Google's $45B for Anthropic Follows the Copper-Market Playbook

Google put $45 billion behind Anthropic's chip deal in 2025. Nvidia helped finance CoreWeave, the GPU cloud that buys Nvidia's chips. Each arrangement looks like a circular financing: the same handful of firms sitting on every side of the table as investor, customer, and supplier.

The Emerging Trajectories essay pushing back on the bubble take argues that the history of how real commodity markets get built is mostly the history of circular deals. Copper, natural gas, electricity, and aluminium all required buyers, sellers, lenders, and infrastructure operators to commit to each other before the underlying demand existed. A copper mine cannot be financed without a long-dated offtake (a multi-year purchase commitment) from a smelter. A natural gas pipeline cannot be built without a utility agreeing to take the gas for twenty years. The same handful of banks and operators often sit on every side of those contracts, and the structure looks incestuous only until the project produces. The pattern is the financing.

Google's backstop of Anthropic's $45 billion chip agreement is one of the larger recent deals. Anthropic needs compute. It needs chips to run that compute. Google is both a chip financier and a compute provider. The structure is the same one a bank uses when it lends to a mine and takes a copper offtake as collateral: a long-dated commitment that lets the underlying project exist. Whether the deal is healthy or ugly depends on what backs the contract and what happens if demand is not there.

The TeraWulf arrangement with Fluidstack is a ten-year, 200+ MW hosting commitment. TeraWulf announced the contracts in 2025, a power-purchase-style agreement long enough to underwrite a new data-center build. Long-dated offtakes are how the electricity market works. The risk is whether the offtake buyer can actually pay, or whether power prices move against the host.

The Hyperion financing, which one analyst newsletter pegs at roughly $80 billion rather than the $50 billion figure in earlier coverage, is the larger variant. Global Data Center Hub argues the structure is a signal that AI compute is shifting from product-purchase to commodity-purchase, where customers lock in long-dated capacity and finance providers take the demand risk. Nvidia's role in backstopping GPU debt at CoreWeave, covered by SemiAnalysis, is the same template at smaller scale.

Community reaction on Hacker News was largely skeptical, with most discussion drifting to model-training data and copyright rather than the financing structure. The 1999 comparison the skeptics reach for collapses the moment you look at contract length. Most of the recent deals run for a decade or more. That is a long way from a quarter-to-quarter revenue model.

The deals worth flagging are the ones that break the commodity pattern. A circular structure with no long-dated offtake is just a capital loop. A backstop from a firm that cannot absorb the underlying risk is a derivative in disguise. A deal where the same company is the only buyer, the only supplier, and the only lender is a house of cards dressed in contract language. The 1999 comparison applies to those structures, and they deserve scrutiny.

For the next AI compute headline, the question is what kind of circular structure each deal is. If the contract is long-dated, the offtake is real, and the backstop is from a party that can absorb the loss, the deal is closer to a copper-mine financing than to Pets.com. If any of those is missing, the bubble critique lands. The next test will be the first circular deal where one of those three legs cracks under stress: an offtake buyer default, an absorbing party calling a backstop, or a contract that turns out to be shorter than advertised.