The market may be pricing AI infrastructure stocks with a dangerous mix of short-term optimism and long-term skepticism. A new analysis from Michael Hillaert suggests that while analyst forecasts for Nvidia and TSMC through 2028 are already aggressive, the actual duration of the earnings opportunity could extend far beyond what current valuations imply. For builders and infrastructure investors, the key insight is that more AI usage does not automatically translate to more chip spending, but the eventual payoff from autonomous agents could stabilize demand in ways traditional cloud metrics miss.

The Productivity J-Curve and Telecom Parallels

Hillaert draws on the 'Productivity J-Curve' concept from economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson to explain the current disconnect. Early AI adoption often hurts profitability as companies retrain staff and rebuild workflows, creating a lag between hardware spending and revenue generation. This mirrors the late-1990s telecom boom, where infrastructure was built for demand that arrived slower than expected, even though the internet itself eventually grew massively. The warning for investors is that capacity built for rapid adoption can sit idle if the productivity gains take longer to materialize than the depreciation schedule allows.

Agents Change the Compute Equation

The long-term bull case for Nvidia and TSMC hinges on the shift from chatbots to autonomous agents. Microsoft Research data from June 2026 shows that GitHub Copilot agent sessions averaged 59 model calls per session, indicating that a single user task can trigger massive, repeated inference workloads. Salesforce reported 3.2 billion Agentic Work Units in the quarter ending July 2026, with Agentforce ARR growing over 240% year-on-year. This suggests that future demand will be driven by delegated work volume rather than just user count, potentially decoupling chip sales from the volatility of individual user adoption rates.

Margin Risks and Geopolitical Realities

Despite the potential for sustained growth, the near-term risks are severe. Nvidia disclosed $279 billion in supply and capacity commitments in July 2026, tying its fate closely to the pace of the buildout. A ten-percentage-point drop in gross margin could cut operating profit by 15% at current revenue levels. Meanwhile, TSMC faces the challenge of funding massive capital expenditures while managing geopolitical exposure in Taiwan. Although TSMC holds a 72.5% share of the top-ten foundry market and benefits from competition among chip designers, its reliance on a small number of large customers means an industry-wide spending pause would still hit hard.

Key Takeaways

  • Falling inference costs (down ~47% per quarter since 2023) do not guarantee proportional revenue growth due to limited price elasticity.
  • Autonomous agents generate significantly more compute per task than chatbots, potentially stabilizing long-term hardware demand.
  • Nvidia and TSMC could deliver attractive returns by 2035 even if 2028 earnings miss consensus, provided growth resumes afterward.
  • Geopolitical risks and massive capital commitments remain the primary threats to TSMC’s profitability and Nvidia’s supply chain stability.

The Bottom Line

The market is currently too pessimistic about the long-term duration of AI infrastructure demand, but too optimistic about the immediate profitability of current deployments. For infrastructure builders, the real opportunity lies not in the current chatbot wave, but in the inevitable, compute-hungry shift toward autonomous agents that will outlast the current valuation cycle.