AI Infrastructure Economics Series · Tool 4

Multi-Cloud CapEx Comparison Calculator

Three vendors. One workload. The bill rarely comes out the same. Model 3-year cloud TCO across AWS, Microsoft Azure, and Google Cloud — with reserved-instance commitment strategies, enterprise discounts, and the egress-trap math that decides more vendor selections than the compute rate ever does.

On the defaults — and their honesty. List prices for major SKUs were verified against Vantage Instances, DoiT, and CloudPrice on April 30, 2026 (us-east-1 / East US / us-central1, Linux on-demand). Reserved-instance discount tiers, inter-region rates, managed-DB and other-services rates, and negotiated EDP/EA discounts are reasoned approximations — directionally correct, not authoritative for any specific contract. The full sourcing table below shows which is which. Cloud list prices change frequently; verify against the vendor pricing page before signing anything.

Six reference workloads. Each sets compute mix, storage, egress, and managed-service counts. Every value remains editable. CDN / Media Streaming and Batch / ETL are deliberately included to demonstrate the Egress Trap and commitment-ladder charts respectively. "Custom" leaves current values unchanged.

A single workload specification applied uniformly across all three vendors. The tool maps to the most-equivalent SKU per vendor; if you are running different sizes per cloud, override the per-vendor rates in the next section.

Compute

FamilyReference SKUAWSAzureGoogle CloudInstance countHours / mo
General purpose4 vCPU · 16 GBm7i.xlargeD4s v5n2-standard-4
Compute optimized4 vCPU · 8 GBc7i.xlargeF4s v2c2-standard-4
Memory optimized4 vCPU · 32 GBr7i.xlargeE4s v5n2-highmem-4
GPU (8× H100)192 vCPU · 1.5 TB · 8× H100p5.48xlargeND H100 v5a3-highgpu-8g

Storage, network, and managed services

How compute consumption is split across pricing models. The same mix is applied to all vendors for fair comparison. Storage, network, and managed-service rates do not benefit from this mix.

On-demand
List rate, no commitment
%
1-year reserved
No-upfront convertible / standard
%
3-year reserved
All-upfront, deepest discount
%
Spot / preemptible
Interruptible, batch / dev only
%

List prices for the equivalent SKUs. Discounts are negotiated reductions. Override any value with your published quote, EDP/EA agreement, or committed-spend rate. The "1-year RI %" and "3-year RI %" are the discount off on-demand for that commitment.

AWSAmazon Web Services

Compute on-demand $/hr
Reserved discounts (%)
Storage $/GB-mo
Network & services

AzureMicrosoft

Compute on-demand $/hr
Reserved discounts (%)
Storage $/GB-mo
Network & services

Google CloudGCP

Compute on-demand $/hr
Committed-use discounts (%)
Storage $/GB-mo
Network & services
AWS · 3-year TCO
Azure · 3-year TCO
GCP · 3-year TCO
Cheapest vendor
Spread (max − min)

Price is one input. The factors below reshape the verdict without altering the math — they surface what the calculator cannot compute. Toggle anything that applies; the verdict line below adjusts to acknowledge what the price comparison alone cannot answer.

Microsoft-stack heavy estate
Windows Server, SQL Server, .NET, or Active Directory dominate the workload. Azure Hybrid Benefit (~30–55% discount on Windows / SQL VMs under existing licensing) is not modeled in the price calc — and it is often decisive for Microsoft-stack workloads.
Existing committed-spend agreement
EDP / EA / Google committed-spend contract already in place — generally with:
Hard compliance gate
FedRAMP High, ITAR, sovereign data residency, sector certifications (HIPAA-BAA, PCI, IRAP, etc.) pre-filter the candidate list before price even matters. Verify the cheapest vendor qualifies in the specific regions you operate in.
Vendor-specific service dependency
Workload requires a service with no direct equivalent on other clouds — BigQuery, Vertex AI, Bedrock+Claude, Azure OpenAI, AWS Trainium, GCP TPUs, Cosmos DB, etc. Forces a hand:
Multi-cloud already required
Business continuity, M&A inheritance, regulatory redundancy, or negotiating-leverage policy already mandates multi-cloud. The single-vendor framing of this tool then applies per workload, not enterprise-wide — the question is which workloads to which vendor, not which vendor wins overall.
Significant existing data on a vendor
More than ~500 TB already stored on one of the three. Egress to leave is real cost not modeled here:
Decision read

3-year TCO by vendor & category

Stacked breakdown across compute / storage / network / managed services. The shape of the bar tells you where the negotiation lives.

Effective compute rate by commitment mix

Blended $/hour for general-purpose VMs. The gap between vendors at 100% on-demand often closes — or inverts — at 3-year reserved.

Egress trap

Egress as a share of total monthly cost. Above 8% is a yellow flag for vendor lock-in; the lever to pull is architecture, not procurement.

Sensitivity tornado · cheapest vendor

3-year $ delta for the cheapest vendor as each lever is moved alone. Negative bars (green) reduce the bill; positive (orange) increase it.

A worked example with sample values, ending with the punchline most cloud-cost calculators miss: when "cheapest vendor" is and isn't the right answer.

What this walks through. A 50-person enterprise SaaS company evaluating AWS, Azure, and Google Cloud for production. Microsoft-stack heavy estate (Windows, SQL Server, .NET) under an existing 3-year Microsoft EA. Two runs: first without Decision Context, second with two flags toggled. End result — a decision read that contradicts the headline price answer, and shows you exactly why.
Step 1Pick a preset

Sets the workload shape. Every value remains editable. Six reference workloads cover web platforms, enterprise mixes, CDN/media (egress-heavy), data platforms (storage-heavy), batch processing (spot-friendly), and AI training (GPU-heavy).

ActionSampleWhy
Click presetEnterprise General MixClosest match to a 50-person SaaS company with mixed compute, modest GPU usage, and balanced storage. Override individual fields if your actuals differ.
Step 2Confirm the workload

Production fleet, storage volumes, network usage, managed-service counts. Same workload spec is applied identically to all three vendors for fair comparison.

AttributeSampleNotes
General-purpose VMs50 × 730 hrsWeb tier, application servers (always-on)
Compute-optimized VMs20 × 730 hrsAPI workers, task queues
Memory-optimized VMs15 × 730 hrsCache layer, in-memory analytics
GPU (8× H100)2 × 500 hrsInference workloads, ~70% utilization
Block / Object / Archive storage50 / 200 / 500 TBPersistent volumes + content + cold backup
Egress / Inter-region10 / 5 TB · moCustomer-facing API + cross-region replication
DB / K8s / Other services5 / 2 / $2,000Catalog DBs, two prod K8s clusters, monitoring stack
Step 3Set the commitment strategy

Defines the mix of pricing modes used to run that workload. Same mix applied to all vendors. Must sum to 100%.

ModeSampleUse when
On-demand30%Burst capacity, dev environments, unpredictable spikes
1-year reserved30%1-year horizon confidence on production fleet
3-year reserved30%Steady-state production with confidence in capacity
Spot / preemptible10%Batch processing, fault-tolerant async workers

Heuristic: if your spot share is 0% and 3-year reserved is <30%, you're probably overpaying. If 3-year is >60%, you're betting the architecture won't change for three years — verify that bet against your roadmap.

Step 4Confirm per-vendor pricing

Defaults are verified Q2 2026 list rates. Override with your negotiated quotes as procurement evidence lands. The AWS H100 price cut from June 2025 (~44% off list) is reflected in the default — this materially shifts AI-heavy workload economics versus pre-cut analysis.

LeverAWSAzureGCP
General-purpose $/hr0.20160.1920.194
GPU 8×H100 $/hr55.0498.3288.48
3-year reserved discount60%55%55%
Egress $/GB0.0900.0870.085
Enterprise discount10%10%12%
Step 5Read the headline KPIs (first run, no Decision Context)

Scroll up to the KPI tile row. With these defaults, the calculator returns three vendor totals and a verdict on the spread.

AWS
3-year TCOcheapest
Strengthdeepest 3y RI
GPU rateleads post-cut
Azure
3-year TCOmiddle
Strengthgeneral compute
WatchGPU ~78% above AWS
GCP
3-year TCOmiddle
Strengthdeepest 1y CUD
Watchblock storage rate
First-run verdict (price-only): AWS comes out materially cheapest by a single-digit-percent over 3 years. Verdict box border: green. Decision read: "AWS is the answer." If you stop here, you ship the AWS recommendation. The next two steps explain why that's premature.
Step 6Read the four charts

The numbers tell you the size of the bill. The charts tell you why one vendor wins — and which lever moves which bill.

3-year TCO by category
Compare bar shapes, not just heights. Same total can hide different cost shapes — and shape decides which lever helps which vendor.
Effective compute rate ladder
The vendor cheapest at 100% on-demand may not be cheapest at 60% three-year reserved. Watch where the lines cross.
Egress trap
If any vendor's egress share exceeds ~8% of total bill, you have an architecture decision (CDN, edge caching, regional placement), not a procurement decision.
Sensitivity tornado
The longest bar is your biggest decision. If "EDP discount" dominates, that's a procurement problem. If "GPU hours" dominates, that's an architecture problem.
Step 7Second run — apply Decision Context

Toggle the qualitative factors that price alone can't see. The price KPIs do not move. The verdict text and verdict box border do.

Decision Context flagSet toWhy this matters here
Microsoft-stack heavy estateONProduction runs on Windows / SQL Server / .NET — Azure Hybrid Benefit (~30–55% off Windows VMs for owned licenses) is not in the price calc
Existing committed-spend agreementON · Azure3-year Microsoft EA already in place — switching costs (remaining commit, retraining, ecosystem rebuild) are real but unmodeled
Hard compliance gateoffNot applicable in this scenario
Vendor-specific service dependencyoffWorkload is general-purpose, no vendor-only service required
Multi-cloud already requiredoffSingle-cloud production is acceptable here
Significant existing data on a vendoroffLess than 500 TB of existing data — not gravity-bound yet
What changes in the verdict: the price KPIs do not move — AWS still shows lowest 3-year TCO. But the verdict box border turns amber, and two caveats appear under the price decision read:
  • "Microsoft-stack flag changes the math. Azure Hybrid Benefit (~30–55% off Windows VMs) is not modeled here. Re-run with a 35–45% discount applied to Azure compute to test whether the spread closes or inverts."
  • "Existing Azure commitment costs to leave. Switching cost not modeled: remaining commit, egress to migrate data, retraining, ecosystem rebuild. Heuristic: switching pays back only if the new vendor saves >20% AND migration burden is <12 months."
Step 8The actual decision

After Decision Context applies, the math typically lands here:

AWS — price answer
Headlinecheapest
Butno MS license credit
And~12 mo migration off Azure
Netprice ≠ decision
Azure — context answer
Headline2nd on price
+ Hybrid Benefit−30–55% Windows
+ Zero migrationexisting EA
Netlikely lowest TCO
GCP — context answer
Headlinemiddle on price
No advantageno MS credit, no commit
Netout, unless other factors flip
The point: AWS won the price comparison. Azure wins the actual decision. The Decision Context flags surfaced what the headline number couldn't. This is why "cheapest vendor" is rarely the same as "right vendor."
Step 9Export

Capture configuration, computed totals, and the active Decision Context flags for finance review or RFP appendix.

FormatWhat it's for
JSONFull nested structure including Decision Context flags. Re-importable into the calculator. Best for programmatic use, version control, or carrying scenarios across team members.
CSVFlat key/value file. Opens in Excel or Sheets. Best for finance modeling, RFP appendices, audit trails.
What you walk away with

A defensible 3-year TCO range across three vendors with documented assumptions. A clear view of which lever — pricing, commitment mix, architecture, or qualitative context — is doing the deciding. The number for the board memo is rarely the cheapest vendor. It's the vendor that's cheapest under the commitment strategy you actually intend to operate, with the qualitative constraints your business actually faces. The Decision Context flags are how price-comparison output becomes vendor-selection input.

Download the current configuration. JSON preserves full structure for re-import or programmatic use. CSV is a flat key/value file that opens directly in Excel or Google Sheets.

Show every default value, its source, and its rationale
ValueDefaultSource & rationale
AWS general-purpose $/hr$0.2016Verified Apr 30, 2026. AWS m7i.xlarge (4 vCPU / 16 GB) Linux on-demand, us-east-1.
Azure general-purpose $/hr$0.192Verified Apr 30, 2026. Azure D4s v5 (4 vCPU / 16 GB) Linux pay-as-you-go, East US.
GCP general-purpose $/hr$0.194Verified Apr 30, 2026. GCP n2-standard-4 (4 vCPU / 16 GB) Linux on-demand, us-central1.
AWS GPU 8×H100 $/hr$55.04Verified Apr 30, 2026. AWS p5.48xlarge us-east-1. AWS announced ~44% H100 price reduction June 2025; some third-party trackers still show pre-cut $98.32 figures.
Azure GPU 8×H100 $/hr$98.32Verified Apr 30, 2026. Azure ND96isr H100 v5 East US — notably ~78% higher than AWS post-cut.
GCP GPU 8×H100 $/hr$88.48Reasoned approximation, partially verified. GCP a3-highgpu-8g us-central1 list. Wide variance — verify the SKU and region for your case.
1-year reserved discount25–37%Reasoned approximation. AWS Standard RI no-upfront ~30%; Azure RI ~25%; GCP CUD ~37%. Real discount varies by region, instance family, payment option.
3-year reserved discount55–60%Reasoned approximation. AWS RI all-upfront ~60%; Azure 3-year RI ~55%; GCP 3-year CUD ~55%.
Spot / preemptible discount65–70%Reasoned approximation. Spot pricing is dynamic and capacity-dependent.
AWS S3 Standard $/GB-mo$0.023Verified Apr 30, 2026. First 50 TB/month tier in us-east-1.
AWS Glacier Deep Archive $/GB-mo$0.00099Verified Apr 30, 2026. Retrieval costs and minimum retention NOT modeled here.
AWS egress $/GB$0.090Verified Apr 30, 2026. First-tier internet egress (after free 100 GB/month), us-east-1.
Other vendor storage / egressvariesReasoned approximation. Verify against vendor pricing pages.
Inter-region transfer $/GB$0.020Reasoned approximation. Real cost varies significantly by source/destination region pair.
Managed K8s $/cluster-mo$73Reasoned approximation. $0.10/hour control-plane fee × 730 hours.
Managed DB $/hr$0.30–$0.34Reasoned approximation. RDS db.m6g.large equivalent. Variance across edition and HA.
Enterprise discount (EDP / EA / negotiated)10–12%Industry estimate, not vendor-published. Real range 5–25% depending on commit size and competitive pressure. Override with your actual quoted rate.
Hours per month730Standard cloud-bill convention.
3-year horizon36 monthsMatches RI / CUD term length.