The Chip Rally Isn't Over: Here Are My 3 Top Picks

Hyperscaler capex is projected to approach $700 billion in 2026, up more than 60% year over year, with spending concentrated in chips, networking, and memory. The article argues AI investment still sits below prior technology cycles as a share of GDP, implying room for further expansion rather than a near-term peak. Longer-term estimates of trillions needed by 2030 suggest the AI infrastructure buildout is still in the middle of its cycle.

Analysis

The key second-order effect is that capex intensity is shifting from “model training” to the full industrial stack: power delivery, photonics, high-speed interconnect, advanced packaging, and memory bandwidth. That broadens the beneficiary set beyond the obvious GPU complex and makes the cycle less prone to a single-vendor bottleneck; when one constraint eases, spending simply moves to the next bottleneck. The result is a longer runway for suppliers with pricing power in infrastructure rather than just exposed to AI inference hype.

The market is still underappreciating how much of this spend is effectively non-discretionary once clusters are planned. Hyperscalers can slow headline growth, but they are now in a phase where underbuilding would risk ceding model performance and enterprise adoption to peers, so capex cutbacks are likely to be lumpy rather than structural. That argues for a multi-quarter earnings tailwind for the picks-and-shovels names, especially where lead times and qualification cycles protect margins.

The main risk is not demand collapse but digestion: if supply chain capacity expands faster than utilization, the trade can rotate from scarcity winners into quality at a lower multiple. Another risk is power or grid constraints, which can delay monetization and push out revenue recognition even as capex stays elevated. A meaningful slowdown would likely require a step-down in hyperscaler free cash flow discipline, not a change in AI conviction, so the reversal trigger is probably measured in quarters, not days.

Consensus is probably too focused on whether AI demand is “real” and not focused enough on how capital intensity itself creates a self-reinforcing industrial cycle. In that sense the move is underdone in networking, memory, and outsourced manufacturing/assembly names with less obvious AI branding. The asymmetric opportunity is in the vendors that monetize each incremental dollar of spend regardless of which model wins.

Key Decisions for Investors

• Go long a basket of AI infrastructure enablers on any 3-5% pullback over the next 1-2 weeks; prefer networking, memory, and advanced packaging over pure-play GPU exposure. Risk/reward: 15-25% upside over 6-12 months if capex remains >$650B, with ~10% downside if the market digests near-term order strength.
• Pair trade: long AI infrastructure suppliers / short software names with stretched AI monetization assumptions over 3-6 months. The thesis is that capex converts to revenue faster than application-layer monetization, so near-term EPS revisions should favor the supply chain.
• Use call spreads on a diversified semiconductor equipment or networking ETF for 2-4 quarter exposure rather than outright equity beta. This captures the capex continuation while limiting drawdown if hyperscalers pause after a big spend quarter.
• Avoid chasing the most crowded GPU leader after strength; instead rotate into second-order beneficiaries where estimates have not fully caught up to the buildout. The setup is better on lagging names with improving order books than on consensus winners already reflecting perfect execution.
• If power-grid or data-center bottleneck headlines accelerate, add to the infrastructure trade; if hyperscaler commentary shifts toward capex discipline, reduce exposure quickly because the sector will re-rate on guideposts before the revenue data confirms it.