Nvidia’s $110B AI Bet: Is Circular Financing a Repeat of the Telecom Bubble?

Nvidia’s $110 billion commitment to OpenAI in September 2025 has reignited comparisons to the telecom bubble of the late 1990s, particularly due to the rise of vendor financing—where suppliers lend money to customers to purchase their products. While the parallels are striking, key differences suggest this cycle may not end in the same collapse. In the late 1990s, Lucent Technologies, the dominant telecom equipment maker, extended $8.1 billion in vendor financing to cash-strapped carriers. Alongside Nortel and Cisco, Lucent fueled a massive buildout of fiber-optic networks. When demand evaporated, the bubble burst. Lucent’s revenue plummeted 69% from $37.9 billion in 1999 to $11.8 billion by 2002, and it eventually merged with Alcatel in 2006. Behind the scenes, accounting fraud—channel stuffing, side agreements, and reserve manipulation—masked deteriorating fundamentals, leading to $1.1 billion in fraudulent revenue and $470 million in inflated income. The fallout included $3.5 billion in bad debt write-offs and the collapse of numerous competitive local exchange carriers (CLECs), including WinStar, which Lucent refused to bail out. Fast forward to 2025. Nvidia’s $110 billion in direct investments—most notably the $100 billion commitment to OpenAI—far exceeds Lucent’s peak financing. OpenAI’s first $10 billion tranche was priced at a $500 billion valuation, with subsequent tranches tied to deployment milestones. Payments are structured as lease arrangements, meaning most capital flows back to Nvidia. Beyond OpenAI, Nvidia holds stakes in CoreWeave ($3B), other AI startups ($3.7B via NVentures), and has enabled over $15 billion in GPU-backed debt, including $10.45 billion at CoreWeave and $500 million at Lambda Labs. Despite the scale, the context differs fundamentally. Lucent’s customers were speculative telecom firms burning cash with no clear path to profitability. Today’s buyers—Microsoft, Google, Amazon, Meta—are highly profitable hyperscalers generating $451 billion in operating cash flow in 2024. OpenAI, while unprofitable with a $4.7 billion loss in H1 2025, is backed by deep-pocketed investors and has a clear mission. AI is already deployed: 40% of U.S. workers used AI at work by September 2025, up from 20% in 2023. Studies show AI users see up to 40% performance gains and wage growth twice as fast in AI-exposed industries. Still, risks remain. Nvidia’s top two customers account for 39% of revenue, and the top four for 46%—nearly double Lucent’s concentration. The rise of GPU-backed debt at 14% interest rates, triple investment-grade levels, hinges on the assumption that GPUs hold value for 4–6 years. Yet real-world data from Google and Meta shows 60–70% utilization and 9% annual failure rates, suggesting a 1–3 year useful life. Amazon reversed its depreciation policy in 2025, reducing server lifespan from 6 to 5 years—its first pullback in decades. Hyperscalers are also using Special Purpose Vehicles (SPVs) to build data centers, keeping debt off their balance sheets. This allows massive capital expenditures—projected at $300–400 billion in 2025—without triggering credit downgrades. But thin equity layers (10–30%) mean losses hit equity holders first if utilization falls or GPUs depreciate faster than expected. Another risk: custom silicon. Microsoft is moving toward “mainly Microsoft silicon” with Maia accelerators. Google uses TPUs, Amazon has Trainium and Inferentia, and Meta is developing MTIA. If these shift from experimental to production use, Nvidia’s vendor financing becomes exposure to competitors building alternatives. In short, while vendor financing has grown to unprecedented levels, the underlying demand is real and backed by revenue-generating businesses. Unlike the telecom bubble, where demand was speculative and customers were cash-burning, today’s AI buildout is driven by actual usage, productivity gains, and enterprise demand. The risk isn’t the absence of demand—it’s whether the current pace of spending can be sustained if monetization lags or technology shifts. For now, the merry-go-round has paying riders.