Capex is a promise; revenue is the receipt — and inference unit cost is the bridge between them. Amazon's grant US10678508B2 (“Accelerated quantized multiply-and-add operations,” issued 2020-06-09) sits squarely on that bridge. The invention, assigned to Amazon Technologies, Inc. and classified in CPC G06N 3/063, accelerates the multiply-and-add operations that dominate neural-network workloads by performing them in quantized, lower-precision form.
The way this matters financially is direct. Every matrix multiply a model runs costs power and chip time. Doing the arithmetic in 8-bit instead of 32-bit precision — quantization — cuts both, often with negligible accuracy loss. When you rent that compute out by the hour, as AWS does, the gap between full-precision and quantized cost is margin.
Amazon's filings describe its custom-silicon ambitions in general terms; the company has publicly tied its Inferentia and Trainium lines to lowering machine-learning cost, and its annual reports flag infrastructure investment as a major use of capital. The patent is the specific, dated mechanism under that general story — inventors including Ron Diamant and Dana Vantrease are names associated with Amazon's accelerator work.
For the capex analyst the lesson is about denominators. A hyperscaler's AI buildout is usually argued in terms of how many chips it bought. The quieter question is how much useful inference each chip delivers per dollar, because that is what eventually has to pay the buildout back. Quantization IP like this is one of the levers on that ratio.
The disclosure discipline: the patent proves invention and ownership, not a revenue figure. No Amazon filing breaks out earnings attributable to this technique, and we don't imply one. What it documents is that the cost engineering behind AWS inference pricing was patented and dated in 2020, well before the current demand cycle.