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

The article highlights a favorable long-term setup for AMD, Broadcom, and Micron as AI shifts from training to inference and agentic AI. AMD is positioned to benefit from rising CPU demand in AI servers and new rack-level solutions, Broadcom expects more than $100 billion of ASIC revenue in fiscal 2027, and Micron sees secular HBM/DRAM demand with a forward P/E below 8x. The piece is primarily an investment thesis rather than fresh company-specific news, so near-term price impact is likely limited.

Analysis

The market is moving from a single-vendor training capex boom to a more fragmented inference buildout, which is structurally better for suppliers with differentiated packaging, networking, memory, and system integration. That shifts incremental economics away from raw GPU share and toward the stack around the chip: CPUs, HBM/DRAM, custom interconnect, and full-rack design. In that regime, the winners are likely to be the firms that can monetize attach rates and design wins across more sockets per deployment, not just the fastest accelerator.

AMD looks underappreciated not because it beats Nvidia on peak performance, but because inference economics reward total system cost and memory density. The ZT acquisition is strategically important: it gives AMD a way to capture margin beyond silicon and to bundle rack-level solutions, which can compress customer procurement cycles by 2-4 quarters. The main risk is execution timing; if hyperscalers continue using Nvidia as a default while waiting for software maturity, AMD’s upside may be pushed into 2026 rather than the next two quarters.

Broadcom is the cleanest second-order beneficiary because custom ASIC adoption is a response to compute saturation and power constraints, not just a search for cheaper chips. The hidden issue is that ASIC revenue tends to be lumpier and more design-win dependent, so the market may underprice the optionality until orders convert into multi-year production ramps. Micron’s setup is even more asymmetric: inference and agentic workloads increase memory intensity faster than wafer supply can respond, so pricing power can persist longer than the historical DRAM cycle suggests.

The contrarian risk is that consensus may be extrapolating a smooth transition when the real path is uneven: AI capex budgets could rotate, not expand, and one or two large model-efficiency breakthroughs could temporarily reduce accelerator and HBM demand. That would pressure AMD and MU first, while AVGO would likely lag later because ASIC programs have longer cancellation latency. The best hedge is to stay long the infrastructure bottlenecks, but size them as a basket rather than a single-name momentum trade.

Key Decisions for Investors

• Long AMD vs short NVDA as a 6-12 month relative-value trade: thesis is that inference/rack integration narrows the gap in TAM participation, while NVDA remains exposed to any moderation in training-led capex; use a 1:1 notional pair and cut if NVDA re-accelerates software monetization.
• Add AVGO on pullbacks for a 6-18 month horizon: best risk/reward is owning the custom-ASIC tollbooth with networking attached; target continued multiple support if design-win disclosures convert into visible backlog.
• Buy MU on weakness, but only as a cyclical-secular hybrid: 3-9 month trade on HBM pricing strength with a longer-duration thesis on AI memory intensity; prefer staged entries because the stock can re-rate sharply on any inventory scare.
• Initiate a basket long of AMD/AVGO/MU funded by a small short in INTC for the next 2-4 quarters: the market is likely to reward exposure to AI attach rates and punish legacy CPU share loss, even if INTC’s downside is slower than headline sentiment implies.
• Use call spreads rather than outright longs for AMD and MU if entering after a strong tape: upside remains meaningful, but implied volatility can compress quickly once the market recognizes the story; spreads preserve convexity while reducing premium bleed.