Marc Ganzi sold his data center business to SoftBank for $4 billion and immediately began comparing the current AI infrastructure boom to the late-1990s fiber bubble. The timing is precise. The warning is public. The message from a founder who just secured a ten-figure exit is that the capital flooding into power-hungry compute facilities has decoupled from demand visibility.

Ganzi built a career financing the physical layer of connectivity—towers, fiber, colocation. He knows what speculative infrastructure looks like when it moves faster than the applications it is meant to serve. In 1999, carriers laid 80 million miles of fiber across the United States, betting on traffic growth that took a decade to materialize. Today, hyperscalers and sovereign funds are pre-leasing data center capacity in markets where power substations have not yet been upgraded and where cooling systems are still on the drawing board. Ganzi is noting the pattern: capital arrives before the use case proves out, and the gap between buildout and utilization becomes the risk.

This matters because SoftBank is now the entity holding the asset. The firm has a well-documented history of buying at peak narrative—WeWork at $47 billion, Arm before the mobile boom plateaued, and now data centers as every pension fund and sovereign vehicle is underwriting AI as if large language models are a natural law. Ganzi's exit is not just a liquidity event. It is a re-rating signal from someone who knows infrastructure cycles and who chose this moment to convert equity into cash. The fact that he is publicly naming the risk within days of closing the transaction suggests he sees a gap between current pricing and the pace at which AI workloads are actually being deployed at scale.

What allocators should watch is the lag between data center commissioning and tenant occupancy. If power infrastructure cannot keep pace with facility construction, or if model training shifts toward inference at the edge rather than centralized compute, then the $4 billion SoftBank just deployed becomes a forward indicator of how much capital is chasing a thesis that has not yet been tested by a full economic cycle. Ganzi's comparison to 1999 is not casual. Fiber assets eventually became valuable, but only after a decade of bankruptcies, debt restructuring, and capacity sitting dark. The question for 2025 is whether AI demand is durable enough to absorb the supply being financed today, or whether this is another case of infrastructure arriving five years early.

Operators should track power grid upgrade timelines in Northern Virginia, Phoenix, and the Dallas-Fort Worth corridor, where data center construction is outpacing substation capacity. Watch for lease-up rates on newly commissioned facilities over the next eight to twelve months. If occupancy lags, or if hyperscalers begin renegotiating pre-lease terms, that will confirm Ganzi's thesis. If utilization accelerates and power constraints become the bottleneck rather than demand, then the $4 billion will look like the floor, not the ceiling.

SoftBank now owns an asset that a career infrastructure investor chose to sell at what he is calling a late-1990s moment.