Twelve banks just lent Nscale nine hundred million dollars to buy depreciating Nvidia chips. A revolving credit facility syndicated by J.P. Morgan, Goldman Sachs, and Morgan Stanley — twelve lenders in total, including RBC Capital Markets and TD Securities — is not an equity investment. It is a line of credit secured by the company’s assets, extended on the assumption that the assets will generate enough revenue to service the debt. The trouble, and it is a trouble the press release papers over, is the mismatch between the lifespan of the collateral and the lifespan of the debt. A data centre is a warehouse that lasts forty years. An Nvidia Rubin processing unit is a chip that will be commercially obsolete in three. Lending nine hundred million dollars against a pool of GPUs is lending against collateral that loses half its value while the bank is still processing the paperwork.
It is true, and the documentation supports it, that Nscale is not a shell company. The U.K.-based startup raised two billion dollars in equity earlier this year at a $14.6 billion valuation. It is backed by Nvidia, Dell Technologies, and Nokia. It already secured $790 million in May for a data centre project just outside Narvik, Norway, which will house more than thirty thousand Rubin units. The equity is real, and the engineering object is real.
The physical reality of the AI data centre is that it is a power plant that also processes information. The thirty thousand Rubin units Nscale plans to install in Narvik will draw an amount of electricity that requires dedicated grid infrastructure, and they will generate heat that requires dedicated cooling apparatus. The $900 million credit facility is paying for the concrete, the copper, the cooling systems, the power substations. The financial structure layered on top of it is what is precarious. The originating banks capture the fees at the point of sale, then distribute the leveraged paper to institutional investors — pension funds, collateralised loan obligation buyers, the downstream machinery of the credit market — who are pricing the chips as permanent fixtures rather than expiring software licences. Nvidia captures the revenue at the point of sale, before the depreciation begins. The risk — that the AI revenue does not materialise, that the compute price collapses, that the $14.6 billion valuation was a function of the bezzle and not the underlying economics — is not absorbed by the abstract structure of the debt. It is absorbed by the institutional investors holding the paper, and by the public utility ratepayers who will be left holding the grid upgrades for a facility that goes dark.
A $14.6 billion valuation for a company whose business model is building warehouses to house other people’s chips is a valuation that requires the artificial-intelligence buildout to continue compounding at its current rate without a corresponding collapse in the price of compute. It requires the interval that John Kenneth Galbraith named the bezzle — the magic period when the confidence trickster has the money but the victim does not yet understand the loss — to last longer than the operational life of the hardware.
Cory Doctorow, comparing the current AI infrastructure buildout to the telecom fibre buildout of the late 1990s, has pointed out that bubbles in physical infrastructure leave a durable residue. Worldcom laid millions of miles of fibre optic cable; the equity went to zero; the debt was restructured; but the fibre stayed in the ground. The AI data centre is the same category of asset. It is a Worldcom-scale buildout of real physical infrastructure, financed by a syndicate of banks — including two Canadian chartered banks in RBC and TD — who know the collateral is expiring but are passing the tail risk down the credit chain.
There is a structural remedy available, if the regulators choose to pick it up. The Office of the Superintendent of Financial Institutions in Canada, alongside the Federal Reserve in the United States, could adjust the capital-requirement risk weights for loans collateralised by rapidly depreciating compute hardware. If a bank originates a facility against GPUs that will be commercially obsolete in thirty-six months, the originator should be required to hold capital against the full depreciation curve, not the nominal value of the silicon. Make the syndication desk hold the tail risk, and the economics of the bezzle will look rather different.
The data centres will be built in Norway, in Europe, in the United States. The thirty thousand Rubin chips will be installed. The power will be drawn. But when the silicon is worth less than the concrete it sits in, the institutional investors will restructure the paper, and the public will be left paying for the grid upgrades of a stranded asset. The work is to be done.