Key Takeaways

  • DeepSeek ordered 160,000 of Huawei's flagship 950DT AI chips for a new data center cluster in Inner Mongolia.
  • DeepSeek is deploying these 950DT chips for inference rather than model training, despite Huawei designing them specifically for training workloads.
  • Huawei's total manufacturing run for 950DT chips remains capped in the low hundreds of thousands, with company executives projecting they cannot meet domestic Chinese demand until 2030.
  • When Huawei submitted bids for sovereign AI projects in Egypt and Malaysia, Washington assembled a consortium of American companies to launch an emergency counter-bid.

DeepSeek Repurposes 160,000 Huawei Chips

Nvidia executives have spent months warning Washington that export curbs carry unintended fallout. Their core position is simple: strict export restrictions do not halt foreign AI progress; they accelerate it by forcing international buyers straight to Chinese hardware suppliers. As Bloomberg reporter Mackenzie Hawkins reported, Nvidia views Huawei as a genuine rival: “They sort of view Huawei, their primary competitor in China, as this sort of existential threat to US dominance.” (play from 27:55)

The reality on the ground inside China confirms that shift is already happening at scale. DeepSeek, the Chinese AI research lab, placed an order for 160,000 of Huawei's top-tier 950DT chips. The hardware is earmarked for a massive data center DeepSeek is constructing in Inner Mongolia.

The surprising detail is how DeepSeek plans to use them. Huawei built the 950DT as a high-performance training processor. Yet DeepSeek chose to reassign all 160,000 units to run inference. That pivot reveals two operational realities. First, Chinese frontier labs still face major software and networking bottlenecks when attempting to coordinate tens of thousands of non-Nvidia processors for frontier model pre-training. Second, running inference on domestic silicon frees up their existing, restricted inventory of Western accelerators strictly for model development.

Huawei still faces harsh manufacturing ceilings. Hawkins noted that “Huawei's overall production capacity of that hardware is in the low hundreds of thousands.” (play from 47:52) Supplying DeepSeek alone absorbs a massive slice of Huawei's available stock. Senior leadership at Huawei is open about that bottleneck. As Hawkins pointed out, “Huawei's senior executives themselves say it'll take until 2030” (play from 44:56) to meet domestic demand.

The Proxy Fight for Sovereign Data Centers

While Huawei struggles to produce enough silicon for mainland customers, it is aggressively competing abroad. The company is actively pursuing contracts to build sovereign AI infrastructure across the Global South, directly targeting developing economies that want local computing hubs.

Hawkins pointed to direct clashes in markets like Egypt and Malaysia. Huawei submitted formal proposals to build out sovereign infrastructure for national clients. American officials noticed the move and intervened immediately. Hawkins noted that “when they learned of this bid that Huawei had made in Egypt, they like gathered this consortium of American companies and they were like counter bid.” (play from 49:03)

The clash shows that hardware sanctions cannot be separated from international tech diplomacy. Washington wants American cloud giants and chip designers to control foreign data centers. But if US export rules make access to American silicon too bureaucratic or uncertain, foreign governments start listening to Beijing. Nvidia argues that pushing too hard on global export restrictions risks handing overseas market share directly to Chinese competitors.

What to Do With This

If you run an AI startup with heavy inference costs, audit your hardware roadmaps this week. Do not assume frontier model training hardware must run training; split your compute architecture so cheaper or alternate domestic chips handle inference while top-tier Nvidia clusters run fine-tuning. Test whether your serving stack can run across heterogeneous chip clusters by running your next small model deployment on a secondary hardware provider.