AI too vital to slow, too fast to regulate: BofA

Bank of America remains bullish on AI semiconductors, projecting AI capital expenditure could triple to more than $3 trillion by decade-end despite calls for slower AI development. The Philadelphia Semiconductor Index is up 67% year-to-date but trades at 19x forward earnings, in line with the S&P 500, despite 139% year-over-year earnings growth; its 24-month forward P/E of 14.8x represents an 11% discount to the broader market. BofA expects near-term volatility from rates, Middle East tensions, U.S. midterms and data-center resistance, while identifying Micron's September 30 earnings as a key memory-pricing catalyst.

Analysis

The relevant differentiation is not broad AI demand but the ability to monetize constrained inputs. MU’s upside depends on whether contracted HBM and DRAM pricing converts into higher realized gross margin rather than merely securing volume; a favorable print would validate that memory has moved from the most cyclical part of the AI stack toward a capacity-disciplined oligopoly. That would also support SK Hynix exposure indirectly, while pressuring legacy-memory suppliers with weaker HBM qualification and lower-value product mix.

Near term, AI semis remain vulnerable to a factor-driven de-risking rather than a demand break: higher real yields, geopolitical risk, or election-related positioning can compress high-duration multiples regardless of utilization. Within the complex, ANET and AVGO should hold up better than GPU- or memory-pure plays if investors rotate toward revenue visibility, while MU is likely to exhibit the greatest upside/downside around its September 30 results. The key falsifier for a constructive memory view is a sequential decline in bit-price guidance, weaker HBM mix commentary, or evidence that customers are accumulating inventory ahead of new platform launches.

The contrarian risk is that consensus is treating long-term supply agreements as proof of durable pricing. Those contracts can protect revenue but may cap incremental upside if supply normalizes, while concentrated hyperscaler demand leaves the group exposed to any pause in cluster buildouts. A better signal than headline AI spending projections is whether customers accept higher-priced next-generation memory without offsetting cuts in conventional DRAM/NAND orders over the next one to two quarters.

Key Decisions for Investors

• Maintain a tactical long MU position only into the September 30 earnings catalyst, sized at half normal risk; add after results only if management guides sequential ASP and gross-margin expansion. Target a 10-15% post-print rerating versus a 7-8% stop on adverse pricing or HBM-mix commentary.
• Express relative AI durability through long ANET / short SOXX over the next 1-3 months. Networking revenue is less exposed than memory to spot-price reversals; exit if ANET underperforms SOXX by 8% or if hyperscaler capex guidance is cut.
• For existing broad semiconductor longs, rotate part of high-beta GPU and memory exposure into AVGO ahead of macro/election volatility. The trade favors proprietary connectivity and custom silicon cash flows over components with more cyclical inventory risk.
• Do not buy MU earnings options without checking implied volatility against the last eight post-earnings moves. If implied move exceeds the historical median by more than 25%, wait for confirmation in guidance rather than paying for pre-event convexity.