Skip to Content

Why is DeepSeek cutting AI prices when everyone else is raising them?

Is the Huawei Ascend 950 supernode actually powerful enough to run DeepSeek V4?

DeepSeek’s 75% price cut is more than a sale—it’s a move to local Huawei Ascend 950 supernodes. See how domestic chips are now driving AI cost deflation.

Why is DeepSeek cutting AI prices when everyone else is raising them?

Key Takeaways

What: DeepSeek permanently cut V4-Pro API prices by 75% to roughly $0.0035 per million tokens.
Why: To leverage domestic Huawei Ascend 950 silicon and bypass US export restrictions.
How: Months of hardware-specific tuning for upcoming supernode deployments optimized throughput, slashing operational infrastructure costs.

The Supernode Dividend: Why DeepSeek’s 75% Price Cut Isn’t a Marketing Stunt

Most industry watchers assume that China’s AI sector is currently in survival mode, struggling to stay afloat under the weight of US export restrictions. When DeepSeek recently slashed the price of its flagship V4-Pro model by a massive 75%, many saw it as a move to grab market share. However, the reality is more calculated and hardware-driven. This price drop is a “supernode dividend” triggered by a shift to local silicon.

While the Western narrative focuses on how sanctions limit access to Nvidia hardware, these same restrictions have created a massive domestic market for Huawei. DeepSeek spent months reworking its entire software stack to move beyond basic compatibility and into hardware-specific tuning for Huawei’s Ascend chips. This deep integration allows the V4-Pro to run with a level of efficiency that generic software cannot match. The 75% price cut is a direct result of the projected rollout of Huawei Ascend 950 supernodes in the second half of 2026, which is expected to increase throughput and lower operational costs.

Architecture of Autonomy

The V4-Pro is a heavyweight, boasting 1.6 trillion parameters, yet it remains efficient in its operations. It uses sparse attention mechanisms, a technical approach that allows the model to process huge amounts of data—up to a million tokens at a time—without needing an impossible amount of memory or computing power. By only “activating” about 49 billion parameters during any single task, the model keeps costs low while maintaining high performance.

The real breakthrough is what engineers call “day-0” adaptation. Usually, when a new chip comes out, it takes months or even years for software to run well on it. DeepSeek and other domestic vendors like Cambricon and Hygon managed to have their models ready the moment the hardware was available. This level of coordination was previously a unique advantage held only by Nvidia. By building specifically for the Ascend 950, DeepSeek has optimized for a local infrastructure that is immune to foreign trade policy.

Turning Sanctions into a Competitive Edge

There is a persistent belief that losing access to Nvidia’s top-tier GPUs would leave Chinese AI firms permanently behind. In reality, these restrictions created an exclusive market for Huawei. Because DeepSeek cannot rely on the “black market” for expensive foreign hardware, they have pivoted to local silicon, which has lowered their long-term infrastructure costs.

This shift is backed by massive capital. The China Integrated Circuit Industry Investment Fund, often called the “Big Fund,” is reportedly interested in leading a major funding round for DeepSeek, which has already seen its valuation climb toward $45 billion. This state-backed support ensures that the transition to local chips is a permanent structural change in how AI is built and priced in the region.

What This Means for the Global Market

For developers and businesses, the result is a price crash in the token market. API costs for the V4-Pro have plummeted to as low as $0.0035 per million tokens. At this price point, integrating advanced AI becomes incredibly cheap, undercutting both domestic rivals and Western firms.

Furthermore, DeepSeek has made the model easy to adopt by supporting standard OpenAI and Anthropic-compatible API formats. This makes it simple for companies to swap out expensive Western cloud models for more affordable local alternatives without rewriting their entire codebase. While independent third-party benchmarks at scale are still pending, the initial data suggests that the move to Huawei Ascend supernodes is creating a self-contained, resilient ecosystem that functions independently of the global semiconductor supply chain.