NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

A Newly Listed Chinese GPU Maker Reported Its First Numbers. The Growth Is Real; the Profit Isn't.

Denglin Technology's first update since its Hong Kong listing shows revenue up nearly sixfold and board shipments above 27,990 — but no profit figure, and peers four to six times its size.

2026-11-20 · 1185 words · NeuroAI
A Newly Listed Chinese GPU Maker Reported Its First Numbers. The Growth Is Real; the Profit Isn't.

A company that listed in Hong Kong a week ago published a short note on a Wednesday morning. It contained no product photograph and no victory speech, just a few operating numbers that investors had been waiting to see.

The figures were strong in one direction and unfinished in another. That combination describes much of China's domestic inference-chip industry right now.

What the Announcement Actually Says

Suzhou Denglin Technology Co., Ltd. (苏州登临科技股份有限公司, 09917.HK) filed a voluntary announcement on 18 November 2026, roughly a week after its listing. The document covers the nine months to 30 September 2026 and states four things:

  • Revenue: more than RMB 500 million (≈ US$75M / HK$585M) on an unaudited basis, up about 592 percent year on year.
  • Customers: 123 during the period.
  • Concentration: the three largest customers accounted for 50.7 percent of revenue; the five largest for 68.8 percent.
  • Shipments: more than 27,990 boards delivered.

The filing carries a standard warning: these figures are unaudited, have not been confirmed by the company's auditor, and may be adjusted.

Set against the prospectus published on 30 October, the shape of the growth becomes clearer. Denglin booked roughly RMB 28 million (≈ US$4.2M / HK$33M) in 2023, RMB 83 million (≈ US$12M / HK$97M) in 2024 and RMB 171 million (≈ US$26M / HK$200M) in 2025. In the first four months of 2026 alone it took in about RMB 200 million (≈ US$30M / HK$234M). Since the nine-month figure exceeds RMB 500 million and the four-month figure was about RMB 200 million, the May–September period contributed well over RMB 300 million — a rough estimate from two documents with different cut-off dates.

Denglin makes inference (推理) GPUs for industry large models (大模型), enterprise private clouds, smart-city systems and industrial inspection. Inference is the unglamorous half of AI computing: running a trained model millions of times, where cost per request matters more than peak benchmark scores.

The Customer Mix Is the More Interesting Number

A 592 percent growth rate invites excitement. The customer concentration figures invite a different kind of attention.

In the four months to April, Denglin's top five customers made up 90.4 percent of revenue. By September that had fallen to 68.8 percent, with the top three at 50.7 percent. The four-month window showed 90.4 percent and the nine-month window 68.8 percent; because the periods differ, the decline is indicative rather than a like-for-like comparison.

That is a meaningful change, but it is also a reminder of where the company started. A business selling expensive accelerator boards to a short list of large clients can look like a platform while remaining a project business. Diversification is real progress; it is not yet proof of repeatable, product-led demand.

The 27,990 board figure points the same way. It is a cumulative count, not a quarterly run rate, and it covers all of the first nine months. Without a breakdown by product or customer, it is hard to tell how much of that volume came from a single deployment.

The Peers Who Got There First

Denglin is not entering an empty market. Three domestic GPU companies listed on the mainland reported their own nine-month figures over the same period:

  • Moore Threads: revenue of RMB 3.289 billion (≈ US$492M / HK$3.84B).
  • MetaX: RMB 2.824 billion (≈ US$422M / HK$3.29B).
  • Enflame: RMB 2.332 billion (≈ US$349M / HK$2.72B).

Each is roughly four to six times Denglin's size. The gap matters because inference customers tend to buy across a portfolio, and the larger vendors can spread software-adaptation costs over far more units.

There is a second difference. Moore Threads reported a first quarterly profit on a non-recurring basis in the third quarter, on revenue of RMB 1.552 billion (≈ US$232M / HK$1.81B). MetaX said its third-quarter non-recurring net profit was positive for the second quarter in a row. Both have passed a threshold Denglin has not yet reached.

The Profit Question

Denglin's announcement contains no net profit figure. The prospectus does: losses of RMB 332 million (≈ US$50M / HK$388M) in 2023, RMB 297 million (≈ US$44M / HK$347M) in 2024 and RMB 281 million (≈ US$42M / HK$328M) in 2025, narrowing to RMB 16.47 million (≈ US$2.5M / HK$19M) in the first four months of 2026.

The direction is encouraging and the absolute numbers are not small. Losing nearly RMB 300 million a year while selling RMB 171 million of hardware is a business still being financed by capital markets, not by customers. The narrowing in 2026 suggests improving unit economics, but four months is a short window and the latest announcement deliberately does not extend it.

This is why the filing is more useful than a product launch. It tells you where the company is on the two axes that matter: commercial traction and the path to profitability. Strong revenue growth plus falling concentration is good news. No updated profit figure is a gap, not necessarily a warning — but it is a gap a reader should keep in view.

Key takeaways

  • Denglin's first post-listing filing reports more than RMB 500 million (≈ US$75M / HK$585M) in nine-month revenue, up about 592 percent.
  • It shipped over 27,990 boards and served 123 customers in the period.
  • Customer concentration fell: the top five were 90.4 percent of revenue in the four-month window and 68.8 percent in the nine-month window; the periods differ, so the decline is indicative rather than a like-for-like comparison.
  • The announcement includes no profit figure; the prospectus shows losses in every full year from 2023 to 2025.
  • Three listed peers reported nine-month revenue four to six times larger.

Honest limitations

All figures in the 18 November filing are unaudited and flagged by the company as subject to adjustment. I have not seen a subsequent audited statement.

The year-on-year growth rate of 592 percent compares a nine-month period with the equivalent period a year earlier; the underlying base was small, so a high percentage is less informative than the absolute revenue. The prospectus figures and the filing figures come from two different documents with different cut-off dates, and I have not reconciled them line by line.

The peer comparisons use nine-month revenue figures reported by other companies and covered in the financial press. They are not like-for-like in product mix: some peers sell training as well as inference hardware. The inference-versus-training distinction is therefore a broad characterisation, not a precise segmentation.

Conversions use approximate rates of US$1 = RMB 6.69 and US$1 = HK$7.8. Actual reported values will differ. No share-price movement is discussed here; a one-week-old listing is too volatile for that to be useful.

What readers can do now

  • Read the concentration figures before the growth rate. A large customer share that is falling is a better sign than a large share that is merely growing in absolute terms.
  • When a company omits profit from an update, check what the prospectus said. The absence is not an answer, but the earlier document gives you a baseline to compare against.
  • Compare inference vendors on shipped units and software support, not launch claims. Adaptation work across model frameworks is where much of the real cost sits.

Related coverage

Cite this article

NeuroAI (脑机网), "A Newly Listed Chinese GPU Maker Reported Its First Numbers. The Growth Is Real; the Profit Isn't," 20 November 2026, https://neuroai.site/a/na-chips-denglin-inference-gpu-results

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Chips & Compute” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories