DeepSeek has ordered more than 160,000 Huawei Ascend 950DT accelerators for a data center in Ulanqab, Inner Mongolia, planned at 1 gigawatt at full buildout, according to reports on 4 September citing Bloomberg.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source 2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source The site is for inference. Training stays on Nvidia hardware, because earlier attempts to train on Huawei chips failed on software and interconnect maturity.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source 3 The Arabian Post 2026-09-05 Inference primarily, training on Nvidia because of a performance gap; deliveries likely staged through 2027; memory and component supply as the constraint; broader Chinese pattern of domestic serving and imported training. Open source Huawei plans to ship the 950DT in the fourth quarter of 2026, memory shortages will hold output to the low hundreds of thousands this year, and partial operation is targeted for late 2027 or early 2028.2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source Our assessment, with high confidence, is that this is the largest single commitment by a Chinese frontier lab to domestic silicon for production serving; with moderate confidence, that the binding constraint is Huawei's high bandwidth memory supply rather than DeepSeek's money or Beijing's will; and with moderate confidence that the training and serving split described here is the template Chinese labs will follow for at least two years.
What the order is and is not
The Ascend 950DT carries 144 gigabytes of high bandwidth memory at 4.0 terabytes per second and sells for about 111,000 yuan, roughly 16,000 dollars.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source At that price 160,000 units is a chip bill in the region of 2.6 billion dollars before servers, networking or the building, a figure this outlet derives from the reported unit price rather than one any source states. DeepSeek raised more than 7 billion dollars in June 2026, which Forkast put at about 7.4 billion at a valuation of 52 to 59 billion dollars, so the order is affordable.2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source 4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source
The workload matters more than the count. Inference is where a lab's revenue is served and where latency and cost per token decide margins; it is also the workload that tolerates a weaker software stack, because a model is compiled once and run many times. Training is where the interconnect, the compiler and the debugging tools are stressed continuously, and where TFTC reports DeepSeek's Huawei experiments broke.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source The Arabian Post frames the split as a general Chinese pattern: domestic hardware for serving, imported chips retained for the demanding work.3 The Arabian Post 2026-09-05 Inference primarily, training on Nvidia because of a performance gap; deliveries likely staged through 2027; memory and component supply as the constraint; broader Chinese pattern of domestic serving and imported training. Open source That is the honest reading of the order. It is not a declaration of independence from Nvidia; it is a decision about which half of the business can afford to leave.
Why Ulanqab, and why the dates slip
Ulanqab sits about 350 kilometres northwest of Beijing, offers cheap wind and solar power and an average temperature of 4.3 degrees Celsius, and already hosts about 89 data centers under the East Data, West Compute initiative the National Development and Reform Commission launched in February 2022.2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source 4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source DeepSeek plans to self build while leasing additional capacity.4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source A gigawatt site draws power comparable to roughly 750,000 homes, which is why it sits where the wind is rather than where the users are.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source
The timeline is Huawei's, not DeepSeek's. The 950DT does not ship until the fourth quarter of 2026, high bandwidth memory shortages cap this year's output in the low hundreds of thousands across all customers, and filling this one order could take more than a year with deliveries staged through 2027.2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source 3 The Arabian Post 2026-09-05 Inference primarily, training on Nvidia because of a performance gap; deliveries likely staged through 2027; memory and component supply as the constraint; broader Chinese pattern of domestic serving and imported training. Open source TFTC reports DeepSeek is lobbying Beijing for allocation priority, which is what a buyer does when the seller cannot make enough for everyone.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source
Who gains and who loses
Huawei gains its anchor customer for the 950DT and a reference deployment at gigawatt scale. Its AI chip revenue is projected to rise from 7.5 billion dollars in 2025 to about 12 billion in 2026, and its share of the China AI chip market is projected near 50 percent.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source 2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source Beijing gains a showcase for East Data, West Compute and for the industrial policy behind domestic accelerators.4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source DeepSeek gains a serving stack that no export rule can switch off, ahead of an IPO filing Forkast reports is targeted for late 2026.4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source
Nvidia loses a customer's serving business it had in practice already lost. TFTC puts Nvidia's China AI chip share at about 8 percent in 2026 from about 40 percent in 2025; tech-ish quotes Jensen Huang describing a move from 95 percent to zero.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source 2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source The two figures are not reconcilable and neither is independently sourced, so the direction is well established and the magnitude is not. The other loser is any Chinese cloud that hoped to rent DeepSeek inference capacity: a lab that owns a gigawatt is a lab that no longer needs to lease one.
The counter case
The thesis that this is a durable template could fail in either direction. If Huawei's memory constraint persists, the 160,000 chips arrive late and in tranches and DeepSeek keeps serving on whatever Nvidia and cloud capacity it can find, in which case the order is a statement of intent for the IPO prospectus more than a production plan.2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source 3 The Arabian Post 2026-09-05 Inference primarily, training on Nvidia because of a performance gap; deliveries likely staged through 2027; memory and component supply as the constraint; broader Chinese pattern of domestic serving and imported training. Open source If, on the other hand, US export policy loosens further, as the January H200 rule already did for some parts, Chinese labs may find imported inference chips cheaper per token than domestic ones and the split reverses. There is also the unresolved question of what DeepSeek trains on. US officials have alleged the company used banned Blackwell processors for its latest model, which DeepSeek and Nvidia dispute; if enforcement tightens on training hardware, the inference plan becomes the easy half of a much harder problem.4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source
What to watch
- 950DT volume shipments. Public evidence of Huawei shipping 950DT in volume in the fourth quarter of 2026 would keep the plan on schedule; a slip into 2027 would push partial operation toward 2028.2 tech-ish 2026-09-04 Ulanqab 350 km northwest of Beijing; wind and solar power and a 4.3 degree average temperature; partial operation late 2027 or early 2028; 950DT launch in the fourth quarter of 2026; memory shortages cap output in the low hundreds of thousands; Huang on a share move from 95 percent to 0; DeepSeek raised more than 7 billion dollars in June. Open source
- DeepSeek's IPO filing. A prospectus filed by early 2027 should disclose committed capital spending and supplier concentration; the absence of Huawei as a named supplier would undercut the reports.4 Forkast via Yahoo Finance 2026-07-31 About 89 data centers in Ulanqab; East Data, West Compute since February 2022; self build plus leased capacity; about 7.4 billion dollars raised in June at 52 to 59 billion valuation; revenue near 500 million dollars annualized; IPO filing targeted for late 2026; disputed Blackwell training allegations. Open source
- A training run on Ascend. Any credible report of a DeepSeek frontier model trained primarily on Huawei hardware within two years would collapse the split this brief describes; continued Nvidia training would confirm it.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source
- Allocation politics. If other Chinese labs announce comparable Ascend orders in the next six months, Huawei's memory supply becomes the contested resource and DeepSeek's delivery schedule is the first casualty.1 TFTC 2026-09-04 More than 160,000 Ascend 950DT; 1 GW at full buildout; inference only, training stays on Nvidia; 144 GB HBM at 4.0 TB/s; about 16,000 dollars per chip; more than a year to fulfil; Nvidia China share about 40 percent to about 8 percent; Huawei chip revenue 7.5 to 12 billion dollars; about 750,000 homes of power. Open source
DeepSeek has decided where its tokens will be served. Where its models are trained remains a question it has not answered, and the order does not answer it either.