Nvidia is falling behind as semiconductors surge. Here's what analysts say is going on

Nvidia has lagged peers, flat since April 27 while Intel and Micron are up more than 30%, as investors reassess AI capex, memory-chip bottlenecks, and hyperscaler in-house chip efforts. Hyperscalers raised 2026 capex plans across the board, with Alphabet at $185B, Amazon at $200B, Meta at $135B, and Microsoft at $190B, while Nvidia guided FY2027 capex higher than FY2026. Analysts remain constructive long term, but the stock is under pressure from valuation concerns, competition from TPUs/Trainium, and an apparent rotation into other semiconductor names.

Analysis

The market is re-rating the AI stack from a pure “GPU scarcity” story into a broader infrastructure cycle where beneficiaries are now rotating to the parts of the stack with the tightest near-term supply or clearest spend visibility. That favors memory, networking, and the hyperscalers with the best ability to internalize workload economics, while NVDA is getting punished because it sits at the intersection of every good and bad AI narrative: it still benefits from capex growth, but it also becomes the easiest source of disappointment if customers substitute even 5-10% of demand to in-house silicon.

The second-order effect is that the winners may be the companies enabling AI density rather than the model owners. If capex keeps stepping up into 2026-2028, the bottlenecks likely migrate from GPUs to HBM, advanced packaging, interconnect, and power delivery. That implies relative upside for INTC/AMD/Micron-type beneficiaries can persist for months, but it also means the current move may be partly a squeeze on positioning rather than a clean fundamental inflection; once investors have chased the “everything AI” trade, spreads can reverse quickly if customers signal budget discipline in the next earnings cycle.

NVDA’s main risk is not demand destruction, but mix erosion and multiple compression: even if unit demand stays strong, a few percentage points of workload displacement toward TPUs/Trainium/other custom accelerators can flatten the growth narrative enough to cap upside. Conversely, if hyperscaler 2026 capex revisions continue higher, the market may eventually realize that NVDA’s lag is a timing issue, not a thesis break. The near-term catalyst map is binary: customer commentary on custom silicon adoption, any further capex raises, and memory supply updates over the next 1-2 quarters. The consensus is probably underestimating how much of the current dispersion is a relative-value trade inside semis, not a macro call on AI spending.

Key Decisions for Investors

• Maintain a tactical long INTC / short NVDA pair for the next 4-8 weeks: if memory and PC/server replacement demand keep improving, INTC has more operating leverage to the current bottleneck than NVDA, while NVDA remains exposed to multiple compression if custom ASIC adoption headlines continue.
• Buy 3-6 month call spreads in AMD versus outright stock exposure: AMD has the cleanest second-derivative benefit from AI capex spillover plus supply tightness, but the move already prices in some good news; spreads limit downside if the trade becomes crowded.
• Add a small long GOOGL / short NVDA relative-value leg into the next hyperscaler capex update window: GOOGL has optionality from TPU monetization and better control over internal economics, while any incremental disclosure on TPU adoption could pressure NVDA sentiment without requiring a demand miss.
• Use NVDA downside puts or put spreads into strength over the next 2-3 weeks rather than shorting stock outright: the stock can still rip on capex revisions, but the skew is attractive because the market is pricing execution perfection while allowing for only modest substitution risk.
• Stay underweight UBS-rated “too expensive for growth” names only if they are direct AI-levered suppliers with stretched multiples; otherwise prefer pair trades over outright shorts, because the AI capex cycle can keep momentum names elevated for another 1-2 quarters even if fundamentals are improving.