KEY POINTS
- AMD shares are up 114% in 2026 on record $5.8 billion data-center revenue, with the chip rally now broadening to Intel, Broadcom, and the memory names.
- HBM4 is the structural bottleneck on every Blackwell 300 and Rubin shipment, putting SK Hynix, Samsung, and Micron at the center of the next leg of the trade.
- Computex next week will deliver the next reset, with AMD's MI400 commentary and Rubin specifications the two events traders need to watch.
AMD shares have rallied 114% in 2026, the strongest run in the Magnificent-adjacent semiconductor cluster, and the latest quarter shows why: data-center revenue hit a record $5.8 billion, up 57% year over year, blowing past consensus by roughly $400 million. Intel is up double digits year to date for the first time since the Pat Gelsinger era began. Micron and SK Hynix have moved with them. The AI chip trade in 2026 looks fundamentally different from the AI chip trade of 2023 and 2024. It is no longer a one-name story.
The Real Number Inside AMD's Quarter
The headline AMD numbers — earnings beat, revenue beat, guidance raised — are not the actionable detail. The actionable detail is that AMD's data-center segment now grows at 57% year over year while Nvidia's grew 92% in the same quarter. The growth rates are converging in absolute dollars, not yet in scale. AMD's $5.8 billion quarter is roughly 7.7% of Nvidia's $75.2 billion data-center print, but the gap was closer to 5% a year ago. That is the share-shift narrative bulls have been waiting for, and it is finally showing in the segment data rather than in projections.
Intel is the more contrarian piece of the trade. The stock has rallied on the back of CPU demand from inferencing workloads, where customers running smaller models and agentic workflows are realizing that GPU rack time is wasted on tasks that fit comfortably on a Xeon or Granite Rapids node. Intel's foundry business also got a credibility bump from sovereign-AI buyers willing to take on a non-TSMC fabricator to diversify supply-chain risk after the 2025 export-control fight. CNBC's shift-in-leadership coverage lays out the rotation in detail.
The Memory War Is The AI War
The most underappreciated leg of the trade is high-bandwidth memory. Nvidia's Rubin R100 and R200 GPUs will each carry 288 gigabytes of HBM4 across eight stacks. AMD's MI400 will carry similar quantities. Every Blackwell 300 system shipping today is bottlenecked on HBM availability, not on GPU silicon. SK Hynix has already completed mass-production preparations for HBM4 and shipped its first qualifying samples to Nvidia earlier in the spring. Samsung is plotting a 50% expansion in production capacity this year and is in advanced negotiations to enter the HBM4 supply chain at Nvidia. Micron is the smaller third entrant and has been the surprise winner of 2026, with the stock outperforming both AMD and Nvidia on a year-to-date basis.
The structural reason is straightforward. The semiconductor industry shipped $791.7 billion in 2025, and the Semiconductor Industry Association projects $1 trillion in 2026. Logic capacity is doubling. Memory capacity is doubling. The constraint is fab time, not capital. Investors who own the cycle through SMH or SOXX are now getting paid as the trade broadens beyond the single-name Nvidia concentration that defined 2023 and 2024.
Where The Trade Goes Next
The next inflection is Computex, which opens in Taipei next week. AMD's Lisa Su arrived on May 20 ahead of Huang in a clear positioning move. Watch for AMD commentary on the MI400 rack-scale system, on its custom AI ASIC partnerships, and on networking. Broadcom and Marvell, the two custom-silicon and networking names that benefit from the same capex wave, are likely to move on any Computex commentary that ties their ASIC roadmaps to hyperscaler design wins. Google's TPU v7, Meta's MTIA v3, and Amazon's Trainium 3 are all in production or near production, and Tom's Hardware's roundup of the custom-ASIC field shows just how broad the silicon battlefield has become.
The risk to the broadening thesis is concentration. If Nvidia's Rubin ramps faster than expected on the back of SK Hynix and Samsung HBM4 yield, the company could re-take more than its historical share of data-center capex by Q4 and pull the trade back to a single name. The window for AMD, Intel, Broadcom, and the memory names to consolidate gains is roughly the next two quarters. After that, the Rubin cycle dictates terms.
For traders, the actionable setup is barbell rather than rotation. Own Nvidia for the absolute dollar growth, own AMD and the memory names for the share-shift narrative, and use SMH or SOXX to dilute single-name risk. Watch the SMH/NVDA ratio. It has expanded sharply since March as the trade broadened. Continued expansion confirms the rotation is durable. A sharp reversal would signal Rubin is closer to volume than Computex chatter currently implies.
