The MoatNarrow moat

Here is the puzzle at the center of AMD. It is the greatest comeback in the history of the semiconductor industry — a decade ago a two-dollar penny stock left for dead, today, under Lisa Su, a company worth roughly $1 trillion¹ that took the server market away from Intel and became the only credible alternative to Nvidia in AI chips. And yet it trades at about 159 times trailing earnings and nearly 25 times sales, the richest valuation² of any chip company in this collection, while its returns on capital still sit near or below its cost of capital. My argument is that both things are true at once, and reconciling them is the entire investment case: AMD is a superbly-run number two with a genuinely narrow moat, priced as if it will build a wide one by breaking Nvidia's monopoly. You are not paying for the business AMD is; you are paying for the business the market bets it becomes.

The one part of the moat that is truly structural — not earned, but owned — is the x86 duopoly. By a tangle of decades-old cross-licensing, only AMD and Intel⁷ may legally make the x86 chips that run virtually every Windows PC and most of the world's servers, and no amount of capital lets a newcomer join the club. That single fact is why AMD, through all its near-death years, was never actually displaced: there was no third x86 maker to take its place. Everything else in the franchise — the relentless share gains since the 2017 Zen launch, EPYC's climb from a rounding error to roughly 40% of the server market³ — is magnificent execution layered on top of that license, and execution, unlike a license, has to be re-won every generation.

The reason the stock re-rated into the hundreds of billions is not the CPU business but the AI one, and that is where the moat is thinnest. AMD's Instinct accelerators — the MI300, now MI350, soon MI400 — are the only GPUs bought in volume as an alternative to Nvidia, and a world terrified of depending on a single supplier has anointed AMD the designated second source; data-center revenue more than doubled year-over-year in mid-2026. But being the second source is a role the market grants, not a moat AMD owns, and its Achilles' heel is software. Nvidia's CUDA, built over nearly two decades, is a self-reinforcing ecosystem that AMD's ROCm is still racing to match — and an ecosystem lead can widen faster than a challenger closes it.

What made the comeback possible was engineering, not structure: AMD pioneered chiplets — building a processor from small modular dies instead of one big one — which handed a far smaller company a cost and yield edge over Intel, and it went fabless, letting TSMC manufacture its designs while Intel's own fabs stumbled. Both were decisive. But note what kind of advantage they are. Chiplets are now an industry standard rivals have copied, and the TSMC edge is rented, not owned — Nvidia builds at the very same foundry. The most durable piece is the sticky, high-margin embedded franchise AMD bought with Xilinx in 2022: reconfigurable chips locked into seven-year design sockets in industrial, aerospace, and automotive gear. It is the closest thing AMD has to a wide moat — but it was purchased, not built, and its roughly $49 billion price tag⁴ is exactly what still drags returns on capital below the cost of capital.

So hold the two truths together. The business is real and improving — 2025 revenue up 34% to $34.6 billion⁵, a record 2026 on doubling AI sales — and the moat is under permanent assault from every side: Intel is a wounded but larger incumbent with state backing, Nvidia dominates AI in silicon, software, and networking, the hyperscaler customers are designing their own chips, and Arm is encroaching on x86 itself. The verdict is narrow, and strengthening, and dangerously priced. The number that decides whether the price was sane is AMD's AI-accelerator share: at under 10% today⁶, the valuation only makes sense if that climbs durably into the double digits against a rival pulling away in software. If it does, the narrow moat widens and the multiple was foresight; if it stalls, this is a cyclical challenger priced for a coronation that didn't come.