Technology & Science
DeepSeek Unveils V4 LLM Optimized for Huawei Ascend Chips, Undercutting U.S. Rivals on Price
On 24 April 2026, DeepSeek released preview versions of its open-source V4 model that run natively on Huawei Ascend NPUs, matching top U.S. systems in benchmark tests while charging roughly one-sixth their inference prices.
Focusing Facts
- V4-Pro packs 1.6 trillion parameters, handles 1 million-token context windows, and is priced at $1.74 per million input tokens and $3.48 per million output tokens—versus $5 and $30 for OpenAI’s GPT-5.5.
- Huawei’s Ascend A2/A3/950 processors were used in portions of V4’s training and are fully validated for ‘day-zero’ inference, making V4 the first leading Chinese LLM launched without primary reliance on Nvidia GPUs.
- Huahong Semiconductor and SMIC shares rose 15 % and 10 %, respectively, after the launch on expectations of wider domestic chip adoption.
Context
China’s decision to pair a frontier AI model with home-grown silicon echoes the Soviet Union’s 1957 Sputnik launch: a technologically symbolic moment that signaled capability parity rather than outright dominance. Just as Sputnik spurred the U.S. to pour resources into space and STEM, DeepSeek-Huawei cooperation pressures Washington to rethink export-control leverage built on Nvidia’s hardware monopoly. Technologically, the move accelerates two long-running arcs: the open-sourcing of once-proprietary AI (mirroring the shift from IBM mainframes to commodity PCs in the 1980s) and the broader decoupling of U.S.–China tech stacks that began with the 2018 ZTE sanctions. Over a century horizon, the event matters less for today’s leaderboard scores than for the precedent that advanced AI can be trained and served economically outside the U.S. hardware ecosystem; if replicated, it could erode the geopolitical ‘compute choke-point’ that has shaped digital power since the first silicon foundries in the 1960s.
Perspectives
Asia-Pacific mainstream media
e.g., The Straits Times, Newsday — Portrays DeepSeek V4 as a milestone proving China can now field world-class AI on Huawei chips, signalling growing technological self-reliance and narrowing of the gap with U.S. labs. Coverage leans into the narrative of regional tech ascendancy and downplays unresolved IP-theft allegations and the fact that independent benchmarks are still pending, likely reflecting geopolitical and commercial interest in showcasing China’s progress.
Tech-enthusiast trade press & blogs
e.g., Mashable, Techmeme, MIT Technology Review — Frames V4 as an exciting open-source, cost-slashing rival that could democratize advanced AI and pressure proprietary U.S. models on price and features like long context and agentic coding. These outlets often amplify vendor benchmark slides and hype to engage their tech-savvy readership, so their reporting tends to accept DeepSeek’s self-reported performance at face value and gives limited scrutiny to technical caveats or alleged data misappropriation.
Western business & tech-skeptical press
e.g., Australian Financial Review, The Register — Argues that while V4 cuts costs, it still fails to close the capability gap with U.S. leaders and its benchmark claims require caution, offering relief to Washington and Silicon Valley. By catering to Western investors and incumbent firms, this coverage may accentuate doubts about Chinese breakthroughs and highlight allegations of stolen IP, reinforcing a narrative of sustained U.S. dominance even when evidence is mixed.
Like what you're reading?