China targets 9,800 exaflops of AI compute by 2030 — 4.5× today's level
China's Ministry of Industry and Information Technology has published the 15th Five-Year Plan for the information and communications industry, reported on September 7 and 8, 2026. Two numbers carry it. The first is a target of 9,800 exaflops of intelligent computing capacity by 2030. The second is 3.8 trillion yuan — about $532 billion — of cumulative information-infrastructure investment across 2026 to 2030. The document itself is dated August 12, 2026. At the end of June 2026 China's intelligent computing capacity stood at 2,185 exaflops, 177 percent higher than a year earlier, and 52 intelligent computing facilities each holding more than 10,000 accelerator cards had been built. The 2030 target is therefore roughly 4.5 times the mid-2026 level. The plan asks for the orderly deployment of clusters at the 10,000-card and 100,000-card scale, and for separate inference-oriented facilities tailored to particular applications. The reason this reaches Korea is simple: intelligent compute is assembled from accelerator cards, and accelerator cards carry memory
The three lines
- Target — 9,800 exaflops of intelligent compute by 2030, about 4.5× the 2,185 exaflops recorded at end-June 2026
- Money — 3.8 trillion yuan (~$532bn) of information-infrastructure investment across 2026–2030, per MIIT's five-year plan
- Method — orderly deployment of 10,000- and 100,000-card clusters; 52 facilities at 10,000-card scale already exist
Key questions
- What exactly did China announce
- **One capacity target and one investment figure.** | Item | Value | As of | |---|---|---| | Intelligent compute target | **9,800 exaflops** | by 2030 | | Current level | **2,185 exaflops** | end-June 2026 | | Year-on-year growth | **177%** | mid-2026 | | Cumulative infrastructure investment | **3.8 trillion yuan** (~$532bn) | 2026–2030 | | Facilities with 10,000+ accelerator cards | **52** | end-June 2026 | The issuing body is the **Ministry of Industry and Information Technology (MIIT)**, and the document is the **15th Five-Year Plan for the information and communications industry**. It is **dated August 12, 2026**, was posted publicly on **September 7**, and drew coverage on **September 8**. **'Intelligent computing' is a Chinese administrative term** for accelerated compute used in AI training and inference; it is counted separately from general-purpose server capacity.
- How large is 9,800 exaflops
- **One exaflop is 10¹⁸ operations per second.** So 9,800 exaflops is **9.8×10²¹ operations per second**. One way to get a feel for it is to convert to racks. Nvidia's **Grace Blackwell NVL72 rack is specified at about 1.4 exaflops** (「What Grace Blackwell NVL72 is」). Divide and you get **roughly 7,000 racks**. **That division should not be trusted as it stands.** AI compute figures depend entirely on the precision at which they are measured — FP8, FP16, INT8 — and can differ by multiples. **The precision basis of the Chinese statistic is not public.** **What is reliable is the ratio.** Going from 2,185 exaflops in June 2026 to 9,800 by 2030 is **about 4.5×** over four and a half years. **But the preceding twelve months already delivered 177% growth.** At that pace the target reads as conservative rather than ambitious.
- How does this connect to Korean semiconductors
- **Intelligent compute is built from accelerator cards, and accelerator cards carry memory.** The plan's own wording makes the link explicit: it calls for the **orderly deployment of 10,000-card and 100,000-card accelerator clusters.** Policy is being written in units of cards. And high-end accelerator cards do not exist without HBM (「What HBM is」·「What a memory supercycle is」). **That demand does not convert into Korean revenue automatically.** Three valves sit in between. **① Export controls** — top-tier AI chips and some HBM are restricted (「What AI chip export controls are」). **② Domestic substitution** — China is scaling its own accelerators and its own HBM in parallel, and Chinese HBM3E has already moved Korean share prices (「Samsung and SK Hynix fall 4% (September 2)」). **③ Plans are not orders** — a five-year target is a target, not a purchase. **The announcement tells you the direction of demand. It does not tell you who fills it.**
A country has set a numerical target for computing capacity. The target is 4.5 times what it has now.
1. The numbers
China's Ministry of Industry and Information Technology (MIIT) published its 15th Five-Year Plan for the information and communications industry on September 7, 2026; it was widely reported on September 8. The document is dated August 12.
| Item | Value |
|---|---|
| 2030 intelligent compute target | 9,800 exaflops |
| End-June 2026 actual | 2,185 exaflops |
| Year-on-year growth | 177% |
| Cumulative infrastructure investment 2026–2030 | 3.8 trillion yuan (~$532bn) |
| Facilities with 10,000+ accelerator cards | 52 (end-June 2026) |
The gap between target and current level is about 4.5×.
2. Policy written in units of cards
The most revealing detail in the plan is not the target but the unit. The document calls for the orderly deployment of 10,000-card and 100,000-card accelerator clusters, and for inference-dedicated facilities tailored to specific applications.
| Unit used in the plan | Meaning |
|---|---|
| 10,000-card clusters | the scale 52 sites already occupy |
| 100,000-card clusters | frontier-model training scale |
| Inference-dedicated facilities | serving, not training (「Training versus inference」) |
Separating training from inference at the level of a policy document matters. It signals that the market for inference-specific silicon is now planned for on its own terms (「What an AI inference chip is」).
3. Reading exaflops carefully
One exaflop = 10¹⁸ operations per second. 9,800 exaflops is 9.8×10²¹.
| Conversion | Arithmetic | Caveat |
|---|---|---|
| In Nvidia NVL72 racks | 9,800 ÷ 1.4 ≈ 7,000 racks | precision basis may differ |
| Against June 2026 | 9,800 ÷ 2,185 ≈ 4.5× | same basis — reliable |
Only the second line is safe. AI compute figures swing by multiples depending on whether they are quoted at FP8 or FP16, and China's basis is not published. Use the ratio; treat the rack conversion as intuition only.
4. The route to Korea, and the valves on it
More accelerator cards means more HBM (「What HBM is」). Formally, this plan is good news for Korean memory. Three valves sit on the pipe.
| Valve | Content | Direction |
|---|---|---|
| US export controls | top-tier AI chips and some HBM restricted | closes (「What AI chip export controls are」) |
| Domestic substitution | Chinese accelerators and Chinese HBM3E scaling | closes |
| Plan versus order | a five-year target is not a purchase | delays |
Nvidia's decision to model China revenue at zero in its third-quarter guidance (「Nvidia's Q2 revenue of $96.2bn (August 26)」) is what those valves look like in a financial statement. Demand growing and Korea capturing that demand are different propositions.
5. What remains unresolved
- The precision basis of the compute figures is not public. The 4.5× ratio holds; cross-system comparisons do not.
- The composition of the 3.8 trillion yuan is unclear — state fiscal spending or an aggregate including carriers.
- The original Chinese document was not read directly. Figures come from English coverage and state media summaries.
- The definition of '52 facilities' is ambiguous — operating sites only, or including construction.
- In the same week, Europe's largest private AI round was €3 billion (「Mistral raises €3bn led by Samsung (September 8)」). One company's record raise and one country's five-year infrastructure budget are two orders of magnitude apart — which is itself the story of AI competition moving from firms to states.
Sources
- South China Morning Post — China targets fourfold boost in AI computing capacity by 2030 in major tech push
- Unite.AI — MIIT plan targets 9,800 eflops of intelligent compute by 2030
- SCIO — China releases 5-year development plan for information, communications sector
- Success Stories News — China targets fourfold boost in AI computing capacity by 2030
- AI Weekly — AI News Today, September 8