Moore Threads Q3: Sales Up 1,774% and a 100,000-GPU Cluster With JD

Moore Threads, the Chinese GPU maker listed on Shanghai's STAR Market, reported third-quarter revenue of 1.552 billion yuan, up 1,773.92% from a year earlier, and a net profit of 49.26 million yuan. A year ago it lost 453 million yuan in the same quarter, per ChainCatcher's summary of the filing. Bloomberg rounds the revenue to about 1.6 billion yuan, around $240 million, and Techmeme's summary of that report adds that the company is building a domestic 100,000-GPU cluster with JD.com. The growth rate is dramatic mainly because the base was tiny. What matters for AI builders is that a Chinese GPU vendor is now selling at a scale that shows a real second hardware track.

We cover the numbers, the cluster, and the limits of what the filings prove.

TL;DR: the Moore Threads numbers

What do the figures actually say?

The quarterly jump comes against a weak comparison period. Moore Threads reported a 453 million yuan loss in the third quarter of 2025, and its first-half 2026 revenue of 1.736 billion yuan had already exceeded its full-year 2025 total, according to reporting on its semi-annual results. Quarter on quarter the picture is steadier: first half revenue averaged about 868 million yuan a quarter, so the third quarter's 1.552 billion yuan is roughly 1.8 times that average. Still strong, but a more sober way to read it than a four-digit percentage.

The profit is small. Forty-nine million yuan on 1.55 billion yuan of revenue is a margin of about 3%, and one outlet, Futu News, gives a non-GAAP profit figure of 20.07 million yuan for the quarter, so definitions differ. The company had been loss-making through 2025, with research spending running above a billion yuan a year, per TrendForce's earlier report on its first-quarter turn to profit. A profit on this scale tells you the business covers its costs at current volumes; it does not yet show strong pricing power.

What is the JD Cloud 100,000-GPU cluster?

On September 9, 2026, at JD's Global Technology Explorers Conference, JD Cloud and Moore Threads announced plans for a cluster of 100,000 GPUs built from Moore Threads silicon, designed for large-model training, inference and embodied-intelligence workloads, per DatacenterDynamics and AI Weekly. The coverage describes it as the first deployment of Chinese GPUs at that scale inside a major domestic AI cloud. The two companies already operate a 10,000-GPU cluster together.

What we could not find: the specific chip model for the 100,000-GPU system (the MTT S5000 is Moore Threads' flagship accelerator, but none of the JD sources we read confirm it is the part used), a delivery timeline, interconnect details, or a price. So treat "100,000 GPUs" as an announced plan. Our reading of Bloomberg's Oct 28 report comes from its headline summary, not its full text.

Why are Chinese buyers moving to domestic GPUs?

Two pressures push in the same direction. US export controls limit which Nvidia parts can be sold into China, and one German report says Beijing has guided state-linked projects away from Nvidia's H20 and toward domestic accelerators; the sources we found do not detail that guidance, so consider it reported rather than confirmed. Moore Threads itself is on the US Entity List, which restricts its access to some foreign technology and manufacturing.

The competitive field is crowded. Huawei and Cambricon also chase the same spending, which we looked at in the IDC-based comparison of Nvidia and Huawei's China AI chip share and in the DeepSeek and Huawei Ascend co-design report. Moore Threads' market value reportedly fell from about 37 billion euros to under 24 billion euros after its lock-up expired on September 7, per one German outlet, which shows investors price the stock on expectations, not only on revenue. We could not verify that figure independently.

How should you read a 1,774% growth headline?

Percent growth from a small base is the least informative number in an earnings release. Moore Threads listed on the STAR Market in December 2025, per Bloomberg's description of this quarter as its fastest growth since the listing, and its third-quarter 2025 revenue was only about 83 million yuan if you back it out of the reported 1,773.92% increase. That is why a quarter that adds roughly 1.47 billion yuan produces a figure that looks absurd.

Better questions to ask of the release:

• Absolute scale. Nine-month revenue of 3.289 billion yuan is a few hundred million US dollars. That is meaningful for a company that was loss-making a year ago, and small beside the data-center revenue of the largest US chip vendors.
• Concentration. Large cluster orders, such as the 660 million yuan KUAE order reported in March, can swing a quarter. We did not find a customer breakdown for Q3, so we cannot say how much came from a handful of buyers.
• Repeatability. A second quarter above 1.5 billion yuan would suggest a run rate; one quarter could be a lumpy delivery.
• Margins. A profit of about 3% of revenue leaves little room for price pressure from Huawei, Cambricon or discounted Nvidia parts.

The same caution applies to the stock. A 5.84% rise on the day, to a market value of about 179.3 billion yuan, tells you investors rewarded the report, not that the valuation is justified by current earnings.

What are people asking?

Is this an AI-model story or a chip story? A chip story with model consequences. Training and serving frontier-scale models needs clusters of tens of thousands of accelerators, and until recently Chinese labs relied heavily on Nvidia hardware. Each domestic cluster that works reduces that dependence.

Can developers use Moore Threads GPUs? Moore Threads sells its own software stack and has said its clusters operate in several Chinese cities, per reporting on its half-year results. Most readers outside China will not buy them directly; the effect arrives through Chinese cloud providers and model labs.

Does the cluster count as a training run? No. A cluster announcement is infrastructure. Whether a frontier-class model gets trained on it is a separate claim that nobody has made yet.

What does it mean for people building with AI?

Three practical takeaways.

1. Hardware diversity is real inside China. Models trained there may increasingly run on non-Nvidia stacks. If you evaluate Chinese open-weight models, expect inference tooling to target several accelerators, not just CUDA. Our coverage of China's compute capacity shows how large that buildout is.
2. Memory remains the choke point. Chinese GPUs depend on high-bandwidth memory, where domestic yields are still reported as low, as in our post on CXMT's HBM3e yields. Meanwhile the global shortage keeps lifting suppliers such as SK Hynix, covered in our Q3 earnings post.
3. Policy keeps moving. Taiwan has moved to criminalize some AI chip server exports to China, as in the Foreign Trade Act proposal, so supply chains for all of these companies are under pressure from both sides.

Claimed vs. verified

To check the core numbers, read the company's Q3 report on the Shanghai Stock Exchange (ticker 688795). For the cluster, watch for a JD Cloud or Moore Threads technical post naming the chip and network fabric.

What to watch next

• Fourth-quarter guidance and margins. Whether the profit holds as the company scales.
• Independent benchmarks. Training throughput and inference cost per token against Nvidia and Huawei parts.
• The JD cluster. First phase delivery and a named chip.
• Export policy. New US or Chinese rules on what can be sold or used.

Related reading

• Nvidia vs Huawei China AI chip share (IDC)
• DeepSeek and Huawei Ascend co-design
• China's 24 GW compute vs the US
• CXMT HBM3e yields
• China STAR50 AI stocks surge
• SK Hynix Q3 and the memory shortage
• Taiwan's AI chip export proposal

Figures are as reported on October 28, 2026. Currency conversions follow the cited outlets.