Nvidia's $500 billion compute push leans on a fragile asset comparison

Jensen Huang views GPUs as the new treasury bills. Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR have lined up half a trillion in financing to make "compute an asset class." The argument is that H100s and undefineds produce revenue, remain useful for years, work as substitutes, and offer flexible scaling, making them investable assets like real estate or infrastructure.

The analogy fails regarding fungibility. Real estate remains stationary, while GPUs are only interchangeable if the software stack, interconnect, and power envelope are identical. Replacing an H100 with a undefined requires changes to cluster topology, cooling, and kernel tuning, which is a migration project rather than fungibility.

Calling GPUs "long-lived" is difficult because Nvidia releases new architectures every 12-18 months, moving from Hopper to Blackwell to Rubin. Each generation reduces the previous one's performance-per-watt advantage. A 5-year depreciation schedule for hardware that is competitively dead in two years is a lease with extra steps, not an asset class.

Revenue-generating potential is limited to periods of infinite demand. Rental rates for H100s are softening as supply increases, with spot pricing on GPU clouds dropping 30-40% last quarter. An asset producing declining cash flow because the next generation undercuts it is a melting ice cube, not a bond.

The $500B figure is financing capacity across six firms, not committed capital. Apollo's portion might be structured credit, and BlackRock's could involve infrastructure funds; no single $500B check is being written to Nvidia. This number is marketing rather than a balance sheet entry.

Regarding residual risk, limited partners in these funds pay the power bill when a GPU cluster sits idle because a foundation model lab pauses training. Nvidia is paid upfront, while financiers take on duration risk for hardware that depreciates faster than a Maserati.

Huang's CNBC quote claiming technology chips have become an investable asset class for the first time ignores that semiconductor equipment from ASML and Applied Materials has been securitized for decades. Chips are inventory, not infrastructure.

This strategy acts as a demand guarantee. Nvidia secures massive pre-orders funded by other people's money, financiers receive yield backed by take-or-pay contracts from hyperscalers, and hyperscalers obtain capacity without putting capex on their balance sheets. This works until someone is left with 50,000 H100s renting for $0.80/hr.

Compute capacity can be an asset class, but compute hardware is inventory with better PR.