5 AI Stocks to Own for the Inference Age

The article argues that the AI shift from training to inference should benefit Nvidia, AMD, Broadcom, Cerebras, and Micron, with multiple cited growth drivers including $100B+ chip commitments and Broadcom AI revenue projections of $180B in fiscal 2028. It highlights Nvidia's inference-focused end-to-end racks, AMD's memory-heavy chiplet approach and rising CPU demand, Cerebras' 15x faster inference chips, Broadcom's custom ASIC opportunity, and Micron's HBM demand amid soaring DRAM prices. Overall tone is constructive on AI infrastructure beneficiaries, but the piece is largely thematic rather than event-driven.

Analysis

The key shift is not “more AI spend,” but a reallocation of the AI bill of materials from compute toward memory, networking, and custom silicon. That changes the profit pool: hyperscalers will keep pushing down GPU prices where they can, but they cannot compress HBM supply or board-level power constraints as easily, which makes the memory vendors and a few design enablers the cleaner beneficiaries on a 6-18 month horizon.

NVDA remains the platform winner, but the margin mix gets more nuanced as inference economics favor specialized racks and software orchestration over raw training horsepower. The hidden risk is that inference demand is highly price elastic: as tokens get cheaper, usage explodes, but that also invites customer in-sourcing and design substitution into ASICs, which caps the long-run attach rate of merchant GPUs. In other words, NVDA wins the volume race, but AVGO and GOOGL may take share of the most profitable unit economics.

AMD is the most interesting second-order beneficiary because it does not need to beat NVDA outright to matter; it only needs to be “good enough” in a memory-heavy workload where software and packaging matter more than peak FLOPs. The market may still be underestimating the CPU content uplift in inference servers, which can create an earnings tailwind even if GPU share gains are modest. That said, AMD remains execution-sensitive: any delay in ecosystem adoption or a setback in large customer qualification could push the story out by 2-3 quarters.

Cerebras is the optionality trade: if it escapes the premium niche, it could force a repricing of inference architecture economics, but the base rate is still low because deployment complexity and capex intensity are real barriers. Micron looks best on a fundamental basis because the demand impulse is coming into an already tight supply regime, and unlike GPU vendors it has a shorter path to monetization through pricing. The contrarian takeaway is that the trade is not simply “buy AI”; it is “own the bottlenecks,” while being selective on the names exposed to competitive commoditization.

Key Decisions for Investors

• Long MU on a 3-6 month horizon; use any HBM supply-chain wobble to add, as inference-driven memory tightness is the clearest near-term P&L lever. Risk/reward: asymmetry favors upside from pricing power, with downside limited unless DRAM supply normalizes faster than expected.
• Maintain core long NVDA, but hedge with a smaller short in a broad GPU basket or AMD vs NVDA pair if the market begins to price inference share gains too aggressively. Risk/reward: NVDA retains platform value, but valuation dispersion should widen as custom silicon adoption rises.
• Initiate a long AVGO / short semis-beta pair over 6-12 months. Broadcom monetizes the shift to custom ASICs and design wins, while generic compute suppliers face margin pressure as hyperscalers optimize for cost-per-token.
• Trade AMD tactically only on pullbacks and catalyst windows; consider a call spread into next-gen product/customer announcements rather than outright common. Risk/reward is favorable if inference and CPU attach rates ramp, but execution risk remains high.
• Avoid chasing Cerebras common absent a clear deployment inflection; if available, prefer event-driven exposure via structured options rather than directional equity. The odds of a niche-to-mainstream transition are low, but the payoff is large if OpenAI-linked volume converts.