A recent take argued the entire AI trade thesis is wrong and markets aren't repricing
Microsoft has racks of H100s sitting unused because they literally cannot power them; the necessary power infrastructure doesn't exist
That implies analyst models that value companies by chip purchases and GPU counts are focused on the wrong constraint—the bottleneck has moved while markets still act like it's 2023
This rewrites the capex equation: if $50B of GPUs are bought and half remain unpowered for 18 months, the ROI timeline collapses; each idle GPU depreciates versus future releases while incurring data center construction costs with no revenue
The real winners are those who secured power purchase agreements or built generation capacity 3–4 years ago; GPUs can be ordered in months, but adding 500 megawatts takes years of permitting, construction, and grid hookup
With GPU release cycles compressed to roughly annual cadence, a chip bought today may only hold performance leadership for 12–18 months; if deployment is delayed, the asset is already depreciating before it earns anything
Faster energization of capacity is now a direct multiplier on chip purchase ROI, so vertically integrated players with their own power and real estate capture disproportionate value
This changes the competitive moat structure: model quality and algorithm gains used to be the primary edge, but physical infrastructure and energy access are now the durable constraint.
A better model can be developed in months; a powered data center cannot be built that quickly, creating long-lasting separation between winners and losers
This changes the competitive moat structure: model quality and algorithm gains used to be the primary edge, but physical infrastructure and energy access are now the durable constraint.
A better model can be developed in months; a powered data center cannot be built that quickly, creating long-lasting separation between winners and losers