On July 28, 2026, the US semiconductor sector saw a long-awaited broad rally. At the close, the Philadelphia Semiconductor Index rose 2.7%, with AMD (Advanced Micro Devices) surging 5.3%, leading major chip stocks. On the news front, AMD announced pre-market today that its latest AI accelerator MI400 series has secured large orders from multiple cloud computing giants, totaling over $60 billion, far exceeding market expectations. This news instantly ignited investor enthusiasm for the AI computing track.
How Powerful is the MI400?
The AMD MI400 series was officially released in March this year, adopting the new CDNA 4 architecture with 153 billion transistors, achieving FP8 computing power of 1.6 PetaFLOPs, directly competing with NVIDIA's Blackwell B200. More importantly, AMD integrates 288GB of HBM3E high-bandwidth memory in the MI400, with memory bandwidth up to 5.2 TB/s, effectively alleviating the "memory wall" problem that previously constrained AI chip performance. SEMI analysts point out that the MI400's cost-effectiveness in LLaMA-3 large model inference scenarios exceeds NVIDIA's B200 by about 15%, which is the core reason attracting cloud computing giants to place orders.
Supply Bottlenecks Breaking?
Over the past two years, the AI chip production capacity bottleneck has mainly been concentrated on TSMC's CoWoS advanced packaging and HBM memory supply. However, as Samsung, SK Hynix, and Micron successively expand HBM3E capacity this year and next, and TSMC's new packaging plants in Chiayi and Kumamoto gradually come online, the delivery cycle for AI chips has shortened from the previous 40 weeks to 28 weeks. AMD CEO Lisa Su said in today's conference call: "The yield rate of MI400 is ramping up quickly, and Q3 shipments are expected to increase by 20% compared to the original plan." In addition, Micron mentioned in its latest earnings report yesterday that its HBM3E product capacity for fiscal 2026 has been fully sold out, further confirming the strong supply and demand situation.
Is NVIDIA Under Pressure?
Although AMD's market share in the data center GPU segment is still far below NVIDIA (currently about 12% vs 85%), the strong performance of MI400 has led investors to re-evaluate the competitive landscape. Wall Street investment bank Evercore ISI upgraded AMD's target price from $200 to $230 in a research report released today, citing that "the structural shift of the high-end AI chip market from a single supplier to a duopoly will boost the entire industry's profit pool." Meanwhile, although NVIDIA's H200 and B200 series are still in short supply, their order growth is marginally slowing, and some customers are adopting a "multi-cloud + multi-chip" strategy to reduce dependence on a single supplier. For example, Google has already transferred some TPU training tasks to the AMD platform early this year to balance cost and performance.
How Much Room for AI Compute Demand?
Despite market concerns about AI investment returns, actual demand continues to explode. Leading large model companies like OpenAI and Anthropic continue to purchase chips for training next-generation multimodal models; meanwhile, AI inference demand is also rising exponentially with user growth of applications like ChatGPT and Claude. IDC's latest forecast shows that the global AI server market will reach $165 billion in 2026, a year-on-year increase of about 65%. Among them, the proportion of servers used for inference will rise from 30% last year to 45%. Therefore, both NVIDIA and AMD will continue to benefit from this wave.
New Supply Chain Trends: Advanced Packaging and Optical Interconnects
Beyond CoWoS, TSMC is advancing higher-density 3D packaging technology, while Intel is also trying a differentiated route. At the same time, to further increase inter-chip interconnect bandwidth, silicon photonics and CPO (co-packaged optics) technologies are emerging. Broadcom announced today that its 7nm CPO switches for AI clusters have begun shipping to major cloud vendors, with per-port rates reaching 1.6Tbps. This marks an accelerating shift in AI compute clusters from "compute bottlenecks" to "interconnect bottlenecks." In the future, growth certainty is also high for supporting areas such as network chips and optical modules.
How Should Investors View This?
At the US stock investment level, the AI computing sector has moved from "wild growth" to a "differentiation phase." In the past two years, the entire sector rose almost together; now it's necessary to identify who has a true moat. For conservative investors, it is recommended to focus on equipment and foundry leaders like TSMC and ASML; for aggressive investors, companies like AMD and Micron that have expectation gaps in the "catch-up" landscape are worth noting. In addition, AI infrastructure-related sub-sectors such as power, cooling, and server liquid cooling are also worth tracking.
Future Outlook
AMD's MI400 orders exceeding expectations not only boost the company's stock price but also send a positive signal to the market: AI computing demand is far from peaking, and competition will drive accelerated innovation. With more AI chip launches in the second half of the year (such as NVIDIA's Rubin architecture, Intel's Falcon Shores, etc.), volatility and opportunities in the US semiconductor sector will increase. Investors are advised to closely monitor AMD's Q2 earnings report to be released next week, as well as NVIDIA's GTC conference at the end of August.
