The release of Deepseek V4 shakes the market—how will it impact the AI industry chain in Hong Kong stocks?

On April 24, 2026, China's AI sector reached a milestone moment.DeepSeek officially released and open-sourced its new series of models, DeepSeek-V4,which is divided into two versions: Pro (1.6 trillion parameters) and Flash (284 billion parameters), with a context length of one million tokens becoming a standard feature across the series. This seemingly routine technical iteration, however, created massive waves in the Hong Kong stock market.

On the day of the release, the 24th, the Hong Kong stock market’s AI sector showed extreme divergence, with both extremes represented. On one hand, AI star stocks that were all the rage earlier in the year suffered heavy losses: $MINIMAX-W (00100.HK)$ dropping 9.44%, $KNOWLEDGE ATLAS (02513.HK)$ plunging 9.05% in a single day and breaking below the thousand-yuan mark, while the 'Token First Stock' $XUNCE (03317.HK)$ fell sharply by 15.98%; on the other hand, domestic chip concept stocks collectively surged. $SMIC (00981.HK)$ A rare 10% surge drove up sector sentiment, $MONTAGE TECH (06809.HK)$ with a 13.78% jump on the 27th.

The market is voting with real money, revealing a core contradiction: Is DeepSeek V4 a 'depth charge' igniting a valuation bubble or a catalyst driving industry growth? How will this technological revolution reshape the value distribution landscape of the Hong Kong stock market’s AI industry chain?

Price War – From 'burning cash for market share' to an 'efficiency revolution'

The most direct impact of DeepSeek V4 comes from its pricing strategy,On the evening of April 25, DeepSeek announced a limited-time 60% discount on the V4-Pro model API; on the 26th, it further reduced the price of input cache hits across all DeepSeek series APIs to one-tenth of the original price. After the latest adjustment, the cost per million tokens for DeepSeek-V4-Flash is only 0.02 yuan for input cache hits, and 0.025 yuan for DeepSeek-V4-Pro.

This 'rock-bottom' pricing strategy has disruptive significance. Compared with similar overseas models, V4-Pro's output pricing is about one-fifth to one-eighth of foreign models, while V4-Flash offers the lowest pricing in the entire industry. Haitong International analyst Zhang Xiaofei frankly stated: 'DeepSeek V4 achieves top-tier performance at a low cost.'

The immediate effect of the price war is squeezing the profit margins of pure large model companies,Models such as MiniMax’s abab series face direct threats, and Zhipu’s expectations of improving profitability through price hikes have been completely shattered.As foundational models become increasingly open-source and cheaper, the business logic of relying on model service fees faces fundamental challenges.It is reported that since April 24, MiniMax has fallen more than 15%, and Zhipu has dropped over 22%. Notably,In July, two stocks will face share unlockingwith Zhipu's six-month lock-up period ending, 5.8% of its shares will be unlocked; MiniMax's unlocking ratio is even higher at 39.0%. By January 2027, when Zhipu’s 12-month lock-up period ends, 90.3% of its shares will be unlocked.

Token Economics — From 'General Scarcity' to 'Scenario Premium'

Amid a general decline in token prices, a phenomenon is occurring: the total token market is rapidly expanding. According to data from the National Data Bureau, as of March 2026, China’s daily average token usage has exceeded 140 trillion, growing over a thousand times compared to 100 billion at the beginning of 2024, with an increase of more than 40% in the past three months.

DeepSeek V4’s push for lower token prices has actually accelerated the stratification evolution of the token economy:

First Layer: Basic Token Layer. The price of general-purpose tokens continues to drop, becoming infrastructure similar to utilities like water, electricity, and gas. DeepSeek V4-Flash’s input cache hit price per million tokens is 0.02 yuan, already approaching infrastructure pricing levels.

Second Layer: Scenario-Specific Token Layer. Tokens tailored to specific scenarios receive high premiums. In high-barrier sectors such as finance and energy, fragmented, complex, and heterogeneous private data are converted into standardized inputs that large models can efficiently consume, enhancing value output for each invocation.

Third Layer: Application Value LayerAI-native SaaS companies accelerate the embedding of AI capabilities through low-cost tokens, creating new application value. The drop in price lowers the threshold for application development, stimulating innovation.

General data cannot support high-precision decision-making, making it difficult to consistently produce results in high-barrier scenarios such as finance and energy. As a result, specialized tokens have rapidly evolved from being an 'optimization option' to becoming a 'necessity,' with rigid demand for their use rising sharply.

Domestic computing power - from 'import dependency' to 'ecosystem independence'

V4 overturns the 'system table': Amid fierce competition from GLM, Kimi, Qwen, Gemini, and other strong players, V4 not only addresses shortcomings in coding, mathematics, and long-context processing but also completes compatibility with domestic chips. On the same day as the model's release, eight domestic AI chip manufacturers—Huawei Ascend, Cambricon, Hygon, Moore Threads, Muxi Corporation, Baidu Kunlun Chip, Alibaba Pingtouge, and Iluvatar CoreX—all collectively completed full-link adaptation and performance optimization.

This means that domestic large models are rapidly progressing from being 'usable' to 'user-friendly,' while the domestic computing power industry is seeing systematic opportunities moving from single-point chip substitution to full-link 'model-compute synergy.'

DeepSeek V4’s deep compatibility with domestic chips injects a shot of adrenaline into the Hong Kong stock semiconductor sector. Orient Securities believes,DeepSeek-V4's powerful performance and high cost-effectiveness will significantly enhance the global competitiveness of domestic models, potentially driving substantial growth in AI application demand at low cost, thereby boosting the prosperity of the domestic computing power chip and semiconductor equipment supply chain.。

$MONTAGE TECH (06809.HK)$ : Global leader in memory interface chips, with a market share of 36.8%, a key component in AI server memory modules

$SMIC (00981.HK)$ : Leading domestic AI chip manufacturer, handling compute power chip orders

$HUA HONG SEMI (01347.HK)$ : Power semiconductors, MCUs, etc., are synergistic with AI computing power demands

$LENOVO GROUP (00992.HK)$ : Server and PC manufacturers benefit from the AI PC concept

$GIGADEVICE (03986.HK)$ : Memory chip design benefits from AI server memory demands

Additionally, attention can be paid to the continued benefits to underlying AI infrastructure driven by AI computing power demands. SEMI China President Lily Feng stated in March 2026 that global AI infrastructure spending will reach $450 billion in 2026, with inference computing power accounting for more than 70% for the first time, driving strong demand for GPUs, HBM, and high-speed networking chips. $YOFC (06869.HK)$ 、 $CSOP SK Hynix Daily (2x) Leveraged Product (07709.HK)$ 、 $CSOP Samsung Electronics Daily (2x) Leveraged Product (07747.HK)$ 、 $GIGADEVICE (03986.HK)$

The release of DeepSeek V4 marks a new phase of development for China's AI industry. This technological revolution is not a zero-sum game but a catalyst for upgrading the entire industrial chain. For investors, the key now is not to avoid volatility but to identify structural opportunities within the differentiation, capturing the dividend of value redistribution amidst the wave of industrial transformation.

1.3M Views