AI Cloud TCO Model
The ownership economics of AI compute: what a GPU-hour actually costs, what it rents for, and whether the business case clears. Per-accelerator benchmarks, rental price scenarios, and a full cluster finance suite for operators, investors, and compute buyers.
- Cost resolution
- $/GPU‑hr
- Financial model
- 3-statement
- Update cadence
- Monthly
The business case for AI compute, priced line by line.
The SemiAnalysis AI Cloud TCO Model examines the ownership economics of AI clouds that buy accelerators and sell bare metal or GPU compute. It builds the all-in cost of a GPU-hour from server capex, power, colocation, and cost of capital, sets that against current and scenario rental pricing, and carries the result through a full financial model of the cluster.
Underneath sits a per-accelerator benchmark set spanning Nvidia, AMD, and custom ASICs: training and inference throughput, power draw, cost of ownership per hour, inference cost per million tokens, and training cost per FLOP, with install base projected through 2028 and shipments by vendor through 2034.
Who uses it, and what it decides.
The buyers of this model, and the calls they make with it.
- AI cloud & neocloud operators
- The business case for building and running an AI cloud: which accelerators to buy, what a competitive rental price is, and how pricing, utilization, and financing move the returns.
- Equity & debt investors
- Underwriting AI cloud operators and GPU-backed lending: residual value, cash generation, IRR, and how the balance sheet holds up under different rental price paths.
- Compute buyers
- Benchmarking and planning long-term compute procurement: what the compute should cost, when to lock a term contract, and how the cost curve moves as new silicon lands.
Four modules, one business case.
Each module is built from primary research and reconciled against the others, so benchmarks, cost stacks, and financials stay consistent.
Accelerator economics & install base
Per-accelerator benchmarks for the silicon that matters: training and inference throughput, GPU TDP and all-in system power, cost of ownership per hour, inference cost per million tokens, and training cost per FLOP, across four vendor families:
- Nvidia
- AMD
- TPU
- Trainium
Sits on a detailed install base by GPU projected through 2028 and total unit shipments by major vendor through 2034, with market-wide training and inference throughput and the frontier cost benchmarks alongside.
GPU TCO & rental pricing
The all-in cost of owning GPU servers in dollars per hour, built from the components an operator actually pays for:
- Server capex including GPUs, DRAM, NAND
- Networking capex
- Colocation cost
- Power cost
- Cost of capital
Set against an overview of current market GPU rental prices and their variation, plus forward rental price scenarios built from supply-demand analysis and the cost curve implied by upcoming accelerator generations.
Cluster finance suite
The full financial picture of a GPU cluster, from first capex to residual value:
NPV & residual value
Net present value and accelerator residual value from future earnings and cash generation
Cash flow & returns
Cumulative project and equity cash flow, project and equity IRR, ROA, ROIC, ROE, EBIT, and EBITDA
Three-statement model
Income statement, balance sheet, and cash flow, with depreciation, unearned revenue, borrowings, and the other items that matter
Every key assumption is adjustable: capital structure and debt mix, cash versus PIK interest, depreciation periods, customer fixed-price terms and prepays, physical GPU operating lifetime, repairs, and tax.
Price to performance ratios
The amount of compute performance per dollar paid:
- $/PFLOPs
- $/MTok
- $/TB/s
Extends into training and inference economics. Informs purchase and ownership decisions across different workloads, measured from marketed FLOPs to measured token throughputs.
Where a GPU-hour goes.
The model builds the all-in cost of an accelerator hour line by line, then prices the rental scenarios on top of it.
Pairs with our public pricing trackers
How the model is built.
Every dollar is priced where it is spent, then tested against the market.
-
Benchmark the silicon
Each accelerator is measured on training and inference throughput and power, with market rental pricing tracked alongside.
-
Build the cost stack
Capex, power, colocation, and cost of capital roll up into the all-in dollars per hour, per accelerator and per buyer type.
-
Run the business case
Rental price scenarios feed the finance suite: NPV, residual value, returns, and the three-statement model.
Research that ships with the model.
Model subscribers receive the update notes, webinars, and analysis published against each release. A sample of recent coverage:
Common questions.
Anything not covered here, ask the team directly through the form below.
What does the AI Cloud TCO Model include?
Four reconciled modules: per-accelerator performance and cost benchmarks, the all-in GPU cost of ownership and rental pricing, a cluster finance suite through a three-statement model, and workload and LLM economics.
How is the model delivered?
As an Excel workbook with dashboard access, including one year of quarterly updates, an onboarding call with the team to walk through the model and methodologies, and ad-hoc calls for questions that come up in use.
Is it part of the SemiAnalysis newsletter subscription?
No. Industry models are separate institutional offerings and are not included with the annual newsletter membership.
Can it underwrite an AI cloud investment?
That is what the finance suite is for: net present value and residual value from future cash generation, cumulative project and equity cash flow, IRR and the full returns stack, and a three-statement model whose capital structure, contract, and lifetime assumptions you control.
Which accelerators does it cover?
Nvidia and AMD accelerators plus the major custom silicon programs, TPU and Trainium among them, each with throughput, power, cost of ownership per hour, inference cost per million tokens, and training cost per FLOP.
Models that pair with this one.
Pairing depends on the decision in front of you. The cost stacks here price the hardware the other models track.
Accelerator & HBM Model
Demand forecastThe shipments and silicon the cost stack prices: SKU-level volumes, pricing, and HBM supply.
View model
AI Networking Model
Network buildThe switches, optics, and cables inside every cluster this model prices, device by device.
View model
Datacenter Industry Model
Capacity modelThe power, land, and shells the clusters deploy into, tracked site by site.
View modelGet the AI Cloud TCO Model.
Start with the sales team. They come back with scoping, licensing, and pricing for your mandate.
- Scoped to your use case
- Onboarding and ad-hoc analyst calls included
- Custom research engagements available
