On July 25, 2026, as global demand for AI large model training and inference continues to rise, the US stock AI computing power sector once again became the market focus. The competition between chip giants NVIDIA and AMD in the cloud AI chip market has entered a white-hot stage, while the capacity bottleneck of high-bandwidth memory (HBM) has become a key factor constraining computing power expansion. The Philadelphia Semiconductor Index (SOX) rose 4.2% this week, and investors remain confident in the long-term growth of AI hardware.
NVIDIA Blackwell Ultra enters mass production early, consolidating dominance
According to the latest news, NVIDIA CEO Jensen Huang announced during a recent investor conference call that the next-generation AI chip Blackwell Ultra will begin small-volume shipments earlier than planned, at the end of 2026, a full two quarters ahead of schedule. The chip uses TSMC's enhanced 3nm process, integrates over 200 billion transistors, and achieves FP8 performance of 5.6 PetaFLOPS, 2.3 times that of the current H200. NVIDIA stated that Blackwell Ultra is designed for trillion-parameter large models, with memory capacity increased to 288GB HBM4, which is expected to significantly reduce inference costs.
Following the announcement, NVIDIA's stock rose 3.8% on the day, with its market capitalization again approaching $4 trillion. Analysts noted that NVIDIA, leveraging its CUDA ecosystem and early layout, still holds over 80% of the AI training market share, but AMD's catch-up pace is accelerating.
AMD MI400 series makes strong impact, GPU market accelerates differentiation
AMD was not to be outdone, officially launching the MI400 series accelerator cards this week. The series is based on the CDNA 4 architecture, using a chiplet design that integrates 8 compute chiplets and 4 HBM3E stacks, achieving FP16 performance of 3.2 PetaFLOPS, narrowing the gap with NVIDIA's Blackwell Ultra in integer operations. AMD specifically highlighted that the MI400X has a memory bandwidth of up to 6.4 TB/s, thanks to HBM3E memory co-developed with Samsung. AMD CEO Lisa Su stated that the MI400 series will provide samples to customers such as Microsoft Azure and Meta in the third quarter, and is expected to capture 5% of the cloud AI chip market share by 2027.
Market research firm IDC predicts that the global AI chip market will reach $120 billion in 2026, with cloud inference chips accounting for 45%, up from 30% last year. AMD, leveraging the cost-effectiveness of the MI300X, has already won orders in some inference scenarios, and the MI400 series is expected to further erode NVIDIA's monopoly in high-end training.
HBM storage in short supply, Samsung and SK Hynix accelerate expansion
HBM memory, as the "heart" of AI chips, directly determines the ceiling of computing power. Currently, the global HBM market is duopolized by Samsung and SK Hynix, which together hold over 90% share. However, with NVIDIA and AMD launching new-generation chips, HBM demand is growing exponentially. According to Semiconductor Industry Association (SIA) data, total HBM shipments in 2026 are expected to reach 280 million units, an 85% increase from 2025, but capacity expansion is constrained by advanced packaging technology.
Samsung has announced a $5 billion investment in 2026 to expand its Pyeongtaek P3 plant, dedicated to HBM4 production, with plans to increase monthly capacity to 300,000 12-layer stacked chips. SK Hynix is even more aggressive, building its first HBM-dedicated fab in Cheongju, expected to begin production in 2027. However, in the short term, the HBM supply-demand gap remains at 15%-20%, keeping prices high. Samsung also announced a cross-license agreement with Micron this week, aiming to alleviate patent barriers and accelerate technology iteration.
Industry insight: Investment logic under the computing power arms race
From a long-term perspective, capital expenditure on AI infrastructure is far from peaking. Cloud giants such as Microsoft, Google, and Amazon collectively spent over $70 billion in capital expenditures in Q2 2026, with 60% used for AI server procurement. Chip orders from NVIDIA and AMD are booked through mid-2027, but analysts warn that the risk of over-reliance on a single supplier is accumulating. Some companies are beginning to explore in-house chips, such as Google's TPU v6 and Amazon's Trainium3, but in the short term, they are unlikely to displace the x86+GPU combination.
For US stock investors, the AI computing power track remains the most certain growth area, but attention should be paid to the following points:
- Valuation bubble risk: NVIDIA's current P/E ratio exceeds 80 times; if performance falls short, it could trigger a sharp correction.
- HBM supply chain lock-in capability: Chipmakers with stable HBM supply will have an advantage; watch investment opportunities in Samsung and SK Hynix.
- Downstream application deployment speed: The penetration rate of large models in scenarios like healthcare and autonomous driving will directly drive computing power demand.
Outlook: AI computing power competition enters multi-dimensional game
It is foreseeable that the AI chip market in 2027 will feature a pattern of "NVIDIA leading, AMD catching up, and cloud vendors supplementing with in-house chips," while HBM as a strategic resource will become a focus of competition. Technological breakthroughs in advanced packaging by TSMC and Samsung could further rewrite industry rules. Investors should remain vigilant while seizing structural opportunities brought by technological iteration.
