Full Breakdown
DeepSeek’s V4 Model: Aggressive Pricing, Huawei Partnership, and the Ongoing US-China AI Gap
4/29/2026, 10:58:52 AM
Launch of V4 and Pricing Strategy
On 24 April 2026 DeepSeek unveiled its V4 family—V4-Pro and V4-Flash—while announcing permanent price cuts that lower input-cache-hit fees to US $0.14 per million tokens and a 75 % discount on V4-Pro for developers through 5 May, targeting cost-sensitive users.
Background and Key Figures
DeepSeek first entered global headlines in early 2025 with its R1 model, which matched leading Western systems at a fraction of the training cost. Founder Liang Wenfeng’s equity rose to 34 % in early 2026, and senior researcher Chen Deli promoted the V4 launch as staff grew from 212 to 270.
Pricing and Statistics
DeepSeek’s V4-Pro costs US $0.0036 per million input tokens—about 97 % less than OpenAI’s GPT-5.5, which charges US $0.5 per million cached tokens. The resulting per-conversation price is roughly 32 times lower than GPT-5.5.
Hardware Collaboration and Impact
V4 runs on Huawei’s Ascend 950 series and uses the CANN software stack, marking DeepSeek’s first full-stack support claim. Analysts note the co-optimisation improves domestic deployment but leaves the model 3-to-6 months behind U.S. frontier systems.
Official Statements
DeepSeek said the price reductions are “effective immediately” and permanent, and that the V4-Pro discount will run through 5 May. Huawei’s WeChat post confirmed its Ascend processors “offer full-stack support” for V4, emphasizing joint development of a domestic AI stack.
Criticism and Opposition
U.S. officials note DeepSeek still trains models on banned Nvidia Blackwell GPUs and relies on distillation from Western systems, limiting true independence. Industry observers warn that a Huawei-first release could undermine U.S. semiconductor leadership.
Conflicting Reports and Gaps
Benchmark firm Artificial Analysis places V4-Pro just behind GPT-5.5, while Bloomberg and the Council on Foreign Relations claim the model does not meaningfully narrow the U.S. lead. Reports also differ on whether V4 was trained on Nvidia Blackwell chips or solely on Huawei Ascend hardware. Founder Liang Wenfeng’s public profile remains absent.
Verbatim Quotes
- “484 days later, we humbly share our labour of love. As always, we stay true to long-termism and open source for all.” — Chen Deli, senior researcher, DeepSeek
- “It shows how Chinese AI labs are learning to turn hardware constraints into a design problem, but it is still far from complete AI stack self-reliance,” — Poe Zhao, founder, Hello China Tech newsletter
- “The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation (the US).” — Jensen Huang, CEO, Nvidia
Outlook
DeepSeek is courting investments from Tencent and Alibaba, targeting a valuation above $20 billion. The firm plans broader V4 deployment on Huawei Ascend supercomputers while maintaining its low-price API, as analysts anticipate ongoing price competition.
