The Chip Shortage Just Got a 2027 Date Stamp

Lisa Su flew to Taipei this week with a harder message than a victory lap. The Advanced Micro Devices (AMD) CEO told reporters she already lifted supply through 2026, and demand is still higher. She said AMD will **substantially increase supply in 2027** and still “can definitely use more.” She is meeting Foxconn, TSMC, and other partners so CPUs, GPUs, packaging, and memory can ramp together. Planning with those partners used to look one or two years out. Now it stretches **three to five years**. My take is simple. Treat this as a date stamp on a multi-year squeeze. The market keeps arguing about whether AI spending is a bubble. The people building the chips are arguing about who gets the next wafer and the next stack of memory. ## What Su actually put on the table Her message is blunt. Demand for AI and high-performance computing stays **very, very high** for the next several years. AMD needs more advanced wafer capacity. Memory stays tight industry-wide, including the high-bandwidth memory that rides next to AI accelerators. She is raising AMD’s Taiwan supply-chain commitment past the **more than $10 billion** plan announced in May. South Korea is next on the memory stop, where Samsung and SK Hynix sit. That is a CEO talking like a logistics chief. When a top AI GPU rival and a serious CPU house still cannot fill the order book after a full year of ramps, you are looking at factories that cannot spin up as fast as software teams can dream up new models and agents. AMD’s market value recently crossed the trillion-dollar line in the same AI rally that made Nvidia a household name. Su’s trip is the receipt under that valuation. Customers still want more silicon than the chain can ship. ## The real bottleneck is a stack, not one ticker Wall Street loves to turn every chip headline into a single-stock fight: Nvidia versus AMD, AMD versus Intel, TSMC versus everyone. The better map is three layers that have to show up on the same calendar. **Logic wafers.** Advanced nodes at Taiwan Semiconductor Manufacturing (TSM) are the scarce starting point. Su said AMD needs more advanced wafer capacity, period. Reports around the same complex say next-gen server CPUs and AI platforms are already fighting for 2-nanometer space years ahead. **Memory.** Micron has been saying the quiet part on earnings: much of 2027 output is already spoken for, high-bandwidth memory is sold out deep into the book, and the crunch can get tighter into 2028. That matches Su’s memory stop in Korea. If the accelerator ships without the stacked memory, it is a paperweight. **Packaging.** CoWoS-style advanced packaging is where dies and memory become a working AI brick. Capacity is expanding, and it is still a traffic jam. One missing package line can hold a whole rack order. That is why Su keeps saying capacity has to come online together. A beautiful GPU design dies on the loading dock if memory or packaging slips a quarter. ## What this means if you own stocks I am not handing out buy and sell tickets. Here is how I read the board. Names tied to **scarce physical capacity** still sit in the driver’s seat while demand outruns supply. That means advanced foundry, high-bandwidth memory, and the packaging and substrate houses that finish the job. Designer chips matter too, but a great roadmap without allocation is a press release. Hyperscalers and model labs will keep signing multi-year deals because they cannot risk empty racks. That supports revenue stories at AMD, Nvidia, and the memory makers. It also keeps capital spending loud and the debt debate loud at the same time. Main Street feels this when memory gets rationed toward data centers. The same shortage that thrills AI bulls is a tax on phones, PCs, and other gadgets that used to treat DRAM like a cheap commodity. ## What would change my mind I would soften this read if foundry and memory books rolled over together, if hyperscalers cut capex in a cluster instead of one noisy quarter, or if Su and her peers started talking about excess inventory instead of three-to-five-year planning. One soft session in the semiconductor ETF is not that. A CEO on her third Taiwan trip of the year, still hunting wafers and memory, is not that either. As for me, I respect the builders who admit demand still beats supply more than I respect the hot take that every shortage ends on schedule. Factories are slow. Models are fast. That gap is the story. ## Bottom Line Lisa Su just told the market the AI chip shortage still has years on the clock. AMD will ramp hard in 2027 and still wants more, because wafers, high-bandwidth memory, and packaging have to arrive together. Treat that as an operating map for the stack, not a victory lap that the squeeze is over.