The Best Stocks to Buy Right Now: Nvidia vs. AMD vs. Broadcom

The article argues that Nvidia, AMD, and Broadcom are all well positioned to benefit from the AI shift from training to inference and agentic AI. Nvidia is cited as the current leader with a potential $200B CPU market opportunity and $20B in CPU revenue this year, while Broadcom targets more than $100B in ASIC revenue in fiscal 2027. The author’s preferred pick is AMD, which is described as having two major early-stage growth opportunities in inference and agentic AI.

Analysis

The key second-order shift is that AI capex is fragmenting from a single-vendor GPU cycle into a three-layer stack: training, inference, and control-plane/agentic workloads. That broadens the winner set, but it also compresses the moat for any one supplier because hyperscalers are now optimizing for total system cost, not peak FLOPS. The most important implication is that networking, memory attach rates, and software lock-in become more monetizable than raw compute alone, which supports NVDA and AVGO while giving AMD a realistic share-gain path if it can keep software friction low.

AMD’s setup is the most asymmetric because its upside is not just unit growth but mix expansion in a market where memory bandwidth and power efficiency matter more than headline compute. If inference and agentic workloads shift CPU/GPU ratios toward parity, AMD’s underappreciated CPU franchise can become a larger profit pool than its GPU line on a marginal basis. The risk is execution: if ROCm fails to become “good enough” for enterprise deployment, the share gains could stay confined to a few lighthouse customers and the valuation rerate would stall.

Broadcom’s custom silicon story is less about displacing Nvidia than about capturing the hedge that hyperscalers are forced to buy when GPU economics get too expensive. The second-order effect is that every incremental TPU/ASIC deployment increases demand for adjacent networking and interconnect content, making AVGO a picks-and-shovels beneficiary of capex optimization. The main risk is cadence: custom-chip revenue can look explosive on paper but often lags design wins by 12-24 months, so the stock can outrun near-term fundamentals before supply ramps actually hit.

The contrarian view is that the market may be underestimating how durable Nvidia’s ecosystem advantage remains even as workloads diversify. If AI agents become sticky software platforms rather than commodity inference jobs, the winner may be whichever vendor controls the orchestration layer and developer toolchain, not the lowest-cost accelerator. That argues for staying long the whole basket, but being more selective on timing around any pullbacks in NVDA while using AMD and AVGO as higher-beta expressions of capex decentralization.

Key Decisions for Investors

• Stay long NVDA as the quality anchor, but treat it as a lower-beta compounder; use pullbacks to build, not chase, given the risk that consensus is already pricing some of the AI infrastructure leadership.
• Add AMD on weakness for a 6-18 month trade: highest upside if inference share gains broaden and ROCm traction improves; risk/reward is attractive because the market is still valuing it as a follower, not a platform winner.
• Initiate AVGO as a 12-24 month compounding position; prefer staged entry because custom-silicon revenue realization is delayed, but the path to higher mix and networking attach provides durable upside.
• Pair long AMD / short a basket of legacy x86 or smaller AI-adjacent hardware names to express share gain in inference and data center CPU intensity without taking full index risk.
• For tactical traders, use NVDA call spreads into pullbacks rather than outright calls; upside remains intact, but implied volatility likely overstates near-term incremental information flow.