3 AI Chip Stocks Down 15% or More to Buy Right Now

The article argues AI capex is sustaining demand for AI chips and networking, despite volatility, highlighting three buy-the-dip names: Broadcom down ~23% with projected >$100B ASIC revenue in FY2027 (vs Citigroup’s ~$180B in FY2028), AMD down ~18% as inference/agentic AI partnerships (OpenAI, Meta) and CPU demand expand, and SK Hynix down ~20% amid an HBM-driven memory supercycle lasting to at least 2030 with forward P/E around ~6x.

Analysis

This is less a broad AI-equity rerating than a supplier mix shift: every incremental dollar of hyperscaler AI spend is moving more toward custom silicon, interconnect, and memory than toward generic accelerators. That structurally favors AVGO because it monetizes both the design layer and the networking layer, while NVDA faces a longer-term share leak at the margin as customers internalize more of the stack. GOOGL and META are indirect winners because custom silicon lowers marginal inference cost and improves payback on capex, which should extend spend even if headline growth slows.

AMD is a more conditional winner. The market will reward it only if inference utilization, not just design wins, turns into visible revenue per installed rack; otherwise the stock is vulnerable to the usual “AI story vs. earnings” compression. The second-order positive is server CPU attach in agentic workflows, but that benefit likely lands over months, not days, and can be overwhelmed by pricing pressure if inference capacity gets commoditized faster than unit demand grows.

SK Hynix is the cleanest 6-18 month structural beneficiary because HBM tightness is a capacity, not just a demand, story. The risk is that the market extrapolates scarcity too far: any meaningful HBM capex acceleration, packaging substitution, or AI capex pause would knock down the supercycle narrative quickly. Near term, the tape should keep rewarding suppliers with the highest bottleneck leverage, but the decisive falsifier is a slowdown in hyperscaler capex commentary or evidence that custom silicon reduces total HBM intensity per workload.

Key Decisions for Investors

• Long AVGO vs. short NVDA for 3-6 months; thesis is that custom ASIC and networking content captures a larger share of AI spend than merchant GPU units. Target 10-15% relative outperformance; exit if NVDA re-accelerates training demand or AVGO ASIC commentary disappoints.
• Buy AMD only on weakness or via 3-6 month call spreads; this is a utilization-confirmation trade, not a blind momentum chase. Best risk/reward if inference revenue starts showing up in partner disclosures within the next 1-2 quarters; thesis breaks if share gains remain narrative-only.
• Add SKHY on pullbacks as a 6-12 month structural long; HBM scarcity supports pricing and margins longer than a normal memory cycle. Falsifier: signs of HBM oversupply, aggressive competitor capex, or a sharp reset in AI memory demand.
• Relative-value long GOOGL/META against the semiconductor basket if the market starts rewarding capex efficiency over pure spend growth. The beneficiaries of lower AI unit costs can see operating leverage even if semiconductor multiples stay volatile.