Amazon flips the AI capex narrative with surging cloud profits

Citi expects shortages across the global memory semiconductor market to deepen through 2031, driven by rising demand for high-bandwidth memory, DRAM and NAND as artificial intelligence systems adopt continual learning. In a research note published Monday, the investment bank identified continual learning as a potential central theme in AI development over the next five years, an approach that updates AI models with new tasks while retaining prior knowledge and so drives demand for both memory and data storage. Citi forecasts HBM bit demand will increase 62% year over year to 75.2 billion gigabits in 2027, followed by a further 69% rise to 127.0 billion gigabits in 2028, and it expects demand for enterprise solid-state drives to expand as AI systems require additional storage capacity. For DRAM, Citi projects global demand growth of 30% in 2027 and 35% in 2028, against supply growth of 19% and 22%, producing supply-demand ratios of -8.7% and -9.7% for the respective years. The bank anticipates a similar imbalance in NAND memory, forecasting demand growth of 29% in 2027 and 33% in 2028 versus supply increases of 21% and 25%, with NAND supply-demand ratios of -6.1% in 2027 and -5.5% in 2028 compared with an estimated -0.8% in 2026. In its equity research, Citi named Samsung Electronics, SK Hynix, Micron, Sandisk and Kioxia as its preferred memory semiconductor stocks, and highlighted Montage, Applied Materials, Lam Research, TES, Eugene Tech and TechWing within semiconductor equipment and materials.

Warren Buffett and newly appointed Berkshire Hathaway CEO Greg Abel have plowed tens of billions of dollars into Alphabet, making it one of Berkshire's biggest investments, after Buffett finally bought the stock in the third quarter of 2025 with shares trading around $200. Buffett had long said he missed the opportunity, noting at the 2017 shareholder meeting that he would not bet against Google, and he has acknowledged missing out on a 9,000% gain by not buying sooner. Google initially offered shares to the public at $85 each in 2004, and after a 2014 split into two share classes and a 20-for-1 split in 2022, the split-adjusted price is just $2.125 per share. Buffett said he likes the stock now because Alphabet plans to spend around $200 billion in capital expenditures this year on new data centers and AI servers, a use of capital he sees earning a high cash return. Google Cloud CEO Thomas Kurian said the average payback period on its new servers is less than two years, and roughly half that for Google's custom TPUs, while average five-year customer contracts and a two-year lead time on data center construction mean even a worst-case scenario produces a positive return. With the stock trading for less than 17 times forward earnings expectations, Buffett and Abel could keep buying in the third quarter.