Here's Why I Believe AI Is Not a Bubble

Article argues the AI rally has “room to run” rather than being a bubble, citing Micron’s fiscal 2026 Q3 revenue up more than 4x YoY and multi-year customer deals aimed at boosting demand durability. It also highlights hyperscaler capex support—Amazon raised its annual capex forecast from $200B to $220B due to higher memory costs—suggesting AI hardware orders remain resilient. Overall, it frames physical AI (robots, autonomous vehicles) as an emerging demand catalyst for compute, memory chips, and data-center infrastructure.

Analysis

The important shift is not “AI demand exists,” but that demand is becoming contractable and financeable, which changes the earnings quality of the supply chain. That favors memory and adjacent constrained inputs first: MU and SNDK should see better utilization, less inventory risk, and higher willingness from customers to pre-commit even at elevated pricing, which supports multiple expansion beyond the usual cyclical semiconductor discount. NVDA still benefits, but at this point the market is paying for perfection; the incremental upside is less about demand discovery and more about supply discipline.

For the hyperscalers, the near-term issue is not a capex cliff but ROIC compression. AMZN, MSFT, and GOOGL can absorb higher AI spend for several quarters, yet the equity risk is a slower conversion of capex into operating leverage, especially if memory and power costs keep rising faster than AI revenue per deployed rack. That creates a second-order winner in infrastructure and a loser in any vendor whose model depends on customers accepting lower returns on the next dollar of spend.

The contrarian miss is timing: physical AI is a real option, but the monetization window is likely 6-18 months for infrastructure and 5-10 years for autonomy/robotics. That means the current move is probably underdone in contracted AI inputs and overdone in long-duration “physical AI” narratives. Falsifiers are simple: if AMZN/MSFT revise capex down, if MU/SNDK commentary loses contract visibility, or if customer pre-buys revert to spot-style purchasing, the thesis weakens quickly.

Key Decisions for Investors

• Long MU on pullbacks over the next 1-3 months; thesis is that contract visibility supports a higher trough multiple, with a stop if management commentary shows DRAM/NAND pricing or backlog duration rolling over.
• Pair trade: long MU / short NVDA for 1-3 months as a relative-value expression of durable contracted supply vs fully priced growth; risk is NVDA reaccelerating on inference demand or another supply constraint shock.
• Buy SNDK as a smaller-cap analog to MU only if customer-financing terms and backlog duration remain intact; otherwise keep it as a watchlist name because the financing/dilution risk is the main falsifier.
• Add AMZN or MSFT only on weakness, not strength, for a 6-12 month hold; the trade is quality compounding, but the first-order risk is margin compression, not demand destruction.
• Avoid chasing small AI infrastructure names like IREN unless there is proof of durable non-hyperscaler demand and low-cost financing; use options or small sizing because the equity story can break on capital-market access even if the product thesis is right.