Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom

Circular Financing and Nvidia–Neocloud Deals

• Some argue Nvidia’s ~$2B equity in CoreWeave vs. $35B CapEx (2026) shows its role is small; others see a broader pattern where vendors fund customers who then borrow more to buy the vendors’ hardware.
• Concern: this can blur the line between genuine external demand and revenue inflated by vendor-backed financing.
• Counter: Supplier equity in customers and “invest to grow your own market” is a long-standing strategy, not inherently problematic. Scale and opacity are what worry people.

Backstop Agreements and Accounting Questions

• Nvidia’s obligation to buy unused CoreWeave capacity (initially $6.3B, potentially more) is seen by some as de facto consignment or demand padding.
• Others state GAAP treatment is straightforward: Nvidia books real sales; the backstop is a separate purchase commitment and later operating expense if triggered.
• Debate shifts from “is this legal accounting?” to “does current disclosure/GAAP adequately surface the economic risk?”

Profitability, Demand, and the AI Bull Case

• Bull view: anything done by humans can be done by AI; AI is likened to printing press/electricity; startups report real productivity gains and growing AI spend.
• Bear view: valuations price in decades of profit not yet visible; AI may become a low-margin commodity like electricity; only “shovel sellers” (GPU vendors, infra) are clearly profiting.
• Some claim massive ARR growth for model vendors; others ask “where is the value?” and call stats exaggerated or unclear.

Bubble, Systemic Risk, and Historical Analogies

• Comparisons drawn to dot-com, 1929, and crypto: huge capital flows, leverage, and potential for a sharp correction.
• Fears that an AI bust would hit indices, ETFs, pensions, and possibly trigger a broader debt/credit crisis; others argue losses would be mostly contained to willing investors.
• A minority insists “there is no AI bubble” and calls it the biggest gold rush ever; critics note past gold rushes enriched a few and harmed many.

Infrastructure, Hardware Economics, and Overbuild Risk

• Discussion on: datacenter build constraints (power, permits) possibly limiting overcapacity; utilization and price decay of older GPUs (A100/H100 vs B200+); and whether new, more efficient non-Nvidia chips could undercut current investments.
• Unclear how sustainable pricing and utilization will be for older hardware over multi-year payback periods.