Why the Second Wave of AI Will Mint More Millionaires Than the First -- and the Stocks to Own

The article argues that the next phase of AI will broaden opportunity beyond Nvidia, highlighting AMD, Broadcom, and Micron as beneficiaries of inference and agentic AI. AMD is cited for its inference GPUs, AI server CPU share gains, and ZT Systems acquisition; Broadcom is positioned to benefit from custom AI ASIC demand with more than $100 billion in ASIC revenue expected in fiscal 2027; Micron is expected to ride persistent HBM/DRAM demand, with shares described as inexpensive at under 8x forward earnings.

Analysis

The market is likely underestimating how much of the next AI capex cycle migrates from generalized GPU buying to a more fragmented stack where inference efficiency, memory bandwidth, and rack-level integration matter more than raw FLOPS. That shift should compress Nvidia’s relative scarcity premium while improving pricing power for the “picks-and-shovels” layer around accelerators, memory, networking, and system integration. In other words, the next phase is less about a single winner and more about the vendors that sit closest to bottlenecks in power, memory, and deployment speed.

AMD looks like the cleanest leveraged beta to that transition because it can monetize both silicon and full-system design, but the bigger second-order effect is that its upside depends on execution credibility in customer qualification cycles. If it wins even a modest share of inference racks, the operating leverage is significant; if launches slip, the market will punish the multiple before revenue catches up. The key risk is that hyperscalers keep broadening internal ASIC efforts, which could cap the long-run share available to merchant GPU vendors even as total AI spend rises.

Broadcom’s setup is stronger than the headline suggests because custom silicon is not just a cost-saving story; it is a strategic lock-in mechanism for hyperscalers that want to reduce dependence on merchant GPUs and control power envelopes. That should also spill over into networking, where attach rates can rise as clusters get denser and more heterogeneous. Micron remains the most asymmetric beneficiary because memory is the scarcest input in inference-heavy workloads, and the longer-term implication is that HBM scarcity can sustain pricing power even if GPU unit growth normalizes.

The contrarian issue is that consensus is likely extrapolating a straight-line AI spend curve when the real risk is demand digestion after the current buildout wave. The weakest link is not end-demand for AI; it is conversion of committed capex into utilization fast enough to justify continued orders. If utilization disappoints over the next 2-3 quarters, the market could rotate from growth multiple expansion to margin scrutiny, which would hurt AMD most and leave AVGO/MU comparatively better insulated.

Key Decisions for Investors

• Long AVGO vs short NVDA over the next 6-12 months: best risk/reward if inference/custom silicon keeps gaining share; NVDA’s moat is still durable, but relative multiple expansion looks harder from here.
• Initiate a staged long in MU on any 5-8% pullback, with a 6-9 month horizon: memory scarcity and HBM attach rates should keep earnings revisions positive, but expect high volatility around supply announcements.
• Buy AMD only on weakness or via call spreads into next 1-2 quarters: upside is large if rack-level wins translate, but execution risk is high enough that outright long is a lower-conviction entry than AVGO or MU.
• Pair long AVGO / short a basket of hyperscaler capex proxies if you expect custom ASIC adoption to accelerate: this isolates the monetization layer that benefits regardless of whose chip design wins internally.
• Set a tactical stop on AMD if customer ramp timing slips by a quarter: the stock’s multiple is vulnerable to any sign that inference share gains are back-end loaded rather than near-term.