An Illustrated Guide to the Capital Market After the Deployment of DeepSeek V4: Zhipu and MiniMax Plunge, Nvidia Panics

Title: Illustrated DeepSeek V4 Landing in the Capital Market: Zhipu and MiniMax Plunge, Nvidia Panics

Author: Rhythm BlockBeats

Source:

Repost: Mars Finance

DeepSeek V4 has finally launched. This is a moment nearly five months in the making. The full 1.6T Pro version, featuring the 1T parameter MoE main model + 285B parameter Flash version, followed closely behind, all open-sourced on GitHub under the Apache 2.0 license, with weights and deployment code released simultaneously.

Once the model was released, the capital market responded in three independent yet interconnected ways.

Different Reactions in the Capital Market

A-shares in the computing power chain surged almost across the board. Cambricon posted 11 consecutive positive days, rising 3.7% in a single day, with a total increase of over 60% in the month. Haiguang Information touched a 10% daily limit during trading, closing +8.4%. SMIC A-shares rose +4.91%, Hong Kong stocks +8.81%. Huahong Hong Kong stocks soared to +18% intraday, closing at +12%. The Guotai ETF for innovative chips attracted 2.4 billion yuan in a single day, reaching a record high.

In contrast, the Hong Kong large-model companies showed a different picture. Zhipu (02513.HK) fell 8.07%, with a short-selling ratio of 9.9%. MiniMax (00100.HK) dropped 7.40%, with a short-selling ratio soaring to 22.87%. The latter recorded the highest single-day short-selling data in the Hong Kong AI sector over the past three months. Both are representatives of the upcoming AI IPO wave in Hong Kong in late 2025, with their IPO prospectuses emphasizing the core competitive advantage: “Self-developed foundational large models.”

The reaction on the other side of the Pacific was equally specific. Nvidia opened down 1.8% last night, briefly falling to -2.6% intraday, but closed flat for the day. Bloomberg’s market quick analysis compared this correction to the “DeepSeek moment” on January 27 with V3. The difference is that, back then, it was panic selling, evaporating $600 billion in market value in a single day. This time, it appears more like a revaluation, moderate in scale but with a clear direction. A new phrase appeared in buy-side research reports: “China’s AI inference demand is beginning to decouple from North America’s AI inference demand.”

Putting these three market segments together is the first verdict written by the market within 24 hours of V4’s landing. After open-sourcing triumphed, money began to re-align, and what could be priced was no longer just the model itself, but where the model runs—on which card, in which industry chain.

11 New Models in 30 Days, V4 Adds Fuel to the Open-Source Camp

The timing of V4’s release itself amplified this reaction.

Zoom out to the past 30 days. Between March 26 and April 24, at least 11 major influential models were released or significantly updated worldwide, covering nearly all major players. Anthropic Opus 4.6, Google Gemini 3.1 Pro, OpenAI GPT-5.5, Mistral Large 3, Meta Llama 4, Moon’s Dark Side Kimi K2.6, Alibaba Qwen3-Next, ByteDance Doubao 2.5 Pro, Tencent Hunyuan 3.0, Kimi K2.6 Plus, and finally, DeepSeek V4 released in the early hours of April 23.

On average, a new model emerges every 2.7 days. This is a pace that even fund managers don’t have time to fully read the release notes. But reviewing the AI asset K-line charts of Hong Kong and China over these 30 days, only one name leaves a continuous mark on the market: GPT-5.5 on April 8 drove Nvidia’s stock up 4.2% in a single day, peaking that day. Then, on April 23-24, DeepSeek V4 prompted consecutive jumps in the computing power chains of China and Hong Kong.

The difference isn’t in the models’ capabilities themselves. In most cases, the gap between these 11 models on the LMArena leaderboard is less than 50 points, placing them in the same tier. The difference lies in two overlapping factors.

First is open source. Among the top 10 models, only Llama 4 is open source, but its weight license includes a long list of commercial restrictions, leading to a lukewarm reception from the Western developer community. OpenRouter dropped out of the top ten just three days after launch. V4’s license is Apache 2.0, with no access barriers, unlimited commercial use, and inference code released simultaneously. This is the first flagship open-source model in the past six months to simultaneously pressure the closed-source camp in performance, price, and openness.

Second is timing. Against the backdrop of the closed-source camp continuously ramping up, the open-source narrative is being repeatedly squeezed. Opus 4.6 pushed the SWE-Bench code task to new heights, and GPT-5.5 set the price at $1.25 per million tokens. Whether open source can catch up with closed source has been debated in Silicon Valley for two years. V4, with an estimated 90 million monthly active users for its open-source flagship, has put this debate on hold.

According to a senior domestic fund manager during a roadshow, “Before V4, we discounted the valuation of large open-source models; after V4, that discount has started to work in the opposite direction.”

DeepSeek Changes the Pricing Table of the Computing Supply Chain

A line in V4’s release notes had never appeared in any official Chinese large-model documentation before: “Day 0 full-stack adaptation to Cambricon MindSpore 590 and Huawei Ascend 950PR, deployment code released simultaneously.” The significance of this line becomes clear only when connecting the three parallel threads over the past 12 months. These threads belong to hardware, software, and Silicon Valley’s reactions.

The first thread is on the hardware side. Huawei Ascend 950PR was officially mass-produced in December 2025, with FP4 compute power of 1.56 PFLOPS, HBM capacity of 112GB, marking China’s first AI chip to benchmark Nvidia’s B series in hardware metrics. In V4’s 1T parameter MoE inference tasks, single-card throughput is 2.87 times that of the H20. The supporting CANN 8.0 software stack optimizes the LLM inference framework down to the operator level. Benchmark results from DeepSeek show that, on Ascend supernodes (8 cards, 950PR), end-to-end inference latency is 35% lower than on comparable H100 clusters. Cambricon MindSpore 590’s data is even more aggressive, with FP8 compute power per chip comparable to H100, at less than half the price.

The second thread is on the software side. The vLLM mainline merged Cambricon MLU backend PR on April 22, enabling native support for non-Nvidia domestic GPUs in the inference framework for the first time. Haiguang Information’s DCU, through the ROCm ecosystem, takes another route but can fully run V4’s MoE routing layer. This means V4’s deployment is no longer limited to “only run on certain domestic cards,” but can be “chosen among multiple domestic cards.” Breaking dependency on single vendors is a key production milestone.

The third thread comes from Silicon Valley. On April 15, Jensen Huang was asked by analysts at TSMC’s earnings call about China’s domestic computing power progress. His response was cold and specific: “If they can really get LLMs to ditch CUDA, it would be a disaster for us.” Nine days later, DeepSeek provided an answer with a Day 0 announcement.

The phrase “domestic substitution” has been overused and lost meaning in the past three years. But after the morning of April 24, this issue finally had concrete data that the capital market could price. Single-card throughput, end-to-end inference latency, inference costs, and deployable commercial code quietly pushed this long-running rhetoric war into the realm of production.

The logic behind Cambrian’s 11 consecutive days of rising stock prices lies here. It is no longer just a “domestic GPU concept stock,” but a “DeepSeek V4 inference infrastructure supplier.” The same logic explains Huahong’s 12% stock increase, as it manufactures the 950PR with a 7nm equivalent process. Every token running on China’s domestically produced Ascend means some capacity originally destined for Nvidia and TSMC is now partially retained in the Pearl River Delta.

And the next steps are already laid out. Huawei’s roadmap plans for the 950DT (training version) to be delivered in Q4 2026, targeting “full-stack training of V5 or equivalent models on a 10k-card cluster.” If this path is successful, CUDA’s moat in China’s large-model training will downgrade from “necessary” to “optional.”