Cambricon, Moore Threads and MetaX, three Chinese AI chipmakers listed on Shanghai’s STAR Market, reported combined revenue of 6.58 billion yuan for the third quarter of 2026, about 3.1 times the 2.13 billion yuan of a year earlier, by our arithmetic from their quarterly reports. Cambricon’s revenue rose 104.47% to 3.53 billion yuan, Moore Threads’ 1,773.92% to 1.55 billion and MetaX’s 367.12% to 1.50 billion.
The numbers sit under a claim made on SemiAnalysis’s emergency episode with ChinaTalk, posted on YouTube on 13 November after Nvidia’s chief executive resigned: that China’s move onto its own accelerators is happening “with or without the remote access laws being passed”. We checked the episode’s three examples against the companies’ own documents. Zhipu’s seven chip platforms and DeepSeek’s Ascend kernels check out; Meituan’s own text says its model trained on “AI ASIC superpods” without naming Huawei.
Business Pill · Switching Costs Explained
An evergreen Business Engineer explainer on switching costs, why it is hard to leave an AI supplier. It teaches the concept only and does not discuss Nvidia, Huawei or any company in this piece.
The key insight: As we read the primaries, the episode’s claim that China is moving off Nvidia holds at the software and revenue layers: Zhipu and DeepSeek published Ascend-ready code, and three listed chipmakers tripled their quarterly revenue to 6.58 billion yuan, by our arithmetic. Meituan’s own text does not say Ascend.
The episode, titled “Is NVIDIA Over? EMERGENCY Pod with ChinaTalk”, features SemiAnalysis’s Dylan Patel and Jordan Nanos and ChinaTalk’s Jordan Schneider. Its description says they discuss “China’s push onto domestic accelerators” and adds that “Opinions are the speakers’ own”. We quote YouTube’s automatic captions, which do not label speakers.
In the chapter on China’s chips, a speaker said “China is trying to get off GPUs” and, citing SemiAnalysis tracking, that in February “GLM was released with support for seven domestic accelerators”. The same speaker said DeepSeek had published a “mega kernel with Huawei Ascend for their V4 model” and that Meituan claims to have trained its LongCat model entirely on Ascend.
Later, a speaker said “the gap in the model capabilities is wider at this point, but the gap in chip capabilities is like as close as it’s ever been”, adding that Ascend chips are “pretty comparable to Hopper chips”. The policy fight behind the episode, including the Remote Access Security Act, is mapped in our guide to 15 chip export bills.
Third-quarter revenue of Cambricon, Moore Threads and MetaX, billion yuan, 2025 against 2026. The 2026 figures are from each company’s Q3 2026 report on the Shanghai STAR Market; the 2025 figures and the combined totals (2.13 billion to 6.58 billion yuan, about 3.1 times) are our arithmetic from the growth rates each report gives.
Three Claims Checked at the Source
Zhipu, seven platforms: confirmed. The GLM-5 technical report by Zhipu AI and Tsinghua University, submitted to arXiv on 17 February 2026, says GLM-5 is “full-stack adapted to Chinese GPU ecosystems”, with optimisation “across seven mainstream domestic chip platforms, including Huawei Ascend, Moore Threads, Hygon, Cambricon, Kunlunxin, MetaX, and Enflame”. The work it describes runs “from underlying kernels to upper-level inference frameworks”, so it is about running the model, and the report uses Huawei’s Ascend Atlas series as its case study.
The report says that, through those optimisations, “GLM-5 on a single Chinese node achieves performance comparable to dual-GPU international clusters, while reducing deployment costs in long-sequence scenarios by 50%”. It fitted the model onto “a single Atlas 800T A3 machine” with W4A8 quantisation.
DeepSeek, Ascend kernels: confirmed, with a caveat. DeepSeek’s DeepGEMM-Ascend repository, first released on 30 September 2026, is “a port of DeepGEMM to the HUAWEI Ascend platform” that supports “BF16, FP8, FP4 GEMM, MQA logits, and MegaMoE”. It describes MegaMoE as a kernel that “fuses EP dispatch, two grouped GEMMs, SwiGLU, and combine”. A companion DeepEP-Ascend library handles expert-parallel communication on Ascend NPUs.
The README benchmarks cases that “cover inference and training workloads of the DeepSeek model series”, measured on Ascend 950DT; it speaks of the model series rather than V4 by name. Our earlier piece on DeepSeek V5 and Ascend pretraining covers reports about the next model.
Meituan, trained on Ascend: not in its own text. Meituan’s LongCat-2.0 blog says the 1.6-trillion-parameter model was “pre-trained on over 50K AI ASICs” and that “Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods”. It says its accelerators have “significantly less per-device memory than an H800 (80 GB)”. The blog, as we read it, does not name the chip’s maker. Meituan’s NPU serving code says it was “Tested on the Ascend A2”, which covers deployment, not training.
Ascend vs Hopper on Paper
Huawei Ascend 950DT figures from Huawei’s Huawei Connect 2025 keynote of 18 September 2025 (144 GB, 4 TB/s, 2 TB/s total interconnect, 1 PFLOPS FP8) against Nvidia’s H200 SXM page (141GB, 4.8TB/s, NVLink 900GB/s, 3,958 TFLOPS FP8 with sparsity). Ratios are our arithmetic; the two companies define interconnect and compute differently, so the comparison is on paper only.
To test “pretty comparable to Hopper chips”, we set Huawei’s published figures for the Ascend 950DT against Nvidia’s for the H200 SXM, a Hopper-generation part. Huawei’s Huawei Connect 2025 keynote of 18 September 2025 gives the 950DT “144 GB of memory and 4 TB/s memory access bandwidth”, total interconnect bandwidth of 2 TB/s, and 1 PFLOPS in FP8. It said the chip would be available in the fourth quarter of 2026.
Nvidia’s H200 page gives the H200 SXM 141GB of memory, 4.8TB/s of memory bandwidth, NVLink at 900GB/s and 3,958 TFLOPS in FP8, a figure it marks “With sparsity”. By our arithmetic, the 950DT has 102% of the H200’s memory, 83% of its memory bandwidth and 2.2 times its NVLink figure. Huawei’s 1 PFLOPS is 25% of Nvidia’s with-sparsity FP8 figure, which is not a like-for-like comparison.
DeepSeek’s own measurements on the 950DT give a working number. Its README lists a dense BF16 matrix multiply at 431 TFLOPS against a hardware limit of 432, and FP8 at 861 TFLOPS against 865, and says Dense GEMM “reaches up to” 99.8% of the hardware limit.
The Revenue Line
As we read them, the three STAR Market reports show the demand side. Cambricon’s report gives third-quarter revenue of 3,530,828,294.64 yuan and net profit attributable to shareholders of 1.40 billion yuan, up 147.74%. Moore Threads’ report gives 1,552,307,664.26 yuan, and quarterly net profit attributable to shareholders of 49.26 million yuan. MetaX’s report gives 1,500,183,034.23 yuan and net profit attributable to shareholders of 26.53 million yuan.
Earlier in the episode, a speaker, presenting what they called the case Nvidia’s former chief executive had made, said that if you add up the application-layer companies running on open models, “it’s actually well north of $10 billion” in annual recurring revenue. The speaker framed it as playing devil’s advocate. We did not verify the figure, and none of the documents we read gives it.
The Structural Read
As we read the code releases, the cost of leaving Nvidia is moving into open repositories: DeepSeek says its Ascend port is “fully API-compatible with DeepGEMM”, so the same calls run on either platform, and Zhipu adapted GLM-5 to seven Chinese platforms rather than one.
As we read the spec ratios, the gap on paper is in memory bandwidth and in how the two companies state compute, not in memory size: the 950DT lists 144 GB against the H200’s 141GB, and 83% of its bandwidth, by our arithmetic.
DeepSeek, DeepGEMM-Ascend README, GitHub
“On Ascend platforms, users can simply install the package and use the same APIs and development workflow as DeepGEMM on other supported platforms.”
Three Implications
FOR NVIDIA INVESTORS The three STAR Market reports put Cambricon, Moore Threads and MetaX at 6.58 billion yuan of third-quarter revenue combined, by our arithmetic, up from 2.13 billion a year earlier.
FOR MODEL BUILDERS DeepSeek’s README says its Ascend kernels keep the same APIs as DeepGEMM, and Zhipu’s report says GLM-5 was adapted to seven domestic chip platforms.
FOR POLICYMAKERS The episode’s speakers argued that the shift is happening with or without remote-access rules; the bills themselves are tracked in our export-bills guide.
The framework opens: “Every technology stack has layers. Some are scarce, some are abundant.”
As we read it, the Chinese releases in this piece open the software layer around domestic chips, from GLM-5’s adaptation to DeepSeek’s Ascend kernels, which makes the accelerator underneath easier to swap. The framework does not discuss Huawei, DeepSeek or export controls.
What Is Not Established
The captions are machine-made, do not identify speakers and garble some names, so we attribute lines to the episode rather than to a person. Zhipu’s report describes adapting GLM-5 to run on the seven platforms; the excerpt we read concerns inference. DeepSeek’s README speaks of the DeepSeek model series rather than V4, and Meituan’s blog does not name the vendor of its ASICs.
The spec ratios compare figures each company chose to publish, with different definitions: Nvidia marks its FP8 figure with sparsity, and the two interconnect numbers may not measure the same thing. Huawei’s 950DT figures date from a September 2025 keynote. The prior-year revenue figures are our arithmetic from the growth rates the reports give. We did not contact the companies for this piece.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Three listed Chinese chipmakers tripled their combined quarterly revenue to 6.58 billion yuan, by our arithmetic, while Zhipu tuned GLM-5 for seven domestic platforms and DeepSeek published Ascend versions of its kernel libraries. On paper, Huawei’s 950DT has slightly more memory than the H200 and 83% of its memory bandwidth, by our arithmetic. The next data points are the STAR Market companies’ annual reports and the 950DT itself, which Huawei’s keynote said would be available in the fourth quarter of 2026.
A note on sourcing. We watched SemiAnalysis’s episode with ChinaTalk through YouTube’s automatic captions and description on 13 November 2026, and read, the same day, the GLM-5 technical report on arXiv, the READMEs of DeepSeek’s DeepGEMM-Ascend and DeepEP-Ascend repositories on GitHub, Meituan’s LongCat-2.0 blog and the README of its NPU serving branch, Huawei’s Huawei Connect 2025 keynote text, Nvidia’s H200 specifications page, and the third-quarter 2026 reports of Cambricon, Moore Threads and MetaX filed on the Shanghai Stock Exchange’s STAR Market via cninfo.
Revenue and profit figures are the companies’ reported, unaudited numbers; the prior-year revenues, combined totals, the 3.1 multiple and the spec ratios are our arithmetic. The app-layer revenue figure is the episode speaker’s and is not verified. We leave out personal material. We did not contact the companies for this piece.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
