Nvidia is facing a challenge: demand is exploding, but the market has already started pricing for Nvidiaization On April 30th, Nvidia fell 4.63% in a single day, with its market value evaporating by approximately $230 billion overnight and its market value breaking through $5 trillion. On the same day, the financial report still exploded: FY2026 Q4 revenue was around 68.1 billion US dollars (+73% year-on-year), and FY2027 Q1 guidance was around 78 billion US dollars, not including China's data center revenue. Blackwell's monthly production capacity is approaching 1 million pieces, with a target shipment of~20 million pieces by 2026. The GAAP gross profit margin is still as high as 75.2%. It is simply a legal money printing machine to have such high profits after printing for several years. There is no problem with the demand, the problem is that the market sees something else. 1、 The four major clients in the United States are voting with their feet Four ultra large scale cloud vendors are collectively doing one thing: moving inference workloads away from NVIDIA GPUs, and NVIDIA's moat in the training field remains unshakable 1. Google TPU v7: self-developed seventh generation, combined with the Four party Supply Chain Combination Fist, with almost full stack self-sufficiency in internal reasoning Meta MTIA: announces the launch of 4th generation self-developed chips within two years, and signed a joint development agreement with Broadcom in April 3. Amazon Trainium/Inferentia: Meta just signed a large order of millions of Gravitons with AWS this month 4. Microsoft/OpenAI: Advancing the self-developed ASIC roadmap simultaneously, OpenAI is still betting on multiple lines with Broadcom and AMD According to data provided by TrendForce, the shipment of self-developed ASICs by Hyperscale is expected to increase by 44.6% in 2026, while GPU shipments will only increase by 16.1%; The customized chip market has surged to approximately $118 billion this year. Morgan Stanley's share structure is also changing: Nvidia~85%, ASIC slightly over 10%, AMD slightly below 5% -10% may not seem like much, but last year this number was only in single digits. 2、 What Lao Huang was worried about in China has still happened What is severely underestimated by the market is the rise of Chinese inference chips, and the pace is stuck in the large-scale shipment of 2026 H2. 1. The Ascend 950PR inference chip will be mass-produced in March 2026, with a single card inference performance claimed to be 2.87 times that of H20. Tencent and ByteDance have already ordered 100000 units; Ascend 920 and 950 series enter the window of large-scale shipment in the second half of the year 2. Ascend 384 Super Node: Accumulated shipment of~300 sets in September 2025, reached~550 sets in January 2026, and increased by 1.8 times in 4 months 3. Huawei's 910C series production capacity target for 2026 is around 600000 pieces - this is just one company 4. At least 9 Chinese AI chip companies have orders or shipments exceeding 10000 kcal: Ascend, Cambricon, Haiguang, Kunlun Chip, Pingtouge, Mole Thread, Muxi, Biren, Tiantianzhixin This is another comprehensive field after the Chinese model More importantly, there is a comprehensive reversal on the model side. On April 24th, DeepSeeker V4 was released, with 1.6T parameters that can run on domestic chips without restrictions. Chinese AI chip companies immediately made comprehensive adaptation. Moreover, DeepSeek's computing power has always been abundant, and it can even do large-scale promotional activities. V4 Pro is 25% off until May 31st. The model Day 0 has been successfully implemented on domestic chips, which means that the moat of CUDA on the inference side has essentially failed in the Chinese market. Superimposed export controls also block H20, a special supply card. After 2026 H2, the main battlefield of China's reasoning power is basically not NVIDIA's the final say. 3、 What is the market pricing The extreme scenario of the market two years after pricing. NVIDIA on the training side will not be replaced in the short term, and the CUDA ecological moat is still there; But reasoning is the real mainstay of AI implementation, and it has the lowest requirement for generality and is most sensitive to unit token costs. This is exactly ASIC's home ground. Once a large-scale customer switches over 50% of their inference computing power to self-developed chips, Nvidia's 75% gross profit margin immediately faces repricing. This is a typical expectation gap moment, where the fundamentals are at their best, but the market begins to discount the growth rate from 70% to 30%. What is bought is redeemed, what is sold is the slope. 4、 What is the conclusion? My three judgments are: 1. Nvidia will not be able to sustain its training market and Rubin roadmap for another two to three years. Trillion orders are real 2. But the end of Davis' double-click is likely to come, and what he wants to earn in the future is no longer valuation expansion, but EPS growth money 3. The true alpha lies in the card players, Broadcom (doing ASIC for Meta, Google, OpenAI), Marvel, TSMC Advanced Packaging Production Capacity, and HBM. These are the tickets that truly benefit from the narrative of de Nvidia