In September 2026, Bloomberg and specialist outlets such as TrendForce and The Decoder reported that DeepSeek plans one of the largest known Huawei Ascend silicon deployments: at least 160,000 Ascend 950DT processors at a gigawatt-scale datacenter in Ulanqab, Inner Mongolia. Founder Liang Wenfeng is cast as driving early adaptation to domestic chips.
The supply calendar places new deliveries between Q4 2026 and Q1 2027. Huawei, per the same sources, has accelerated the training-oriented Ascend 960DT toward early Q1 2027 to absorb demand. DeepSeek aims to use the silicon not only for inference—as earlier readings of the 950DT suggested—but also for training.
The model target is ambitious and should be read with calendar caution: train toward a system of about 2 trillion parameters and, in a later phase, aim at 8 trillion. Neural Pulse does not treat those figures as milestones already hit, but as capacity goals that justify the chip order under U.S. export controls on cutting-edge Nvidia GPUs.
The correct narrative is not “DeepSeek abandons Nvidia tomorrow.” It is a gradual migration: a dual track in which U.S. silicon remains critical for sensitive training workloads while Ascend capacity is built at cluster scale. Export controls make dual-track a strategic necessity, not an autonomy slogan.
For China’s AI ecosystem, a 160k Ascend order in a single DC signals supply maturity and operational confidence in the Huawei stack—software, networking, and scheduling included. Software-stack gaps have historically been the Achilles heel versus CUDA; a deployment this size implies DeepSeek believes enough of that gap is closed to justify capital.
Geopolitical implications: every megawatt and every Ascend rack reduces the elasticity of sanctions as a sole tool. At the same time, it does not erase residual dependence on tools, IP, and global supply-chain nodes. Regulators in Washington and Brussels will keep measuring not only chip counts but effective training performance per dollar and per watt.
From the vantage of open-model buyers, more Ascend capacity at DeepSeek can translate into faster V4.x family iterations and price pressure on closed labs. From a security vantage, it concentrates frontier compute in an actor already central to open weights.
Neural Pulse links this announcement to DSec: ordering chips is not enough; the sandbox and RL platform that turns FLOPs into useful agents is required. DeepSeek is publishing both faces—domestic compute and training plumbing—in the same September window.
Risks to watch: Q4/Q1 delivery slips, real 950DT performance versus H100/B200 on MoE workloads, and possible network/interconnect bottlenecks in a 160k cluster. An announced order is not an operational cluster; the proof will be the first stable training run at scale.
Liang Wenfeng has insisted, per coverage, on adapting to Huawei early. That urgency fits uncertainty over new control rounds and China’s internal race not to depend on a single foreign accelerator vendor.
For Nvidia and U.S. allies, the message is that China’s compute ceiling does not freeze—it reconfigures. For Huawei, DeepSeek becomes an anchor customer and showcase for Ascend in training, not only cloud inference.
Neural Pulse closes without triumphalism: 160,000 Ascend 950DT chips would, if delivered on schedule, be an infrastructure milestone. Until then, the responsible reading is industrial planning under sanctions—with Nvidia still in the equation—and a 2T→8T model horizon that must be verified run by run, not by headline.
Investors and regulators in 2026–2027 should track three indicators: real rack-in dates, DeepSeek-published training benchmarks on Ascend, and whether the 960DT arrives in Q1 2027 with software ready for MoE at scale.
