators, 26 key tasks, six dimensions. For anyone building AI infrastructure, the line to underline is this: by 2030, smart-compute scale reaches 9,800 EFLOPS.
The numbers, on the table first
By 2030, industry revenue hits 4.1 trillion yuan. Cumulative infrastructure investment over five years reaches 3.8 trillion yuan. Total telecom business volume grows 7% a year. Smart compute reaches 9,800 EFLOPS, a cumulative gain of 8,210 EFLOPS over the period. 5G including 5G-A user penetration hits 95%. Half a million new 5G-A base stations get built. Advanced storage reaches 1,700 exabytes. Fifty 5G base stations per 10,000 people. One million 50G-PON ports.
What is 9,800 EFLOPS? An EFLOPS is 100 quadrillion floating-point operations per second. A cumulative gain of 8,210 means compute is being laid down as a strategic resource at the national level. This is not a slogan.
The directions matter more than the digits
The plan is explicit: launch 6G commercially at the right time, accelerate a national integrated compute network, explore space-based compute, push 5G plus Beidou fusion, advance eSIM and off-grid communication in an orderly way, back artificial-intelligence-plus-manufacturing, build an industry anti-fraud lab, construct an identifier system for agent internetworking, and stand up an emergency-communications large model.
Several phrases matter directly to AI practitioners. An identifier system for agent internetworking means rules get set first for how agents talk to each other. AI-plus-manufacturing and AI safety governance are listed as standalone tasks. A power-usage-effectiveness ceiling below 1.2 on new large compute facilities puts a hard energy bar on data centers.
The red lines are written too
The plan repeats technologically advanced and securely controllable. Adapting to domestic compute chips is named explicitly. For players who stacked foreign accelerators and waited for policy to loosen, the balance is tilting toward domestic silicon. At the same time, steadily and prudently expanding autonomous opening in telecom means the door is not fully closed, but the initiative sits on the domestic side.
What this means for you
If you run an AI-related business, this plan is a five-year demand map. Compute, storage, 6G, satellite communications, agent internetworking: these directions will keep drawing projects and budgets.
But do not read the plan as an order book. Execution depends on follow-on ministry rules and on how central enterprises and local governments move, and there is distance in between. What you can actually do: lean your product toward the two certain directions, domestic compute adaptation and agent internetworking, rather than betting on a specific subsidy.
What these numbers mean as business
9,800 EFLOPS is not an abstract metric. Behind it sits a whole supply chain of orders: data centers, servers, optical modules, liquid cooling, power. A cumulative gain of 8,210 EFLOPS means new compute built every year of the plan. Firms in compute operations, liquid cooling, storage and networking have five years of reasonably certain demand ahead.
A 95% 5G-A penetration target and half a million new 5G-A base stations feed the radio-frequency, antenna and small-cell chain. Advanced storage of 1,700 exabytes maps directly onto the data flood from AI training. You cannot train what you cannot store.
A note for startups
The plan's named identifier system for agent internetworking and its AI safety governance are slots not yet occupied by the giants. Issuing identities to agents, authenticating them, auditing them: this picks-and-shovels work could spawn new companies in the next two years. But the pace follows the detailed rules. Do not read the wind as a contract.
One overlooked point: a PUE below 1.2 is now a hard bar. That means high-energy legacy data centers must be retrofitted or retired. Green compute is not a slogan; it is an entry threshold. For training jobs sensitive to power cost, siting and efficiency will shape margin more than simply stacking accelerators.
How to read the red lines
The repeated emphasis on securely controllable is not decorative. Adaptation to domestic compute chips is named outright, which signals where procurement will lean. Foreign-accelerator stacks are not banned, but the initiative sits with domestic silicon. If your roadmap assumes unfettered access to the latest foreign hardware, the plan tells you to build a fallback.