A split like this is easy to miss at first. Now, its starting to become clearer.
The companies supplying the critical hardware semiconductors, high-bandwidth memory, networking silicon, and power infrastructure are still showing strong demand and, in many cases, expanding margins. Their customers, the hyperscalers and model builders writing the checks, face a different set of questions: how quickly free cash flow is converting, how much of the spend is generating incremental revenue, and whether the capital intensity remains sustainable. For most of the year those two groups moved together. Over the past several weeks that correlation has broken.
My process in these moments is straightforward. I favor businesses with pricing power, multi-year demand visibility, and the ability to turn revenue into free cash flow without constant equity raises or heavy leverage. I avoid names whose valuations still assume perfect execution or whose balance sheets look stretched relative to the duration of their contracts. When a theme is broad I will own a diversified set of the best enablers. When differentiation appears, I concentrate.
That framework shaped what I did this year. Early on I held meaningful positions in Nvidia, TSMC, Broadcom. Those were the cleanest ways to participate while demand was unconstrained and supply was the bottleneck. As evidence of sorting emerged, I adjusted. I reduced Super Micro after the multiple expanded faster than the visibility on sustained utilization. I trimmed some pure software names that were trading as AI proxies but showed limited evidence of AI-driven margin expansion. At the same time I added to Micron and AMD memory and competitive GPU capacity that benefit from the same buildout with different risk profiles—and increased my weighting in Broadcom as custom silicon demand from the largest buyers became more concrete.
Current positioning reflects the same discipline. My largest AI-related holdings remain Nvidia and TSMC for the core compute and manufacturing layer, Broadcom for custom accelerators and networking, and Micron for high-bandwidth memory. I hold a smaller but meaningful position in AMD as a diversifier and potential share-gainer. I have limited exposure to the most leveraged neocloud operators and have stayed selective on pure application-layer names until unit economics become clearer. I also maintain positions in a couple of the higher-quality power and cooling suppliers upstream of the data-center buildout, because those constraints look durable.
The sorting process is still early. Correlations can reassert themselves, and strong results from the largest buyers can re-accelerate the entire complex. But the questions have changed, and that is the signal I pay attention to. Who is generating returns on the capital being deployed? Who is funding the buildout on attractive terms? Which business models improve as the cost of intelligence declines, and which ones depend on scarcity remaining high? I will keep answering those questions the same way I have all year by following the cash flows and the balance sheets more closely than the narrative.
Disclaimer: Investing carries risk. This is not financial advice. The above content should not be regarded as an offer, recommendation, or solicitation on acquiring or disposing of any financial products, any associated discussions, comments, or posts by author or other users should not be considered as such either. It is solely for general information purpose only, which does not consider your own investment objectives, financial situations or needs. TTM assumes no responsibility or warranty for the accuracy and completeness of the information, investors should do their own research and may seek professional advice before investing.
