The most important AI story of the week wasn’t a model release or a funding round. It was a set of satellite photos. Aerial and drone imagery of US data center sites — the physical substrate of the entire generative AI economy — show construction running months behind schedule across multiple hyperscaler builds. The narrative of infinite, on-demand compute just collided with rebar, transformers, and zoning boards.
The Ars Technica analysis — built on commercial satellite passes and drone overflights — documents shells without roofs, cooling infrastructure not yet on site, and substations that haven’t been energized at facilities scheduled to be operational by mid-2026. These aren’t speculative builds. These are the named projects on Microsoft, Meta, Amazon, and Oracle capex slides. The ones that justified the $300B+ in AI infrastructure spend telegraphed to investors over the last 18 months.
The bottleneck is no longer chips. Nvidia is shipping. The bottleneck is everything around the chips: power transmission, water rights, cooling systems, switchgear, and the unsexy logistics of pouring concrete in jurisdictions that didn’t plan for gigawatt-scale loads.
Why this breaks the unit economics
Hyperscaler AI economics depend on a brutal assumption: depreciate $40B of GPUs over 5-6 years against revenue that materializes the moment the building turns on. Every quarter of construction delay extends the depreciation clock against zero revenue while the GPUs themselves continue their march toward obsolescence. An H100 that ships in Q2 2026 and sits in a warehouse until Q4 2026 because the data center isn’t ready is a $30,000 asset losing roughly 15% of its lifetime monetization window before it ever inferences a single token.
This is the moment the “scale at all costs” thesis gets quietly repriced. The capex was justified to Wall Street on the premise that compute would be turned into revenue at industrial speed. Concrete doesn’t move at industrial speed.
Who wins, who loses
Winners: Anyone who already owns operational compute capacity. CoreWeave, Lambda, and the colocation providers with energized facilities suddenly hold scarce inventory in a constrained market. Expect spot pricing for H100/H200 capacity to firm up through 2026 rather than collapse as bears predicted. Inference-optimized startups built on existing capacity get a quiet tailwind. The model labs that locked in long-term capacity contracts in 2024 look prescient.
Losers: The hyperscalers carrying the largest unbuilt forward commitments. Their marginstory for FY2026 was predicated on bringing capacity online to absorb GPU depreciation. If 20% of planned capacity slips two quarters, that’s a measurable EPS event. Watch the next earnings cycle for the language shift from “on track” to “phasing” — the universal corporate euphemism for slipping.
The deeper structural lesson: the AI industry has spent two years optimizing the digital layer of the stack — model architecture, distillation, MoE routing, KV caching — while the physical layer was treated as a solved problem you simply order more of. It isn’t. Power is a planning-horizon problem measured in years, not procurement cycles. The companies that figured this out early and bought legacy industrial sites with existing interconnect agreements are about to look like geniuses. Everyone else is about to discover that you cannot ship a data center with a software update.
FourWeekMBA AI Business Intelligence — strategic analysis of the moves that matter.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
