Mark Cuban just dropped a take that is hard to ignore. AI data centers are getting built at a pace that looks almost reckless, and he thinks plenty of them could end up as pickleball courts.
The logic is straightforward. Hyperscalers are pouring hundreds of billions into capacity because they correctly expect AI demand to keep rising. Cuban does not argue with that part. What he questions is the assumption that every facility being planned today will still be needed once efficiency improves. Better chips, smarter models, and lower power requirements have a way of turning today’s must-have infrastructure into tomorrow’s excess square footage. He has seen this movie before with fiber in the late 90s. Plenty of dark fiber sat unused for years. The same pattern could play out with data centers if the price-performance curve moves faster than the buildout.
That brings us to a quieter problem. AI itself may already be showing signs of degradation in how people interact with it. When the response is polished and immediate, the temptation is to stop digging. Understanding gets outsourced. The tool becomes a substitute for thought rather than an amplifier of it.
Meanwhile Jensen Huang has been clear about who he thinks is positioned to win the AI race. Not OpenAI. Not Anthropic. Not Google. Elon.
The road fleet is the largest real-world data collection system on the planet. And Elon controls three of the verticals Jensen calls the most important: foundation models through xAI, autonomous vehicles through Tesla, and humanoid robotics through Optimus. Collecting world data is expensive. Most labs rent compute and buy datasets. Musk built both. That structural advantage is hard to replicate.
Which makes the latest partnership feel almost inevitable. SpaceX is teaming up with NVIDIA to put AI data centers in orbit.
Big news here. Efficiency curves will eventually test how much of that capacity is truly needed. At the same time the intellectual habits around AI are shifting, sometimes toward shallower engagement. And the companies best positioned to navigate both the hardware and data challenges are the ones treating the problem as a full-stack systems challenge rather than a pure model race.
