One year after the DeepSeek shock of January 2025, the Chinese startup has released its next flagship model. DeepSeek V4, published on April 24, 2026, processes up to one million tokens in a single context and costs 97 percent less than OpenAI's GPT-5.5 according to official pricing. The model runs entirely on Huawei chips rather than Nvidia hardware, demonstrating that US export restrictions have not halted China's AI development.
What DeepSeek Is and How It Got Here
DeepSeek is a spinoff from the quantitative hedge fund High-Flyer, founded by Liang Wenfeng. The startup shook technology markets in January 2025 when it demonstrated with its model R1 that efficient AI development is possible without Western high-performance chips. Nvidia's share price fell by nearly 17 percent as investors reassessed a core assumption underlying AI investment: that scaling is necessarily dependent on Nvidia's hardware.
V4 is the successor. It comes in two variants: V4-Pro with 1.6 trillion parameters, of which 49 billion are activated per request, and V4-Flash with 284 billion parameters. The Mixture of Experts approach, in which only a fraction of parameters is active at any time, substantially reduces computational requirements compared to conventional full activation.
Why a Million-Token Window Changes the Competitive Landscape
The context window of V4-Pro holds one million tokens, eight times the 128,000-token window of GPT-4o. In practice this means V4 can analyze a novel, complete source code, or months of email correspondence in a single pass. Liang Wenfeng stated at the release that the model requires only 27 percent of the computational resources of its predecessor V3.2 for million-token processing, with a KV cache consumption of just 10 percent.
In standardized benchmark tests for million-token contexts, V4-Pro surpasses Google's Gemini 3.1 Pro according to company data: 83.5 percent versus 76.3 percent in the MRCR-1M test. Independent benchmark analyses, such as those by AkitaOnRails, show that V4 matches or exceeds OpenAI's frontier models in mathematics and coding.
The pricing is the sharpest argument. V4-Flash costs $0.14 per million input tokens and $0.28 per million output tokens. The South China Morning Post calculated, citing DeepSeek's official pricing, a cost advantage of 97 percent over GPT-5.5 in comparable performance tiers. For comparison: OpenAI's reasoning model o1 costs $15 per million input tokens and $60 per million output tokens. DeepSeek's R1, the predecessor reasoning model, sits at $0.55 and $2.19 respectively, a difference of 96 percent.
For companies integrating large language models into their products, this represents a fundamental shift in operating costs. A company currently spending $50,000 per month on OpenAI APIs could, under comparable conditions with V4-Flash, bring that figure to around $1,500. For startups building AI-powered applications, this difference determines whether the business model is viable.
Huawei Instead of Nvidia: The Real Geopolitical Signal
DeepSeek runs its data centers on Huawei Ascend 950 chips, with the Ascend 910C as the compute core. Both are Chinese designs outside the scope of US export restrictions. The United States had progressively imposed export bans on Nvidia's high-performance chips since October 2022, including the A100, H100 and H800. The underlying logic: without top-tier chips, no competitive AI scaling.
DeepSeek shook that assumption with R1 and has definitively disproved it with V4. Semiconductor analyst Linley Gwennap wrote in April that Huawei plans approximately 750,000 units of the Ascend 950 in 2026, for the first time reaching a scale that carries real weight against Nvidia's data center business. Outside China, Nvidia remains dominant in the short term because Ascend hardware is barely available internationally. In the medium term, the dependencies are shifting: Chinese AI providers are no longer reliant on Western exports.
For Western governments, this creates a strategic problem. Chip export restrictions have not secured a development lead but have placed China's semiconductor industry under pressure to develop independent alternatives more rapidly. What was intended as a brake has functioned as an accelerator.
First External Funding at a $45 Billion Valuation
DeepSeek had been entirely funded by its parent company High-Flyer until May 2026. Now the company is raising its first external funding round. Bloomberg and Technode report the targeted valuation at $45 billion, roughly double what it was two weeks earlier. Tencent, Alibaba and China's state semiconductor investment fund Big Fund are in discussions as potential investors. The amount to be raised is reported at three to four billion dollars.
The round is not a conventional startup financing step. It signals that Beijing's strategic industrial policy treats DeepSeek as a national AI platform. For Western companies, this raises a question with new urgency: what performance gap still justifies using OpenAI or Anthropic APIs when DeepSeek delivers comparable results at a fraction of the price? The answer will emerge from corporate adoption over the next twelve months.
For US technology policy, this is a sobering scorecard: four years of export restrictions, multiple updates to the Entity List and billions in CHIPS Act subsidies for domestic chip production have not significantly slowed China's AI development. At the same time, Western companies that need cheaper AI APIs for their products are increasingly tempted to switch to DeepSeek, thereby using Chinese infrastructure. That is the real strategic paradox that V4 makes visible.