OpenAI's Sam Altman Admits the AI Boom Is Running Late: “We've Not Had the iPhone Moment”

Sam Altman says AI adoption is still at the “palm pilot stage” (no “iPhone moment”) and cautions timelines are slipping due to habit-driven inertia. Countering that, NVIDIA’s backdrop shows “demand has gone parabolic,” with FY27 Q1 revenue of $81.6B (+85% YoY) and Q2 guidance at $91.0B (±2%), while management cites compute as the core bottleneck. The article also highlights that capacity constraints trace to the foundry level (TSMC raised its 2026 capex to $60–$64B and said packaging is “so tight” it limits customer growth), keeping near-term economics favorable to capacity sellers even if application adoption lags.

Analysis

The market is likely to misread this as a simple “AI adoption is slower” headline, but the investable implication is narrower: adoption can lag while infrastructure demand remains prepaid and capacity-constrained. That favors the suppliers of scarce compute over the software layer, because frontier labs and hyperscalers still have to secure supply ahead of usage realization. Near term, the biggest beneficiary is the entity with the tightest bottleneck and the longest queue, which argues for continued pricing power in the foundry/packaging chain more than a straight-line extrapolation of GPU demand.

The second-order risk is valuation dispersion. If end-user monetization stays delayed for another 1-3 quarters, the market will start discounting application-side revenue more aggressively than infrastructure cash flows, and that can hit NVDA’s multiple before it hits its revenue base. By contrast, the more durable scarcity rent sits with TSM because advanced-node capacity and packaging are harder to replicate than design wins; that makes TSM the cleaner way to express the AI capex cycle over 6-18 months.

The contrarian view is that consensus is overestimating how quickly demand will compound from “usage” and underestimating how long the pre-build phase can last. What would break the thesis is any evidence that training/inference efficiency is rising faster than model appetite, or that hyperscaler capex guidance rolls over before the next earnings cycle. If that happens, infrastructure multiples compress first, and the market starts to question whether the current AI buildout is a one-time inventory cycle rather than a multi-year capacity supercycle.

Key Decisions for Investors

• Prefer long TSM vs. long NVDA on a 1-3 month horizon: TSM has the more persistent scarcity rent from advanced nodes and packaging, while NVDA is more exposed to any air pocket in sentiment or multiple compression. Use pullbacks in TSM to build; trim NVDA strength into any post-news squeeze.
• If already long NVDA, hedge with short-dated put spreads into the next earnings window rather than exiting outright. The risk is not revenue collapse, but a de-rating if the market begins to price slower enterprise conversion.
• Monitor hyperscaler capex and any commentary on inference efficiency as the key falsifier over the next 1-2 quarters. A downward revision there would be the first sign that the pre-buy cycle is peaking.
• Keep TSM as the cleaner structural long for 6-18 months, but size it below consensus positioning until packaging capacity and 2nm ramp data confirm the supply bottleneck is still binding.