How to Prove an AWS Discount Benchmark to Your Account Team Without Naming the Source

Your benchmark only works if AWS cannot dismiss it as hearsay, and it only survives if it never exposes the peer who gave it to you. This is the evidence architecture that makes a discount claim stick: anonymised published bands, effective-rate math built from public list prices, and competitor quotes reduced to unit economics.

The Real Problem Is Not the Number, It Is Whether AWS Has to Respond to It

A benchmark your account manager can wave away costs you the meeting. Say "we hear peers at our spend get 18 percent" with nothing behind it and you will get the standard reply: every customer is different, workload mix drives the rate, we price on strategic value not on tiers. That answer is not obstruction, it is procedurally correct, and it resets the conversation to zero while burning the one credibility shot you had. What actually happens after you cite a number is that the account manager runs it through three internal filters before deciding whether to spend political capital on it. First, is it verifiable, meaning can the AM point to something a deal desk reviewer can open independently. Second, is it comparable, meaning does it survive the mix, term, and commitment adjustments the desk will apply. Third, is it actionable, meaning does it map to an exception request the desk has a category for: cross-service rate increase, service-level adder, term-length premium. A benchmark that fails any one of the three is not a negotiation input, it is noise. Reframe your goal accordingly. You are not trying to win an argument about an unnamed peer. You are trying to hand your account manager a paragraph they can paste into a deal desk justification memo without their name on an unsourced claim. That is the entire job. Every evidence form in this article exists to survive those three filters, and the [published band tables by spend tier](aws-private-pricing-discount-bands-by-spend-tier) are useful precisely because a reviewer can open them.

You are not arming yourself to debate the account manager, you are arming the account manager to debate the deal desk.

What the AWS Customer Agreement Actually Forbids, and Where the Gap Is

Read section 11.9 of the AWS Customer Agreement for what it protects, not what it prohibits. AWS Confidential Information is defined to include the nature, content, and existence of any discussions or negotiations between the parties, so the negotiation itself is the protected object, not merely the percentage that came out of it. The duty runs through the term and for five years after it ends, which means a discount from a 2021 agreement that expired in 2024 is still covered today. The Service Terms close the obvious escape route: at section 1.7, where the agreement carries no confidentiality provision and no effective NDA exists, the customer still agrees not to disclose AWS Confidential Information except where law requires it. There is no gap to exploit through sloppy paperwork.

The commercial asymmetry is the part that matters at the table. When a named peer number surfaces, the peer is the party in breach, not you. You may face nothing worse than an awkward call. Your peer faces a conversation with their own AWS account team about how their EDP terms travelled, and the intelligence channel that produced the number closes permanently. Sourcing does not expose you to legal risk so much as it destroys the relationship that generated your best information, which is why disciplined buyers never name and never hint.

Two provisions define the usable space. Service Terms section 1.8 expressly permits customers to perform benchmarks or comparative tests or evaluations of the Services. That is a technical permission, not a commercial one: you can publish latency and cost-per-transaction findings, you cannot publish your rate card. The distinction is your operating line. Separately, 11.9 bars press releases and other public communications regarding the agreement, so the tactic of building leverage by publishing your own band is off the table before you start. Everything credible therefore has to come from sources already in the open or from an intermediary that absorbed the disclosure risk under its own NDA, which is the structural argument behind [using independent benchmark evidence rather than anecdote](benchmarking-negotiation-services-value-pillar).

Evidence Form One: Published Peer Bands You Can Cite by Name

The instinct to cite a peer is the mistake. The moment you say "a company our size at $18M is getting 22 percent," you have handed your account manager two gifts: a target to discredit and a reason to change the subject to who told you. Stop citing peers. Cite datasets. A published index is attributable, the account team can pull it up on their own laptop, and no individual customer is exposed. That reframing does more than protect a source. It changes the burden of proof. A peer anecdote invites AWS to say "every environment is different." A dataset built from 300-plus enterprise cloud commitment contracts, reporting by commitment tier and industry sector against list as of Q4 2025 and Q1 2026, forces the rep to either accept the methodology or attack a third party's sample, which they have no standing to do.

The material now exists in citable form. One dataset covering 1,200-plus contracts reports 12 to 18 percent off list at $5M to $20M annual spend and 18 to 28 percent above $20M. Advisory disclosures work the same way and are anonymised by construction: "roughly twenty to twenty-five AWS EDP negotiations benchmarked across 2024 and 2025" is a sample size, not a source leak. Use two conflicting band sets deliberately. When one source says 10 to 16 percent at $3M to $10M and another puts the overall envelope at 5 to 25 percent with service-level adders on top, the account team spends the meeting arguing with the market instead of with you. Pair this with the discount bands by spend tier analysis so your ask sits inside a published range rather than floating as an opinion.

Evidence Form Two: Effective-Rate Math Built Entirely From Public Prices

Effective-rate math is the only benchmark form that carries zero disclosure risk, because every input sits on a public AWS pricing page. Nobody signed anything to learn that m5.large On-Demand runs $0.096 per hour in us-east-1, that egress is $0.09 per GB for the first 10 TB, that cross-AZ traffic costs $0.01 per GB in each direction, that a NAT Gateway is $0.045 per hour plus $0.045 per GB processed, that a public IPv4 address is $3.65 per month, or that an EKS control plane is $0.10 per hour per cluster with Fargate at $0.04048 per vCPU-hour. These are list prices. Building an argument out of them cannot breach anything. More importantly, it changes the question from "what did somebody else get" to "what does my bill actually cost per unit," and the second question is one AWS cannot answer with a comparability objection.

Here is the number that ends the conversation. In a modelled environment of roughly $1,380 per month (ten m7i.large instances, gp3 storage, one ALB, twelve IPv4 addresses, 5 TB egress, 500 GB inter-AZ), instance hours account for $736 and ancillary charges make up 47 percent of the bill. If your headline cross-service discount lands hardest on compute, then a 20 percent headline is applying to roughly half your spend. Your realised rate is closer to 11 or 12 percent. Say that out loud, with the arithmetic on a single page, and the account team has to argue that half your invoice does not count. They will not. They will instead offer service-level adders on the untouched categories, which is exactly the outcome you want. Run the same exercise through the effective rate after Savings Plans model before you quote any number, because Savings Plans stacking compresses it further.

If ancillary charges are 47 percent of the invoice, a 20 percent headline on compute is an 11 percent deal.

Evidence Form Three: Competitor Quotes Reduced to Unit Economics

A live Google Cloud or Azure proposal is the one piece of evidence your AWS account team cannot dismiss on comparability grounds, because it is not AWS Confidential Information at all. It never touched AWS's confidentiality clause, it did not come from a peer who would be in breach by handing it to you, and it describes your workloads rather than someone else's. That makes it the strongest evidence form in the stack and the only one where the account team's usual rebuttal, that your source is unverified hearsay, collapses immediately. The market context is public enough to state plainly: enterprise agreements at Google Cloud typically land 10 to 18 percent at $5M to $20M of annual spend, 18 to 30 percent at $20M to $100M, and 30 to 45 percent above $100M, with Committed Use Discounts delivering 28 to 57 percent off on-demand before a further 15 to 35 percent is negotiated at the $5M+ level. On the Microsoft side, MACC renegotiations in the 2025 to 2026 cycle have been picking up 3 to 7 additional percentage points over the prior commitment. Those are bands you can name without naming anyone.

The discipline is in how you present it. You do not hand AWS the competitor's discount sheet, which almost certainly carries its own NDA and would tell your account team exactly which alternative you are running and how far it has progressed. You reduce it to unit economics for a named workload: cost per vCPU-hour, cost per terabyte of object storage per month, cost per million database reads, cost per egress terabyte, all fully loaded and modelled over 36 months including migration labor, dual-run overlap, and retraining. Present that as a migration cost model for one identified workload, not as a pricing comparison across the estate. The economics survive the redaction; the source obligation does not follow you into the room. AWS will respond by attacking the migration assumptions rather than the rates, which is exactly where you want the conversation, because those assumptions are yours to defend and yours to have already stress-tested. Do this alongside a serious read of what Google Cloud Committed Use Discounts actually deliver at three years, so the model reflects real alternative pricing and not a list-price fantasy AWS can puncture in one slide.

Pre-Empting the Comparability Rebuttal Before AWS Uses It

The account team's fastest counter is never magnitude. It is mix. They will not argue that your 22 percent band is wrong in the abstract. They will argue that the peer behind it had a different workload composition, a longer term, a steeper growth curve, or a single-region footprint that made their commitment easier to underwrite. That rebuttal is cheap, it is often partly true, and if you let AWS raise it first, you spend the rest of the meeting defending a source you refuse to identify. So raise it yourself, in your opening, before the number.

State the caveats as your own analysis. GPU-heavy estates receive roughly 15 to 25 percent less discount than comparable CPU estates at the same commitment tier, so if your mix is GPU-weighted, say so and adjust your ask downward by that amount in advance. Say it plainly and the account team loses its best move. Then state the stacking order, because this is where most buyers lose 5 to 10 points without noticing: the private pricing rate applies on top of pricing that has already absorbed Savings Plans at 20 to 55 percent and Reserved Instances at 20 to 65 percent, which means a headline percentage is not the number that hits your invoice. Walk the room through what your real effective rate looks like once Savings Plans have already landed, and insist that any figure AWS quotes back is expressed on the same basis.

Quote service-level adders separately from the cross-service rate, never blended. EC2 typically carries 5 to 12 additional points, S3 3 to 8, DynamoDB 5 to 10. Blending them lets AWS present a strong-looking composite that hides a weak cross-service floor, and it makes your benchmark unfalsifiable in the wrong direction. Keep the adder conversation and the term-length conversation on separate tracks as well, since one-year and three-year commitments carry structurally different premiums and mixing them gives AWS a free variable to hide behind.

What a Strong Outcome Looks Like and What AWS Will Do to Avoid It

Set the target before the first meeting, in numbers, because the account team will otherwise define success for you. At roughly $5M annual committed spend, published band data puts the achievable cross-service range at 12 to 16 percent off list, and a buyer arriving with single digits on the table has an evidence problem, not a spend problem. Above $20M, the same datasets put the range at 18 to 28 percent, and the top of that band correlates with longer terms and genuine competitive pressure rather than with polite persistence. Treat the cross-service rate and the service-level adders as two separate negotiations on two separate lines: EC2 compute typically carries 5 to 12 percent additional, S3 3 to 8 percent, DynamoDB 5 to 10 percent, and if you let AWS blend them into one headline number you will never know which one they shorted. Our read on how service-level adders stack on top of the cross-service rate is the right companion analysis before you price the trade.

The counter-sequence is predictable enough to script. First, a request for the source, framed as due diligence: decline once, politely, and restate that your evidence is published band data plus your own effective-rate math. Second, a non-cash sweetener: credits, MAP funding, training vouchers, professional services days. Price these honestly. A $500K credit pool against a $20M commit is 2.5 percent of one year, non-recurring, and often expiring; a two-point discount improvement on the same commit is $400K per year, recurring, and compounds through the term. Third, an above-band offer conditioned on five years, or on a 20 percent annual uplift schedule that quietly transfers the discount back through growth. Model the uplift: 20 percent compounding turns a $5M year-one commit into roughly $10.4M by year four, and a shortfall on that curve is your risk, not theirs. Fourth, deal-desk escalation timed to land in the last ten days of your fiscal quarter, when your own board deadline does the negotiating for them. The commit tier break and the uplift cap deserve their own follow-on work; both are where most of the value leaks after the headline is agreed.

What to Do First

Run this in ten days, in this order. Days one to three: pull three months of Cost and Usage Report data and compute your own effective rate per dollar of gross spend, including egress, NAT Gateway hours and processing, public IPv4 charges, and cross-AZ traffic. In modelled environments ancillary charges have run near 47 percent of the bill, so a discount applied only to instance hours is a smaller concession than it appears. Days three to five: assemble two published band citations that bracket your commitment tier, ideally from datasets that disagree, so the account team argues with the market rather than with you. Days four to six: commission a third-party benchmark review that returns a comparison under NDA, typically inside 48 hours, giving you an attributable intermediary who absorbs the disclosure risk. Days five to ten: open one competitor conversation far enough to yield a workload-level cost model, not a brochure. Our guidance on why documented evidence outperforms anecdote at the table covers how to sequence the review against the AWS calendar.

The sentence to say: "Published tier data for our commitment level shows 12 to 16 percent, our current effective rate is under 9 percent once ancillary spend is included, and I need your written proposal to close that gap by our board date." The sentence you must never say: any variation of "a company like ours is getting X." That names a peer by inference, puts them in breach of their own confidentiality clause, and hands AWS a reason to disqualify your entire evidence pack rather than answer it.

Related advisory: the full topic guide.

Frequently asked questions

Can I legally tell AWS what discount another company received?

Not if you learned it under an AWS agreement or from a peer bound by one. The AWS Customer Agreement treats the nature, content and existence of negotiations as confidential for the term plus five years, and the Service Terms impose an equivalent default duty even where no separate NDA exists. The practical exposure sits with the peer who disclosed, which is why sourcing a number by name destroys your intelligence pipeline faster than it wins a concession.

Will AWS actually move on a benchmark I cannot source?

Yes, if the evidence form is citable rather than anecdotal. Account managers need something they can put in front of a deal desk, and a published index built from hundreds of enterprise commitment contracts, or your own effective-rate math derived from public list prices, both qualify. A vague claim that a friend at another company got more does not, and it hands the account team an easy dismissal.

What is the difference between a headline discount and an effective rate?

The headline is the cross-service percentage AWS quotes. The effective rate is what you actually pay per unit after Savings Plans and Reserved Instances have already reduced the base, and after ancillary charges (egress, NAT Gateways, IPv4, load balancers, inter-AZ traffic) are included, which can be close to half the bill in some architectures. Presenting the effective rate is disclosure-safe because every input is public.

Should I show AWS a competitor's actual quote?

Show the economics, not the document. Competitor quotes usually carry their own confidentiality terms, and handing over a PDF invites a comparability fight over line items. A workload-level migration cost model, expressed in cost per compute hour or per terabyte of storage and egress for a named application, delivers the same pressure without the disclosure risk.

How does AWS typically counter a well-evidenced benchmark?

Expect three moves in sequence: a request for your source, an offer of non-cash value such as credits, migration funding or training in place of rate, and an above-band rate conditioned on a longer term or an aggressive annual commit uplift. Price each of those in cash before responding, because a credit pool and a discount point are not equivalent currency.

How many benchmark sources do I need to be credible?

Two published band sets plus your own effective-rate calculation is usually sufficient. Citing two datasets that disagree slightly is deliberately useful because it forces the account team to argue with the market rather than with you, and it signals that you understand the range rather than fixating on a single number you cannot defend.