These 5 AI Stocks Have Returned 300%-Plus Over the Past 3 Years -- and 1 Still Looks Dirt Cheap

The article frames the AI semiconductor boom as lifting major chip names: Nvidia shares are up 448% and it reported revenue +91% to $81.6B with non-GAAP EPS +140% to $1.87, while Micron shares are up ~1,320% and its Q3 revenue rose +345% to $41.5B and non-GAAP EPS increased +1,200%. It highlights strong current profitability for memory from supply-demand constraints (Micron gross margins ~85%), alongside a key risk that AI memory demand and pricing may eventually normalize as new capacity ramps. It also cites attractive valuation for Micron (P/E just over 19 vs sector average P/E ~35), but warns that any decline in data center spending would likely pressure margins and earnings.

Analysis

The trade is no longer just “AI up or down”; it is now about where the bottleneck sits. The most durable economics appear to be shifting toward scarce manufacturing capacity, advanced packaging, and network/custom-silicon integration, which is why TSM and AVGO have better forward visibility than names tied more tightly to spot pricing or headline AI spending. If hyperscaler budgets stay elevated, the next leg of returns should accrue to bottleneck owners rather than the most obvious GPU beta.

The main risk is that investors may be underweighting how reflexive memory economics are. MU can look cheap on trailing earnings precisely at the point where incremental supply becomes visible, and that is where margin compression usually starts before revenue rolls over; a few quarters of flat-to-down memory ASPs would be enough to break the narrative. AMD is also more vulnerable than NVDA to ROI scrutiny because its upside depends on continued share gains, not just secular industry growth.

Contrarian view: consensus may be too focused on whether AI spending grows, and not focused enough on which layer captures the value if spending rotates. Even if training budgets decelerate, inference, networking, and custom ASIC demand can keep AVGO and TSM compounding while MU’s pricing power fades. Falsifiers are straightforward: a guide-down in hyperscaler capex, sequential declines in HBM or DRAM pricing, or any evidence that foundry/packaging lead times are easing faster than expected.

Key Decisions for Investors

• Long TSM / short MU for 1-3 months; express the view that advanced-node scarcity outlasts memory pricing. Risk/reward is favorable if memory ASPs soften, and the thesis is wrong if MU maintains pricing power through the next two quarters.
• Overweight AVGO versus NVDA on a 6-12 month horizon. AVGO has more diversified monetization from custom silicon and networking, while NVDA is more exposed to multiple compression if AI ROI skepticism deepens; add on 3-5% pullbacks.
• Treat AMD as a tactical trade only, not a core allocation, until there is proof that AI capex is converting into sustainable data-center share gains. Reassess if management commentary or backlog conversion disappoints over the next earnings cycle.
• Use SMH or SOXX as a hedge vehicle around earnings rather than a pure long here; the basket can cushion single-name risk while preserving AI exposure if the sector keeps rotating within the stack.