We are in a race for chips and memory, says ODDO BHF's Stéphane Houri

Memory prices are expected to stay elevated for the next 2 to 3 years, supported by ongoing AI investment and hyperscaler capex. Stéphane Houri of ODDO BHF said AI upside is broadening beyond Nvidia to power, connectivity and CPUs, implying wider sector opportunities. The message is constructive for AI-linked suppliers and hardware beneficiaries, though it is more commentary than a direct market-moving catalyst.

Analysis

The more interesting takeaway is that AI capex is morphing from a single-name GPU trade into a broader industrial utility cycle. If memory stays tight for 2-3 years, the market should stop treating DRAM/NAND pricing as a mean-reversion story and instead price it like a capacity-constrained supply chain with multi-quarter earnings persistence, which tends to favor the lowest-cost producers and punish downstream hardware assemblers with weak pricing power.

That has second-order implications for Nvidia: even if demand remains intact, the equity can lag because the marginal dollar of AI spend is flowing into adjacent bottlenecks and enabling infrastructure rather than only into accelerators. In other words, the next leg of AI monetization may accrue to names tied to power delivery, networking, and CPU/server balance rather than the most crowded AI beneficiary, creating a relative-value opportunity within the AI stack.

The key risk to this thesis is not demand collapse but supply response. Memory is historically one of the fastest industries to overbuild once pricing turns up, so the relevant horizon is months to quarters, not days; if capex inflects sharply, the current tightness can unwind faster than consensus expects. The contrarian read is that the market may already be underweight the durability of AI infrastructure spend outside GPUs, while simultaneously overpaying for duration in the most obvious AI winner.

Key Decisions for Investors

• Long a basket of memory suppliers on pullbacks for 6-12 months; prefer the lowest-cost operators over cyclical laggards. Risk/reward improves if spot pricing remains firm through the next two earnings cycles, with downside if capex guidance rolls over.
• Pair trade: long AI infrastructure enablers (power, networking, CPUs) vs short NVDA or a GPU-heavy basket for 3-6 months. Thesis: the marginal AI dollar is broadening, so relative multiple expansion should shift away from the most crowded name.
• Consider call spreads on select memory names with 9-18 month expiry to express the view that elevated pricing persists, while capping premium at risk if supply ramps faster than expected.
• Use any sharp rally in NVDA to trim exposure and rotate into less-consensus beneficiaries; the setup favors relative-value rotation more than outright index beta.
• Set a catalyst watch on hyperscaler capex guidance and memory capex announcements over the next 1-2 quarters; a step-up in supply spending would be the clearest sign the trade is entering the late stage.