My 3 Favorite AI Stocks to Buy on the Continued Chip Sell-Off

The article highlights Nvidia at a discounted forward P/E of 16x (fiscal 2028) as it remains positioned to dominate AI training via CUDA, while also benefiting from inference/agentic workloads and faster-growing networking. AMD is pitched as a winner from inference and agentic AI, supported by expectations that the GPU-to-CPU ratio shifts from 8:1 to 1:1 for agentic AI and that the data center CPU market could double to $120B by 2030. Broadcom is portrayed as having a major custom-chip/TPU opportunity, with estimates that AI revenue could reach $180B (fiscal 2028) and custom AI chips could exceed $100B (fiscal 2027), supported by Alphabet’s reported up to $190B AI infrastructure spend and a separate $30B Apple data-center deal.

Analysis

This is not a thesis break on AI spend; it is a dispute over mix. If inference and agentic workloads keep growing, the marginal dollar shifts away from pure training GPUs toward memory, networking, and custom ASIC content, which structurally favors AVGO and, on share capture, AMD more than NVDA. That means the market may be underpricing a winner-takes-more dynamic inside the stack while simultaneously overpricing the idea that every supplier grows equally.

Near term, the next 1-3 months matter more than the long-term narrative. The cleanest signals will come from hyperscaler capex commentary and next-quarter data-center revenue prints: if spending pauses, the first deceleration should hit incremental GPU orders, while networking and custom silicon can keep compounding off already-committed designs. In that setup, NVDA can still be a good company but a worse stock if growth normalizes faster than the multiple can expand.

The contrarian miss is that "agentic AI = more CPUs = more total silicon" may be too linear. Agentic systems could also improve software efficiency and reduce brute-force compute intensity, which would compress unit demand even as workloads expand. Falsifiers are simple: a guide-down in AI capex intensity, NVDA data-center growth missing expectations, or AVGO AI bookings failing to accelerate into the next earnings cycle.

Key Decisions for Investors

• Put on a 2-3 month, beta-adjusted long AVGO / short NVDA pair: thesis is that custom-chip and networking content outgrows training-GPU share; target ~10-15% relative outperformance if hyperscaler capex stays firm.
• Do not add aggressively to NVDA until the next hyperscaler earnings/capex round confirms spending is re-accelerating; if data-center growth decelerates below consensus, treat rallies as sellable.
• Initiate a small starter long in AMD via call spreads or limited-risk equity size into the next 1-2 earnings cycles; this is a share-gain trade, not a guaranteed fundamental winner, so keep sizing below core semiconductor weights.
• Watch GOOGL and META capex commentary as the leading indicator: if they pull back AI infrastructure budgets, reduce exposure to the whole AI hardware basket and expect AVGO/NVDA multiple compression before fundamentals fully roll over.