The document puts the deal at $7 billion for 100,000 offshore AI chips with Oracle. That figure does not prove a per-hour cost, a chip SKU, a delivery date, or a measured H100 comparison. The headline’s 43 percent estimate is a rental-rate claim, not a benchmark. It also leaves open whether the chips are H100s, or whether the rental comparison uses the same workload.

Die Brief reads the constraint as the buyer’s scale, not the chip’s performance. A 100,000-unit purchase shifts risk from unit price to contract duration. The provider’s spare capacity is untested for the buyer’s side of the deal. The fab gets no yield signal.

What remains unmeasured is the actual hourly price, the accelerator model, the data-center location, and the H100 baseline used for the 43 percent figure. Without those, the deal is a capacity purchase, not a performance result. The baseline stays open, though.

Die Brief reads the constraint as the buyer’s scale, not the chip’s performance. A 100,000-unit purchase shifts risk from unit price to contract duration, while the fab gets no yield signal.

The document does not state the accelerator model, the data-center location, the hourly price, or the H100 baseline used for the 43 percent estimate.

The deal size is a capacity purchase: 100,000 units at $7 billion sets a block price, but it does not establish per-hour cost, SKU, node, or H100 parity.