SemiAnalysis Ships AI Cloud TCO Model Across Nvidia, AMD, Intel

There's a new subscription model out from SemiAnalysis trying to answer, in one place, the question every AI cloud operator and their investors keep asking, which is what does it actually cost to buy the accelerators, stand up the servers, keep the lights on, and sell the compute back out. It covers Nvidia, AMD, Intel, and custom accelerators, and it is pitched at both the operators building these clouds and the equity and debt investors trying to underwrite them.

The metrics tracked are the ones you would expect if you were sizing a real deal, including cost of ownership per hour, inference cost per million tokens, training cost per FLOP, and all-in TDP per GPU. The $/hr calculation is built from upfront server capex, system power consumption, colocation and electricity costs, and cost of capital. The model also analyzes how Pipeline Parallel, Tensor Parallel, Expert Parallel, and Data Parallel schemes shift GPU throughput, and it wraps a three-statement financial model on top, with income statements, balance sheets, and cash flow accounting for server depreciation, unearned/prepaid revenue, and borrowings.

Why this matters if you are not building an AI cloud yourself: this is the kind of scaffolding that lets a procurement team signing a multi-year compute contract, or a debt investor pricing a GPU cluster deal, compare accelerators on the same basis instead of vibes. SemiAnalysis says the model also carries detailed install base projections through 2028 and vendor unit shipment estimates through 2034, which pushes it toward planning artifact rather than snapshot.

The honest caveats are worth naming. This is a paid model, so the specific dollar outputs are not visible in the public description, the reader is buying the framework and quarterly updates rather than a screenshot. The public page does not spell out how inputs like colocation pricing, power costs, or accelerator residuals are sourced, or how sensitive the final $/hr number is to those assumptions. SemiAnalysis's own comparative work has suggested that AMD's MI300X and MI325X generally present lower total hourly costs than Nvidia's H100 and H200, but that perf/$ flips by workload, so the answer any given buyer gets is a function of what they actually run.

For operators, investors, and large compute buyers, the useful move is treating this as one input against a bottoms-up model, not an oracle. As more capital flows into GPU clouds, the operators with defensible TCO math get funded and the ones without do not.