"Every year $5 trillion, or 800 trillion yen, you might think that’s a lie, but I am confident that’s what it will cost." With those words, spoken Tuesday at SoftBank’s annual corporate conference in Tokyo, Masayoshi Son threw down a number that shifts the entire debate on artificial intelligence. He didn’t explain how he arrived at that figure, but the message is clear: anyone who doesn’t invest on that scale will be left behind.

The projection comes at a time when doubts about AI hype are multiplying. Since 2023, every quarter has produced fresh waves of enthusiasm followed by inevitable expectation corrections. Son dismisses the idea of a bubble with a blunt adjective: absurd. This is not a neutral stance. SoftBank is a major shareholder in Arm, the backbone of much of the hardware for inference and training, and has placed billions on companies like NVIDIA. His thesis, then, is also an existential bet for his financial empire.

But the real question is structural: $5 trillion a year means AI would become the highest infrastructure spend in history, surpassing global defense and healthcare combined. If that trajectory materializes, only a handful of players – hyperscalers, sovereign wealth funds, perhaps a few nation-states – will be able to afford frontier model training. Everyone else will have to rent capacity, deepening dependence on cloud providers concentrated in two or three jurisdictions.

This scenario directly challenges those evaluating on-premise or self-hosted deployments today. The prevailing narrative says that the TCO of local infrastructure – GPUs, VRAM, power consumption – can’t compete with hourly rental. But Son’s projection adds a second-order element: spending concentration will accelerate the race toward ever-larger models that are unlikely to run on enterprise clusters without a step-change in scale. It’s not just a matter of unit costs; it’s a problem of access to the technology itself. If models require hundreds of billions of dollars worth of hardware, on-premise risks becoming synonymous with lower service tiers, not greater control.

Paradoxically, however, Son’s move could strengthen the case for local deployment. Data sovereignty, GDPR compliance, and latency won’t vanish because aggregate spending rises. In fact, the more power concentrates in a few centralized data centers, the more regulated enterprises and governments will find strategic justifications to invest in on-premise hardware, even at higher unit costs. It’s a tension that AI-RADAR has long tracked: not about demonizing the cloud, but about understanding where the breaking point lies between economic efficiency and real control.

Son’s statement is less a prediction than a signal. It tells the market: prepare for an unprecedented capital requirement. And, implicitly, it indicates that the future of AI will be written by those with the means to build the infrastructure, not just by those who develop algorithms. Those operating in the on-premise space should read it as a warning, but also as a challenge to make self-hosting credible in a world where scale appears to dictate every rule.