What If the AI Cost Revolution Arrives Late? Interrogating the Industry's Most Confident Bet
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
The most confident assumption running through today's technology coverage — shared by investors, executives, and analysts from Sam Altman to Jensen Huang to Goldman Sachs — is that AI compute costs will fall dramatically and soon. Altman's projection that infrastructure costs will decline toward $20 billion per gigawatt-scale deployment, from a current figure he cited of roughly $70 billion over five years, is treated in many circles as the foundational premise for the entire AI investment supercycle. It is worth examining what would have to be true for that projection to be wrong, or simply significantly delayed.
The cost-decline thesis rests on three interlocking assumptions: that semiconductor manufacturing will continue to follow something resembling Moore's Law; that power generation and distribution infrastructure can scale at the pace AI deployment demands; and that the dominant training and inference architectures remain stable enough that efficiency improvements compound predictably. Each assumption carries a plausible failure mode. On semiconductors, TSMC's progression to 2-nanometer nodes and below involves extraordinary capital expenditure with diminishing yield improvements per dollar compared to previous node transitions — the cost-per-unit-of-compute may keep falling, but perhaps not at the rate the projections embed.
The power constraint may be the most underappreciated obstacle. Data center energy consumption is now measured in gigawatts, and the US electrical grid was not designed for this growth rate. Interconnection queues for new power plants run to seven or eight years in some markets; new nuclear capacity is a decade-plus proposition. If AI deployment continues accelerating, it may encounter a physical energy ceiling before silicon learning rates generate the cost reductions the models predict. Architecture instability introduces a third uncertainty: the transformer paradigm has dominated AI training for several years, but if a fundamentally different architecture emerges, efficiency gains accumulated in transformer-optimized hardware do not transfer cleanly, effectively resetting the cost-improvement curve.
The analogy to solar energy is instructive. The cost of solar generation fell dramatically, and the prediction was directionally correct — but it took longer than many 2010 projections assumed, and those delays had material consequences for climate policy timing. AI compute costs will probably fall. How fast, and what physical or architectural constraints slow the descent, is where the genuine uncertainty lies. For the companies that have committed to $70-billion-scale infrastructure investments in anticipation of rapid cost declines, a meaningful delay in that trajectory means holding expensive assets that monetize more slowly than their models assumed — a risk that the current investment enthusiasm has not fully priced.