9,800 EFLOPS by 2030 Against 2,185 Today (September 2026): China Wrote the 100,000-Card Cluster Into a Five-Year Plan — and the Binding Constraint Is Memory, Not Design

On Monday 8 September 2026 China's Ministry of Industry and Information Technology published its five-year plan for the information and communications industry, and among its 13 key indicators is intelligent computing power of 9,800 EFLOPS by 2030 — against 2,185 EFLOPS at the end of June this year. The plan also asks, in its own language, for the orderly deployment of computing clusters of 10,000 accelerator cards and of those using 100,000 or more. On Tuesday, JD Cloud announced one. The interesting question is not whether Beijing wants a 100,000-card domestic cluster; the document says it does. It is which physical input decides whether the cluster arrives — and on the published evidence that input is memory, not chip design.

What actually happened, in two days

The sequence matters more than either item alone. On 8 September the Ministry of Industry and Information Technology issued its development plan for the information and communications industry across the 15th Five-Year Plan period, 2026 to 2030. Xinhua's summary of the document lists 13 key indicators: industry revenue of 4.1 trillion yuan (about US$604.8 billion) by 2030, cumulative information-infrastructure investment of 3.8 trillion yuan, average annual growth of 7% in total telecommunications business volume, 50 5G base stations per 10,000 people, 95% 5G user penetration — and intelligent computing power of 9,800 EFLOPS. Twenty-six key tasks sit underneath, including starting 6G commercial services at an appropriate time and building a nationwide integrated computing-power network.

Buried in the cluster language is the line the chip industry read first. The South China Morning Post reported that the plan calls for the orderly deployment of intelligent computing clusters with 10,000 graphics processing cards or those using 100,000 or more cards, and noted that China had already constructed 52 intelligent computing facilities each equipped with more than 10,000 accelerator cards.

The next day, at JD's 2026 Global Technology Explorers Conference, JD Cloud said it would build a cluster of 100,000 general-purpose GPUs supplied by Moore Threads. Moore Threads said in its own statement that it would be the first time domestically developed GPUs are used in a 100,000-card core computing cluster at a leading Chinese AI cloud provider. The workloads named were large-model training and inference, embodied intelligence and supply-chain AI; the capacity is to be rented out to third parties in the same way JD already sells cloud built on Nvidia hardware. The two firms have already run a 10,000-GPU cluster together, so this is a scale-up of a tested partnership rather than a cold start.

The arithmetic: 520,000 cards, a 4.5x, and a unit with no precision attached

Start with the national numbers, because they frame everything else. MIIT spokesperson Xie Cun told a 20 July 2026 press conference on first-half industrial development that intelligent computing capacity had reached 2,185 EFLOPS by the end of June, up 177% year on year. The plan's 9,800 target is therefore about 4.5 times the June 2026 level, to be delivered by 2030.

Set that against the card base. Fifty-two facilities at more than 10,000 cards each puts the floor somewhere above 520,000 accelerator cards. If per-card performance were frozen at today's average, a 4.5x capacity target would imply a fleet on the order of 2.3 million card-equivalents by 2030 — arithmetic, not a forecast. Per-card performance will not be frozen, and that is precisely where the slack in the target lives: a large share of the multiple can be delivered by newer silicon rather than by more buildings.

Then the unit itself. An EFLOPS is a rate of floating-point operations, and a floating-point operation is not one thing: the same die scores roughly twice as high at FP8 as at FP16 and higher again against FP32. A national capacity total published without a stated precision can therefore move by a factor of two or four on definition alone. This is not an accusation of bad faith — every country's compute statistics have the same problem — but it is the first question to ask of any headline compute number, including the ones in Western capex disclosures. Moore Threads markets its flagship on hardware-native FP8 support, which improves the score and the workload at the same time.

The vendor's own datasheet stops at 10,000 cards

The most useful document in this story is not the policy plan. It is Moore Threads' own product page for the MTT S5000, its flagship accelerator, built on the PH100 chip and the "PingHu" architecture with full-precision support from FP8 to FP64.

Read what the company claims, and read where it stops. The S5000 is described as delivering near-linear scaling from a single GPU to a 10,000-GPU cluster, with cluster linearity of up to 95% at that scale. In core computer-vision and large-language-model training it is claimed at better than 74% of the performance of leading international flagship products. Model FLOPs Utilisation is quoted above 60% on Llama3-70B and above 40% on DeepSeek-236B. In a DeepSeek-236B run on a 10,000-GPU cluster, the loss curve is said to hold a relative precision deviation of 0.6% against clusters running international flagships over the first 30,000 steps.

Every one of those numbers is scoped to ten thousand cards. Neither company has published a scaling claim at a hundred thousand. That is the honest state of the evidence, and it is the gap between a policy target and a working machine: interconnect topology, collective-communication efficiency and failure rates do not scale linearly with card count, which is why the company's own marketing material treats 10,000 as the demonstrated ceiling and quotes its reliability gain at that scale too. On the vendor's own relative-performance figure, 100,000 of its cards is roughly 74,000 international-flagship equivalents — again arithmetic, and again dependent on the workload the 74% was measured on.

The corporate scale tells the same story. Moore Threads listed on Shanghai's STAR Market in December 2025 after an IPO that raised about 8 billion yuan; the shares closed their debut session at 600.5 yuan against a 114.28 yuan offer price. Its interim report for the six months to June 2026 shows revenue of RMB 1.74bn, up about 147% year on year and already above its RMB 1.51bn full-year 2025 total, with research and development of RMB 769m absorbing 44.3% of revenue and a first quarterly net profit of RMB 29m in the March quarter. A company at that revenue run-rate has never shipped anything close to a hundred thousand flagship accelerators in a year. Growing into that order is possible; it has not happened yet.

Where the constraint actually binds

A training accelerator is a memory machine with arithmetic attached. Bandwidth, not raw FLOPS, is what keeps tensor cores fed during a large-model run, which is why every serious training part is built around stacked high-bandwidth memory. It is also the single component the United States moved hardest to restrict, expanding controls on HBM and on the tooling used to fabricate and package it in December 2024.

Domestic substitution has begun, and it is early. Korea JoongAng Daily reported at the start of September 2026 that ChangXin Memory Technologies had begun producing HBM3E for AI accelerators, is targeting Chinese customers and aims to ramp production next year — while still trailing the leading memory makers by a wide margin in its ability to mass-produce it. Small-scale production with a ramp targeted for next year is not the same input as an existing supply line.

That is the mechanism in one line: the card count in an announcement is set by ambition, and the card count that ships is set by memory. It is the same squeeze this site traced through the cheap end of the market when AI data centres started repricing mature-node wafers — scarcity migrating outward from the leading edge into everything that shares a supply chain with it. A second national buildout bidding for stacked memory does not only affect Chinese clusters; it competes for the same finite packaging and DRAM capacity that every hyperscaler's bill of materials depends on.

What it does, and does not do, to the numbers people trade

The instrument this story touches is the US semiconductor complex and the indices it dominates — not a currency pair. Handle it precisely.

Nvidia reported revenue of $96.221bn for the quarter ended 26 July 2026, up 106% year on year, with Data Center revenue of $89.0bn, and guided the October quarter to $108.0bn plus or minus 2%. The release says, in its own words, that the company is not assuming any Data Center compute revenue from China in that outlook. Guidance and price therefore already carry a zero on the China line. News that a Chinese cloud intends to build a domestic cluster cannot subtract from a number that is already zero — which is why the immediate index reaction to announcements like this is usually nothing, and why reading a same-day price move as confirmation or refutation is a category error.

Where it can reach a valuation is the long horizon, and Nvidia's own chief financial officer put that channel on record. Speaking after the January-quarter results, Colette Kress said the company had yet to generate any revenue from the small volume of H200 products approved for China-based customers, adding that "we do not know whether any imports will be allowed into China" — and, on competition, that competitors in China, "bolstered by recent IPOs, are making progress and have the potential to disrupt the structure of the global AI industry over the long-term", as CNBC reported. Moore Threads is one of the listings she was describing. A market temporarily closed by licensing can reopen; a market that has built a working alternative stack does not.

For an index reader the practical translation is narrow. A US500 or NAS100 position is a position in the size of the global AI capital-expenditure cycle and in how much of it lands on US suppliers. This week's news changes neither the near-term cycle nor the reported numbers; what it moves is the probability distribution around the terminal share, several years out, of one geography. That is worth understanding, and it is not a trade — the same distinction drawn in the August quarter breakdown, where a single stock's 8.7% move accounted for most of an index day.

Where this reaches a currency, and where it does not

Honestly: barely, and the barely is worth stating so nobody stretches it. None of the eight currencies the meter scores is a semiconductor proxy. The memory supply chain runs through Korea and Taiwan; the lithography and deposition tooling runs through the Netherlands and Japan. The won and the Taiwan dollar are the currencies with direct exposure, and neither is a major.

Two indirect channels do exist and both are thin. The first is the dollar through growth: AI-related construction and equipment spending has become a visible contributor to US fixed investment, and growth is one of the five factors the meter scores — so anything that changes the medium-term size of that capex cycle eventually shows up there, with a lag measured in quarters, not headlines. The second is the yen through Japan's semiconductor-equipment export exposure, which reaches the trade balance far more than the rate differential that actually drives the pair. Neither channel is readable off a policy document, and treating them as such is how a good macro idea becomes a bad position.

What would change the picture

Four observable things, in rough order of information value. First, an acceptance milestone: a published benchmark, a customer running a training job, or a stated in-service date for the JD Cloud cluster would convert an intention into a machine. Second, the memory ramp — evidence that ChangXin's HBM3E has moved from small-scale to volume, and at what yield, because that number caps everything downstream. Third, MIIT's next capacity print at the half-year press conference, which tells you whether the run-rate toward 9,800 EFLOPS is tracking or slipping, and whether the ministry starts quoting a precision alongside it. Fourth, and most directly readable in a share price: whether Nvidia reinstates a China Data Center assumption in a future outlook, which would mean the licensing channel reopened, or continues to guide from zero, which means the substitution story keeps running unopposed.

Until at least the first two arrive, the correct reading of this week is modest and specific. A government published a target that includes 100,000-card clusters, and a cloud provider announced one the following day using a supplier whose own published scaling evidence stops at 10,000. Both facts are real. Neither is a machine.

Educational macro context only — not investment advice.