This chart shows just how much bigger the AI infrastructure boom is about to become. UBS estimates the annual rack capacity deployed for AI will increase from 9.5 GW in 2024 to 104.5 GW in 2030. That is an elevenfold expansion in only six years, with most of the acceleration happening after 2026. The important part is that every new gigawatt of AI rack capacity creates demand across the entire infrastructure stack. More racks mean more GPUs and custom chips, more HBM, more networking and optics, and significantly more spending on power and cooling. Now here are some of the stocks that will benefit from this. $NVDA remains the largest winner because its annual rack capacity is expected to increase from 5.9 GW in 2024 to 50.5 GW by 2030. This means Nvidia could still represent almost half of all new AI rack capacity despite growing competition. The biggest change is the rise of custom AI chips because AI ASIC capacity is forecast to increase to 45.1 GW by 2030, nearly matching Nvidia. These chips are designed for large customers such as Google, Amazon, Meta and Microsoft that want better cost and power efficiency for specific workloads. This does not necessarily mean custom chips replace Nvidia. Nvidia should remain the preferred platform for flexible training workloads because of its hardware and CUDA software ecosystem. Custom ASICs are likely to gain share in large, predictable workloads where hyperscalers can optimize the chip for their specific models. $AMD also benefits but UBS expects it to remain a smaller third player. Its projected annual rack capacity reaches 8.9 GW in 2030, giving it significant growth without suggesting it will overtake Nvidia or custom silicon. The larger opportunity extends beyond the accelerator companies. Whether a rack uses an Nvidia GPU, an AMD GPU or a custom ASIC, it still needs HBM, advanced packaging, networking, optics, storage, power delivery and cooling. $MU, $SKHY and Samsung should benefit from rising HBM and server-memory demand. TSMC, ASE and Amkor should benefit from manufacturing and packaging increasingly complex AI chips. Broadcom, Marvell, Credo and Astera Labs should benefit from connecting those chips inside and between racks. Super Micro, Dell and HPE should benefit from building the servers and racks around the accelerators. Vertiv, Eaton, Schneider Electric and Modine should benefit from the power and cooling challenge. AI racks are becoming more power-dense, meaning data centers need more advanced electrical equipment and liquid-cooling systems. Bullish on the entire AI infrastructure stack and I’m already positioned across memory, networking, power, cooling, and neoclouds, and if you want to see exactly what I hold to capture this buildout, check out my full Milk Road Pro portfolio below.