The honest answer is: both things are true simultaneously, and the hard question is the ratio between them.

The justified core is real. Physical data center capacity in the top North American markets — Northern Virginia, Silicon Valley, Chicago — is running near record-low vacancy. Completed megawatts are being absorbed quickly, and 92% of capacity currently under construction is reportedly pre-leased. AI-related capex now represents roughly 5% of U.S. GDP and is growing at roughly 10% per year, a pace comparable to the late-1990s tech infrastructure boom. Hyperscalers are not buying capacity speculatively in the abstract; they are responding to genuine, contracted demand from cloud customers and their own model-training pipelines. KKR's analysis, which is admittedly bullish but grounded in lease data, argues that long-term demand should justify much of the current data center buildout.

The speculative overhang is also real. The pipeline of announced projects is a different picture from the pipeline of projects that will actually get built and filled. There is credible evidence of double-counting future demand across competing announcements, and some of the financing assumptions embedded in planned projects will not survive contact with actual utilization rates. Enterprise adoption — the layer below hyperscaler self-consumption — is softening relative to the most optimistic projections. The bullwhip dynamic is a genuine risk: upstream suppliers (GPU fabs, data center developers, power infrastructure) are scaling to meet orders that themselves reflect precautionary over-ordering by hyperscalers, which can produce a sharp demand correction even without a fundamental collapse in AI usage.

Where the dotcom analogy breaks down. The standard reassurance — 'even if it's a bubble, the infrastructure will be useful for decades, just like the fiber-optic cables' — is weaker here than it sounds. Fiber-optic cable laid in 2001 was still carrying traffic in 2020. NVIDIA H100s purchased in 2024 will be largely obsolete by 2027–2028, given the pace of GPU generational improvement. The residual value of overbuilt AI compute infrastructure is much lower than the residual value of overbuilt physical network infrastructure. This asymmetry matters for how much comfort the historical analogy actually provides.

The crux that remains unresolved. The central uncertainty is not whether there is a speculative layer — there clearly is — but how large it is relative to the justified base, and whether the depreciation cycle hits before revenue from AI applications scales enough to validate the investment. That depends on: (1) how quickly enterprise AI adoption moves from pilot to production at scale, (2) whether inference demand grows fast enough to absorb training-era GPU capacity as training runs plateau, and (3) whether power infrastructure constraints (which are physical and slow to build) end up being the actual binding limit rather than compute. None of these questions have settled answers yet.

The bottom line. The buildout is not irrational speculation — there is genuine demand underneath it. But the announced pipeline meaningfully exceeds what demonstrated demand can absorb, and the hardware's fast depreciation cycle removes the safety net that made past infrastructure overbuild cycles ultimately benign. The risk is not a dotcom-style total write-off; it is a period of significant overcapacity in compute, margin compression across the AI infrastructure stack, and a shakeout among the more speculative project developers — while the core hyperscaler positions remain defensible.