> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # 10,000 of 30,000 Cards: Who Gets Paid While China's First Ascend 950DT Pool Waits on a Proof Run - URL: https://www.sotkn.com/markets/china/china-ascend-950dt-10000-card-pool-october-2026/ - Published: 2026-10-21T08:08:18.000Z - Updated: 2026-10-21T08:08:18.000Z - Description: The first 10,000-card pool of Huawei's new Ascend 950DT chips is due by the end of October in a hall built for 30,000. The building, the power and the cooling are done. What nobody has shown yet is a weeks-long training run on it. - Author: Sean - Tags: #market-china, Markets *A hall on the grasslands outside Hohhot is about to tell us how far China's own AI chips have come. China Telecom and Huawei say the first 10,000-card pool of Huawei's new Ascend 950DT chips is in place there, in a building made for 30,000, with handover due by the end of October. The building, the power and the cooling are finished. What has not been shown in public is a training run that keeps 10,000 of these chips busy for weeks. The stuck step is the proof run.* **CHINA PULSE** · SEMICONDUCTORS · AI DATA CENTERS · Data as at 21 Oct 2026 · Next update: the Hohhot handover due by 31 Oct **Affects:** [Semiconductors](https://www.sotkn.com/stack/semiconductors/) · [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/) · [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/) — Markets: [China](https://www.sotkn.com/markets/china/) · [Taiwan](https://www.sotkn.com/markets/taiwan/) · [US](https://www.sotkn.com/markets/us/) **In short:** In mid-October China Telecom and Huawei said the world's first Ascend 950DT 10,000-card pool had landed at China Telecom's Inner Mongolia cloud park, in a hall able to take 30,000 cards, built on Huawei's 1,024-card supernode design and due for handover by the end of October. China Mobile has a 1,024-card 950DT supernode running in Beijing, and the first full 1,024-card Atlas 950 SuperPoD was switched on on 15 October. On 19 October The Information reported that DeepSeek plans to start pretraining on the 950DT as early as the end of 2026, a job that may need more than 10,000 chips for weeks, with the start depending on tests. Huawei is paid first. Stuck step: the proof run. Map of who gets paid — not a buy list. Education only, not advice to buy or sell any security. Figures are as stated by the companies, the local government or the publications named. Our own sums are labelled. ### The numbers that matter this week Cards the Hohhot hall can take 30,000 China Telecom B12 hall Built First Ascend 950DT pool 10,000 Handover due by end-October Clock Chips a DeepSeek run may need 10,000+ For several weeks · start set by tests Stuck step Cards in one supernode 1,024 1 EFLOPS FP8 · fully liquid-cooled Lit 15 Oct Memory per 950DT 144 GB 4 TB/s · Huawei's own HiZQ 2.0 Spec Hohhot building efficiency PUE < 1.15 Green power above 85% Built Figures from People's Daily Online 14 October and the Helingeer New Area 15 October 2026 (Hohhot), EET China 16 October (supernode), The Information 19 October via TokenPost and Techmeme (DeepSeek), and Huawei's 18 September 2025 keynote (chip specs). Status chips are Second Order’s read. **On this page:** [Why it matters](#why-it-matters) · [What changed](#what-changed) · [The hall](#the-hall) · [The chip](#the-chip) · [The supernode](#the-supernode) · [The customer](#the-customer) · [Our math](#our-math) · [Who gets paid](#money-goes) · [What is stuck?](#still-stuck) · [Dates next](#dates) · [What would change this](#change-view) · [FAQ](#faq) ## Why does one hall in Inner Mongolia matter to who gets paid in AI? Because China's AI build-out now runs on two questions at once. The first is whether there are enough chips. The second, which gets less attention, is whether those chips can be wired together into one machine big enough to train a leading model. A single chip is only useful for training if it can work with thousands of others for weeks without the job falling over. That is a test of the chip, the links between chips, the software and the building, all at once. Until now, most public proof for Huawei's Ascend chips came from smaller systems or from serving models that were trained elsewhere. A 10,000-card pool of the newest Ascend part, in a hall that can hold three times that number, is the first place where the bigger question can be answered in China with Huawei's newest chip. If it works, every operator that bought into the same design has a template. If it stumbles, the money waits. That decides the order of the money. Huawei is paid when it sells the cards and the supernode cabinets. The builders and the power and cooling suppliers have already been paid for a finished hall. The telecom operators are paid as they rent the capacity out. The model makers are paid last, once a model trained on the pool is in use. Each later payment depends on the run in between. We have followed China's AI build from other angles. Yesterday's [Hygon Q3 Pulse](https://www.sotkn.com/markets/china/china-hygon-q3-hbm-october-2026/) looked at memory as the bind for China's chip designers. Our [SMIC utilisation Pulse](https://www.sotkn.com/markets/china/china-smic-utilization-october-2026/) showed the domestic foundry running nearly full. Our [Alibaba 20 GW Pulse](https://www.sotkn.com/markets/china/china-alibaba-20gw-october-2026/) followed a cloud that has the plan but waits on chips. This piece looks at the next step along: what happens once the chips arrive in one place. 1 · The hall **Who:** China Telecom, Inner Mongolia park **Paid:** Built for 30,000 cards; power and cooling done Paid first 2 · The cards **Who:** Huawei Ascend 950DT and supernodes **Paid:** 10,000-card pool in place; handover by end-October Paid on delivery 3 · The proof run binds **Who:** A weeks-long 10,000-card training job **Paid:** DeepSeek's start depends on testing Stuck step 4 · The models **Who:** DeepSeek, StepFun and cloud users **Paid:** Paid once trained models ship Paid last ## What changed between 8 and 20 October? Huawei promised the 950DT for the fourth quarter of 2026 a year ago. In the past two weeks the promise turned into machines in named buildings, and a named customer came into view. - **9 October — Beijing.** People's Posts and Telecommunications News reported that China Mobile's Beijing branch and its cloud arm have a 1,024-card Ascend 950DT supernode in place at the Changping information-port data center. China Mobile said its cloud already supports nine domestic AI chip families and hosts more than 300 models on its MoMA platform. - **14 and 15 October — Hohhot.** People's Daily Online and the Helingeer New Area announced the first Ascend 950DT 10,000-card pool at China Telecom's Inner Mongolia cloud park, in a hall that can take 30,000 cards. The halls, the power plant rooms and the cooling were reported complete, the system in joint testing, and the core build due by the end of October. Huawei's first 3D data center in northern China topped out on the same campus. - **13 to 16 October — the software.** Huawei's Ascend CANN developer community ran a four-day livestream series on training and serving DeepSeek's V4.1 models on the Ascend 950, ending on 16 October with a session on pretraining trillion-parameter mixture-of-experts models on 950 supernodes. - **15 and 16 October — the first full supernode.** EET China reported that the first 1,024-card Atlas 950 SuperPoD, fully liquid-cooled, had been switched on and was serving StepFun's newly released Step 5 Preview model from day one. - **19 and 20 October — the customer.** The Information reported that DeepSeek plans to start large-scale pretraining on the 950DT as early as the end of this year, that a full run may need more than 10,000 chips for several weeks, and that the start depends on testing. It also reported that Huawei engineers are working from DeepSeek's offices. Put together, the hardware arrived on time and the buyer for a frontier-size job was named in the same week. What links the two is a run nobody has reported finishing yet. ## What is in the Hohhot hall? ![Chart of the Hohhot pool: a hall for 30,000 cards, a 10,000-card pool due end-October, an 8,192-chip maximum SuperPoD and a 1,024-card supernode](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/china-ascend-950dt-10000-card-pool-october-2026-print.png) A third of the hall is the first pool. · Education only. The site is China Telecom's cloud computing park in the Helingeer New Area, part of the Inner Mongolia free trade zone near Hohhot. Inner Mongolia is one of the western hubs in China's East Data, West Computing plan, which pushes heavy computing west, toward regions with spare land and green power. The pool sits in a building the reports call the B12 intelligent computing center, which can take a cluster of 30,000 cards. The reports are clear that the shell is finished. People's Daily Online said the core computing halls, the power center, the power distribution and the cooling systems were all built, that the project had moved into joint testing, and that the core work was due to finish and be handed over by the end of October. Next to it, Huawei's first 3D intelligent computing data center in northern China has topped out. It uses stacked liquid-cooled cabinets and liquid cooling throughout, with a power usage effectiveness below 1.15 and more than 85% of its electricity from green sources, according to the same report. A PUE of 1.15 means about 15% extra power goes on cooling and other overheads for every unit used by the chips (our arithmetic from the definition). This matters for the order of the money. In the US and Australia, our recent Pulses found the grid connection was the step that held up the hall. In Hohhot, the reports say the grid, the building and the cooling are done. The question has moved inside the building. What the reports say is done Computing hallsBuilt Power center and distributionBuilt CoolingBuilt Huawei 3D buildingTopped out What is still open Joint testingUnder way in mid-October HandoverDue by end-October A 10,000-card training runNot reported Cards beyond 10,000Room for 30,000; no date ## What is the Ascend 950DT? It is the training version of Huawei's newest AI chip. At Huawei Connect in September 2025, rotating chairman Eric Xu set out a roadmap with two versions of the Ascend 950 sharing one chip die. The 950PR, for the first stage of answering a prompt and for recommendation systems, came first, in the first quarter of 2026\. The 950DT, for the second stage of answering and for training, was promised for the fourth quarter of 2026\. The October launches mean it arrived inside that window. The difference between the two is memory. Each 950DT carries Huawei's own high-bandwidth memory, which it calls HiZQ 2.0: 144 GB per chip, read at 4 TB per second. Each chip can talk to others at 2 TB per second. Huawei says the 950 series does 1 petaflop at FP8 and 2 petaflops at FP4, the low-precision number formats used to speed up AI work. Memory is where China's chip makers have been squeezed hardest. Our [September HBM access Pulse](https://www.sotkn.com/markets/china/china-hbm-access-september-2026/) and yesterday's [Hygon Pulse](https://www.sotkn.com/markets/china/china-hygon-q3-hbm-october-2026/) both found stacked memory to be the bind for domestic designers. Huawei's answer is to package its own. The October reports do not say how many HiZQ stacks Huawei can make, so we cannot tell from them whether memory limits how fast the hall fills from 10,000 cards toward 30,000. Huawei has also been blunt about where it makes its money. In the same keynote, Xu said the company's AI monetisation strategy is focused on hardware, and that it would open up its CANN software. That is why the cards and the cabinets are the first big payment in this chain: the software is given away to make the hardware sell. Huawei does not name the factory that makes the 950 die in any of the October releases. We leave that open. ## Why does the supernode size matter? ![Chart of Huawei supernode sizes: 384 chips in the Atlas 900 A3, 1,024 cards in the first Atlas 950 supernodes lit in October, 8,192 at the Atlas 950 maximum](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/china-ascend-950dt-10000-card-pool-october-2026-shift.png) Supernodes got bigger fast. · Education only. Because Huawei's whole design bets on linking many slower chips into one big machine. Xu said plainly in 2025 that mainland China will lag in chipmaking process for a long time, so Huawei would make up the gap by combining chips. A supernode is a group of chips joined by Huawei's UnifiedBus links so tightly that the software can treat them as one computer. The sizes have grown quickly. The Atlas 900 A3 SuperPoD of 2025 joined 384 of the older Ascend 910C chips; Huawei said more than 300 had been deployed by September 2025\. The first full Atlas 950 SuperPoD switched on in October joins 1,024 cards in liquid-cooled compute and interconnect cabinets, for 1 exaflop at FP8 and 2 at FP4, with 256 TB of shared memory and an end-to-end delay of 3 microseconds, EET China reported. Huawei's plan for the Atlas 950 SuperPoD at full size is 8,192 chips in 160 cabinets, about 1,000 square metres of floor. A 10,000-card pool built on 1,024-card supernodes is about ten of them working together (our arithmetic). That means the pool has two kinds of links: very fast ones inside each supernode and slower ones between them. A big training run has to be laid out so the heaviest traffic stays inside a supernode. Huawei's developer sessions in mid-October were about exactly that. The size also explains the hall. At Huawei's own full-size figure of 160 cabinets for 8,192 chips, 30,000 cards would take roughly 590 cabinets (our arithmetic, a rough scale only, since the Hohhot pool uses 1,024-card nodes). That is why the building, its power and its liquid cooling had to be designed for the full pool before the first card arrived. ## Who will use 10,000 cards at once? DeepSeek is the name now on the table. According to The Information, as summarised by TokenPost and Techmeme, the Hangzhou lab plans to start large-scale pretraining on the 950DT as early as the end of 2026\. A full run may need more than 10,000 chips running without a break for several weeks. Whether it starts before the year ends depends on testing. The same report says Huawei engineers are working from DeepSeek's offices, folding the lab's training needs into the design of future chips, and that DeepSeek is to receive most of the supply of Huawei's next chip, the Ascend 960DT, next year. Caixin reported in September that Huawei had pulled the 960DT forward to the third quarter of 2027, one quarter early. TokenPost also notes DeepSeek is still seeking more Nvidia chips. The lab is hedging, not switching. DeepSeek is not the only user lining up. StepFun's Step 5 Preview was running on the first 1,024-card supernode from day one. China Mobile says its MoMA platform will tune its 300-plus hosted models for the 950DT, and that its domestic chips and in-house inference engine have already cut unit costs by about 30%. The reports do not say whether DeepSeek's run would use the Hohhot pool, another China Telecom site, or Huawei's own cloud. We do not guess. What they do say is that the first place in China with 10,000 950DT cards under one roof is due for handover this month, and that the first buyer for a job that size wants to start by year-end. ## Our math: how big is the first pool? A few simple sums put the pool in proportion. All of them are our arithmetic from the figures above. - **Hall versus pool:** 10,000 of 30,000 cards is one third of the hall's capacity. Two thirds of the room has no dated cards yet. - **Pool versus supernodes:** 10,000 cards on 1,024-card supernodes is about ten supernodes linked together. - **Pool versus shipments:** Shanghai Metals Market estimates Huawei shipped about 572,000 Ascend chips in the first half of 2026\. On that estimate, the whole Hohhot pool is under 2% of half a year's output. Chips in total are not the limit. Having enough of the newest chip in one place, working as one machine, is. - **Pool versus plan:** Huawei's Atlas 950 SuperCluster plan, for the same quarter, joins more than 520,000 950DT chips. The Hohhot pool is about 2% of that plan. - **Memory on the floor:** at 144 GB a chip, 10,000 cards carry about 1.44 petabytes of Huawei's own stacked memory. The point of the sums: the pool is small next to Huawei's plans and its shipments, but it is the right size for one frontier training run. That is why it is the test. ## Who gets paid, and in what order? ![Infographic: who gets paid as China's first Ascend 950DT 10,000-card pool comes online, and why the proof run is the stuck step](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/china-ascend-950dt-10000-card-pool-october-2026-infographic.png) Who gets paid / stuck step in Hohhot. · Education only. - **First — the builders, power and cooling.** The halls, power center and liquid cooling are reported complete. Those contracts are earned whatever happens next. The reports put the green share of the new building's power above 85%. - **Second — Huawei.** Huawei is paid for the cards and the supernode cabinets as they are delivered. It has said hardware is where it makes its AI money. Each 1,024-card supernode is a large single order. - **Third — the operators.** China Telecom in Hohhot and China Mobile in Beijing are paid as they rent the capacity to model makers and companies. Their income depends on the pools being busy, which depends on them working well. - **Last — the model makers.** DeepSeek, StepFun and others earn on the models once trained and in use. For a pretraining run, that is months after the run starts. Compare it with Taiwan, where this morning's [TSMC CoWoS Pulse](https://www.sotkn.com/markets/taiwan/taiwan-tsmc-q3-cowos-october-2026/) found packaging capacity still short. In Taiwan the chips wait on a factory step. In Hohhot the chips have arrived and wait on proof that they can work together at scale. Different bind, same rule: the step that is short decides who is paid first. One-line rule: in Hohhot right now, the builders are paid for a finished hall, Huawei is paid on delivery, the operators are paid as the pool is used, and the model makers wait for a model. The stuck step is the proof run. ## So what is the stuck step? The proof run: a training job that keeps 10,000 950DT cards working together for weeks. Not the building: it is finished. Not the power or the cooling: both are reported complete. Not chips in general: Huawei ships them in the hundreds of thousands. Three things make this step stick: - **It has never been done in public on this chip:** the first full 1,024-card supernode was switched on only on 15 October, and no report describes a 10,000-card training run on the 950DT yet. - **The links between supernodes:** a 10,000-card job spans about ten supernodes, and the slower links between them are where big runs tend to stall. - **The customer's own test:** DeepSeek's start date depends on testing, by its own account as reported. If testing slips, so does the run, and so does the payback on the pool. So-what: whoever has already delivered, the builders and Huawei, is paid before whoever needs the run to work. A 30,000-card hall is worth what its first 10,000 cards can prove. For the general pattern of cash sticking where the bind sits, see [where the stuck step moves](https://www.sotkn.com/analysis/where-does-the-stuck-step-move/). ## Which dates test this read next? ![Calendar of dates that test the Ascend read, from the end-October handover in Hohhot to the Ascend 960DT in the third quarter of 2027](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/china-ascend-950dt-10000-card-pool-october-2026-calendar.png) Dates that test the read. · Education only. - By 31 October 2026 — handover of the Hohhot pool. The test: whether China Telecom confirms it is in use, and by whom. - Fourth quarter 2026 — Huawei's own date for the Atlas 950 SuperPoD and SuperCluster. The test: whether larger 950DT systems ship beyond the first two operators. - End of 2026 — DeepSeek's earliest pretraining start, as reported. The test: whether the run starts, and on how many cards. - 2027 — most Ascend 960DT supply reportedly goes to DeepSeek. The test: whether the 950DT run went well enough to justify it. - Third quarter 2027 — the Ascend 960DT and a liquid-cooled 960 supernode, per Caixin. The test: whether the next chip keeps the pace. - Fourth quarter 2028 — the Ascend 970\. The test: the long clock of Huawei's roadmap. The near dates test the proof run. The 2027 dates test whether the proof earns the next round of orders. ## What would change this view? - A named 10,000-card training job finishes on the 950DT — the proof-run bind would loosen and the stuck step would move to how fast the hall fills toward 30,000 cards. - Huawei or China Telecom date the next 20,000 cards for Hohhot — the bind would move to chip and memory supply. - DeepSeek's run slips well into 2027, or moves back to Nvidia chips — the bind would stay on proof and the operators' payback would wait. - New export controls or a power shortfall in Inner Mongolia — the bind would move outside the building again. ## Practical takeaways - First Ascend 950DT 10,000-card pool at China Telecom's Inner Mongolia park; hall built for 30,000 cards - Halls, power and liquid cooling reported complete; handover due by the end of October 2026 - First 1,024-card Atlas 950 SuperPoD switched on 15 October; China Mobile runs a 1,024-card supernode in Beijing - DeepSeek reportedly plans 950DT pretraining from as early as end-2026, needing 10,000+ chips for weeks - Huawei is paid on delivery; operators on use; model makers last - Stuck step: the proof run ## FAQ **What launched in Inner Mongolia?** China Telecom and Huawei say the world's first 10,000-card pool of Huawei's new Ascend 950DT AI chips has landed at China Telecom's cloud park in the Helingeer New Area outside Hohhot. It sits in a hall called B12 that can take 30,000 cards, uses Huawei's 1,024-card supernode design, and was in joint testing in mid-October with handover due by the end of the month. **Why does a 10,000-card pool matter?** Because 10,000 chips is about the size a frontier model's training run needs. The Information reported that DeepSeek plans to start pretraining on the 950DT as early as the end of 2026 and that a full run may need more than 10,000 chips running for weeks. A pool of that size is where such a run can be proven. **Who gets paid first?** Huawei, which sells the cards and supernodes; Huawei has said its AI money comes from hardware. Then the builders and the power and cooling behind the hall, which are already finished. China Telecom and China Mobile earn as they rent the capacity out. Model makers such as DeepSeek earn last, once a model trained on the pool ships. **What is the stuck step?** The proof run. The hall, the power and the cooling are built and the chips are arriving. What has not been shown in public is a 10,000-card training job running stably on 950DT chips for weeks. DeepSeek's start date depends on that testing. **Is this investment advice?** No. Education only. It maps who gets paid and which step is stuck. It is not a recommendation to buy or sell any security, and we leave out share prices, broker ratings and price targets on purpose. ## Keep reading on the map - [China Markets hub](https://www.sotkn.com/markets/china/) — country page - [Hygon Q3 and HBM Pulse](https://www.sotkn.com/markets/china/china-hygon-q3-hbm-october-2026/) · [China HBM access Pulse](https://www.sotkn.com/markets/china/china-hbm-access-september-2026/) - [SMIC utilisation Pulse](https://www.sotkn.com/markets/china/china-smic-utilization-october-2026/) · [SMIC margin guide Pulse](https://www.sotkn.com/markets/china/china-smic-margin-logicfolding-october-2026/) - [Alibaba 20 GW Pulse](https://www.sotkn.com/markets/china/china-alibaba-20gw-october-2026/) · [China 8-inch power foundry Pulse](https://www.sotkn.com/markets/china/china-8inch-power-foundry-october-2026/) - [China power–computing coordination Pulse](https://www.sotkn.com/markets/china/china-power-computing-september-2026/) · [China data-services power Pulse](https://www.sotkn.com/markets/china/china-data-services-power-august-2026/) - [TSMC CoWoS Pulse](https://www.sotkn.com/markets/taiwan/taiwan-tsmc-q3-cowos-october-2026/) — Taiwan's bind this week - [Tenstorrent AI servers skip HBM Wire](https://www.sotkn.com/wire/tenstorrent-1-5b-6-2b-ai-servers-skip-hbm-wire/) · [How long an AI data center takes to build](https://www.sotkn.com/explained/how-long-does-it-take-to-build-an-ai-data-center/) - [Semiconductors on the stack](https://www.sotkn.com/stack/semiconductors/) · [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/) - [All Markets doors](https://www.sotkn.com/markets/) — Taiwan, Korea, US, China, Singapore, Australia, Malaysia **Get the Wire.** Second Order’s free daily brief on what moved in the AI chip map. [Read the Wire →](https://www.sotkn.com/the-wire/) · [Subscribe free →](https://www.sotkn.com/#/portal/signup) **Members’ Analysis (teaser):** When the stuck step moves, where does cash stick? The free line: whoever clears the bind is paid before whoever waits on it. [Read the Analysis →](https://www.sotkn.com/analysis/where-does-the-stuck-step-move/) **How we check this.** The Hohhot pool, the 30,000-card hall, the 1,024-card supernode design, the liquid-cooled building, its PUE and green-power share, the testing stage and the end-October handover come from People’s Daily Online on 14 October and the Helingeer New Area’s own notice on 15 October. The China Mobile Beijing supernode and its per-card figures come from People’s Posts and Telecommunications News on 9 October. The first 1,024-card Atlas 950 SuperPoD switch-on comes from EET China on 16 October. DeepSeek’s pretraining plan comes from The Information on 19 October, as summarised by Techmeme and TokenPost; we have not seen the original. Huawei’s chip and SuperPoD roadmap comes from Eric Xu’s keynote of 18 September 2025 and Caixin on 18 September 2026\. Shipment numbers are Shanghai Metals Market estimates, labelled as such. Huawei does not name the factory that makes the 950 chip in these releases, and we do not guess it. Share prices, broker ratings and price targets are left out on purpose. Anything we worked out ourselves is labelled “our arithmetic”. The references we used are listed at the end. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) Not investment advice. Do your own research. --- ### References 1. People's Daily Online (Inner Mongolia), via Tencent News: world's first Ascend 950DT 10,000-card cluster lands in Hohhot, 14 Oct 2026\. [news.qq.com](https://news.qq.com/rain/a/20261014A08XF900?ref=sotkn.com) — China Telecom and Huawei pool at China Telecom's Inner Mongolia cloud park; B12 hall able to take 30,000 cards; 1,024-card supernode; Huawei 3D liquid-cooled data center topped out with PUE below 1.15 and green power above 85%; halls, power and cooling built; joint testing; core build due by end-October. 2. Helingeer New Area investment service bureau, Ascend 950DT 10,000-card cluster lands, 15 Oct 2026\. [hlgena.huhhot.gov.cn](http://hlgena.huhhot.gov.cn/zxzx/xqdt/202610/t20261015%5F2043074.html?ref=sotkn.com) — first Ascend 950DT 10,000-card pool at China Telecom's Inner Mongolia park; 30,000-card deployment capacity; FP4 inference, UnifiedBus 2.0, 1,024-card supernode. 3. People's Posts and Telecommunications News, via China Association of Communication Enterprises: China Mobile domestic-stack computing center lands in Beijing, 9 Oct 2026\. [cace.org.cn](https://www.cace.org.cn/News/Ncontent?key=f2017e1d64821e360324c2480fe5cbdf&ref=sotkn.com) — Ascend 950DT 1,024-card supernode at China Mobile Cloud's Changping node; 4 TB/s memory bandwidth and 2 TB/s interconnect per card; 1 EFLOPS FP8; nine domestic AI chip families supported; MoMA platform with 300-plus models, unit cost down about 30%. 4. EET China (Hardware World), first 1,024-card Ascend 950 liquid-cooled supernode powered on, 16 Oct 2026\. [eet-china.com](https://www.eet-china.com/mp/a531734.html?ref=sotkn.com) — first Atlas 950 SuperPoD with 1,024 NPUs, fully liquid-cooled compute and interconnect cabinets; 1 EFLOPS FP8, 2 EFLOPS FP4; 256 TB unified memory; 3 microsecond latency; day-zero support for StepFun Step 5 Preview. 5. TokenPost, DeepSeek targets year-end pretraining run on Huawei Ascend 950DT chips, 20 Oct 2026\. [tokenpost.com](https://www.tokenpost.com/news/technology/35470?ref=sotkn.com) — DeepSeek plans large-scale pretraining on the 950DT as early as end-2026; a full run may need more than 10,000 chips for several weeks; start depends on testing; Huawei engineers at DeepSeek's offices; DeepSeek also seeking more Nvidia chips. 6. Techmeme summary of The Information, DeepSeek and Huawei collaborate on future chip designs, 19 Oct 2026\. [techmeme.com](https://www.techmeme.com/261019/p35?ref=sotkn.com) — DeepSeek to use Ascend 950DT for pretraining as soon as the end of 2026 and to receive most of the Ascend 960DT supply next year, per The Information. 7. Huawei, keynote by Eric Xu at Huawei Connect 2025, 18 Sep 2025\. [huawei.com](https://www.huawei.com/en/news/2025/9/hc-xu-keynote-speech?ref=sotkn.com) — Ascend 950DT due Q4 2026 with HiZQ 2.0 memory, 144 GB and 4 TB/s; Atlas 950 SuperPoD up to 8,192 chips in 160 cabinets on about 1,000 square metres, Q4 2026; Atlas 950 SuperCluster of more than 520,000 chips; AI monetisation focused on hardware; more than 300 Atlas 900 A3 SuperPoDs deployed. 8. Caixin Global, Huawei speeds up AI chip roadmap, unveils optical super node, 18 Sep 2026\. [caixinglobal.com](https://www.caixinglobal.com/2026-09-18/huawei-speeds-up-ai-chip-roadmap-unveils-optical-super-node-102486033.html?ref=sotkn.com) — Ascend 960DT now due Q3 2027, one quarter earlier than planned; liquid-cooled Ascend 960 super node Q3 2027; Ascend 970 Q4 2028. 9. Shanghai Metals Market (SMM), Huawei Ascend 950DT 10,000-card pool and 1,024-card liquid-cooled super node go live, 21 Oct 2026\. [news.metal.com](https://news.metal.com/newscontent/104169283-smm-flash-huawei-ascend-950dt-10000-card-pool-+-1024-card-liquid-cooled-super-node-goes-live?ref=sotkn.com) — SMM tally of the October launches; SMM estimate of 572,000 Ascend shipments in the first half and a 28% share; cluster scale from 384 to 8,192 cards. 10. Ascend CANN community, meetup and livestream list, accessed 21 Oct 2026\. [gitcode.com](https://gitcode.com/cann/community/blob/master/events/meetup/README.md?ref=sotkn.com) — 13 to 16 October series on DeepSeek V4.1 training and inference on Ascend 950, including a 16 October session on trillion-parameter MoE pretraining on 950 supernodes.