A $3.7 billion write-down by SoftBank (TYO: 9984) on power generation equipment is raising uncomfortable questions about whether AI infrastructure demand is as robust as widely believed.

Dan Niles, founder and portfolio manager at Niles Investment Management, pointed to the SoftBank loss as a telling indicator during a conversation with Deirdre Bosa, asking pointedly: “Why did Softbank if things are so wonderful, take a $3.7 billion write down on power generation equipment?”

Niles argued that if genuine demand existed, SoftBank could simply have offloaded the equipment on the secondary market rather than absorbing the loss.

The warning lands at a particularly sensitive moment for neoclouds, companies that have relied on Nvidia (NASDAQ: NVDA) not only for GPU allocations but also for financing arrangements that critics call circular.

Jensen Huang’s departure from Nvidia removes what Niles described as the single most powerful internal advocate for open-source AI and the generous financing structures that neoclouds depended upon to grow.

With Huang gone, attention shifts to Nvidia’s CFO Colette, whom Niles regards as an excellent financial steward but one far less sympathetic to the deal structures that fueled neocloud expansion.

Niles stated that Colette is going to “even more closely examine all of those deals because she wasn’t as big of a proponent for that and all the circular financing,” suggesting a meaningful tightening of terms ahead.

He characterized the shift bluntly, describing what he called the “Bank of Nvidia” as heading toward a temporary closure, with consequences that will ripple across the neocloud ecosystem.

While CoreWeave and Nebius (NASDAQ: NBIS) have both pointed to rising customer prepayments as evidence of greater business maturity, Niles was skeptical that prepayments offer the protection investors assume.

Nebius itself was forced to cut its forecast, even as the company reported covering roughly 100% of its capital expenditures through prepayments in the third quarter, illustrating the limits of that metric.

To explain his skepticism, Niles drew a direct parallel to the semiconductor industry’s post-COVID reckoning, when chip shortages prompted widespread long-term supply agreements that quickly became liabilities once consumer demand normalized.

Microchip Technology served as his sharpest example, with revenues declining more than 40% year-over-year for four consecutive quarters as customers refused to honor commitments for products they no longer needed.

The fallout at Microchip was severe enough that its CEO was removed, with a predecessor returning specifically to renegotiate and abandon those agreements entirely.

Niles suggested the AI infrastructure cycle is following a structurally similar path, where today’s prepayment commitments could prove just as fragile if end demand softens or competitive dynamics shift.

The winners in the new landscape, he argued, are closed-source AI developers such as OpenAI and Anthropic, whose prospects for regulatory advantage improve as Huang’s open-source advocacy disappears from Nvidia’s leadership.