Why this session exists
Session a3 was about who wrote the numbers. This one is about which numbers to look at. A platform budget presented to a board almost always has the right shape and the wrong emphasis: pages of detail on the visible unit prices, one line for storage, and nothing at all for the things that end up dominating the bill. The four leaks below are not exotic. They are ordinary, published, and available to anyone who reads a rate card carefully - which is why finding them is a leadership skill rather than a technical one.
Four leaks into one bill 15 min live
Here are the four, in order of how much money they move and in reverse order of how much attention they usually get. Token spend is the one every executive asks about first. It is the narrowest pipe on the diagram.
LiveLeak 1: it costs more to read your data than to keep it5 min▶
On AWS S3 (Price List API, publication date 2026-08-06), storage is $0.023 per GB-month and egress to the internet is $0.09 per GB after the first 100 GB free. Put those side by side and the ratio is the thing to remember: downloading a byte once costs about 3.9 times what it costs to store that byte for a whole month.
Every architectural argument about data movement is downstream of that ratio. It is why copying data between systems is expensive in a way that surprises people, why "just let the analytics tool read straight from the lake" is a budget decision rather than a plumbing decision, and why leaving a cloud provider is a capital event rather than a migration task.
The headline comparison, on the same fetch dates: egressing 100 TB in one month costs $0 on Cloudflare R2, $0 on Wasabi but only if you also store at least 100 TB, and around $8,563 on AWS. Storage prices across vendors differ by roughly three times. The egress gap is effectively infinite.
LiveLeak 2: the list price is not the price4 min▶
Three numbers for the same accelerator, all fetched 2026-08-07. AWS p5 on-demand is $6.88 per GPU-hour. The same instance on a three-year all-upfront Savings Plan is $2.587 per GPU-hour. The SemiAnalysis composite index for H100 capacity sits at $2.82 per GPU-hour (April 2026), down 57% from $6.62 in the second half of 2023.
Two consequences worth stating out loud in a budget review:
- Published list prices substantially overstate what the market pays. A model built on list is not conservative, it is wrong in a direction that happens to favour whoever built it.
- Azure H100 spot at $2.27 per GPU-hour is below AWS's three-year reserved rate. For interruptible batch work - backfills, embedding regeneration, offline scoring - spot capacity inside an existing contract is frequently the cheapest compute you can buy without leaving your security perimeter.
And prices move violently, which is a governance point as much as a finance one. AWS cut P5 by 44% effective 2025-06-01. Azure's $98.32 per hour for eight H100s is exactly AWS's pre-cut price, never matched. Any budget with a GPU line in it needs a scheduled re-price, not an annual one.
LiveLeak 3: licensed per installed GPU, not per GPU used4 min▶
NVIDIA AI Enterprise is one of the very few published rate cards in this whole subject (docs page updated 2026-06-08). Self-managed is $4,500 per GPU per year; the cloud marketplace rate is $1.00 per GPU-hour on top of the instance cost. Those cross at about 51% annual utilization, since $1/GPU-hour works out to $8,760 per GPU-year.
The trap is the denominator. The licence is charged per installed GPU, not per GPU used, and there is one licence per physical GPU on a multi-GPU card. An eight-GPU DGX B300 therefore carries $36,000 per year in licensing before any hardware cost at all, whether those GPUs are saturated or idle. This is the single line item first-pass budgets miss most often, and it does not shrink when utilization does.
One more detail from the same rate card, purely because it is free money: the five-year term is $18,000 and the four-year term is also $18,000. Five years for the price of four. There is never a reason to buy the four-year term.
Self-studyLeak 4: token spend, and why it is the narrow pipe3 min read▶
Token spend is the leak every executive asks about first, and on most enterprise platforms in 2026 it is the smallest of the four. That is not a reason to ignore it - it is the fastest-growing line and the one most likely to be spent without attribution, which is a governance problem rather than a cost problem.
The structural point for a leader: token spend is the only one of the four leaks with a natural chokepoint already available. Layer seven of the stack, the LLM gateway, exists precisely so that routing, budgets, rate limits and per-team attribution happen in one place. If your token spend is untracked, you do not have a pricing problem, you have a missing layer.
The chokepoint test from a2 applies here exactly as it did there: a gateway that teams can skip by calling the provider API directly gives you cost dashboards covering the compliant half of your spend.
Self-studyThe fifth line: requests, and why it is a rounding error2 min read▶
Object storage bills have a third component after storage and egress: request charges. On S3 (fetched 2026-08-06) PUT, COPY, POST and LIST requests are $5.00 per million and GET requests are $0.40 per million. Cloudflare R2 prices operations exactly 10% under those figures, at $4.50 and $0.36 per million.
Two things worth noticing. First, on a normal analytics workload this line is small enough to be a rounding error next to egress, so it should not drive a vendor decision. Second, and more interesting commercially: R2's 10% discount on operations exists to remove the last counter-argument to its zero-egress pitch. It is a pricing move aimed at a specific objection, and reading it that way tells you more than the discount itself does.
Budget from billing feeds, not aggregators 8 min live
Every price on this page came from a vendor's own machine-readable billing feed rather than a comparison site or a marketing page, and there is a specific reason for that discipline. The most-linked GPU pricing aggregator on the internet is currently wrong by about a third on one of its headline rows, and the error runs in the direction that makes reserved capacity look cheaper than it is.
The aggregator error. The most-linked GPU pricing aggregator (Vantage) labels AWS's three-year, no-upfront rate of $23.777 per hour as "1-Year Reserved." AWS's real one-year rates are in the $32-35 per hour range. That is a 32% error in the single most cited GPU pricing reference, on the exact row a budget owner is most likely to copy into a model. Nobody is being dishonest; a label was mapped wrongly and then linked ten thousand times. Use first-party billing feeds.
LiveWhat is published to the dollar, and what is not4 min▶
Here is a pattern worth carrying into every hardware conversation you have this year. NVIDIA publishes its software prices to the dollar and publishes no list price for any DGX system. Every DGX figure circulating in the market - including the roughly $3M often quoted for a GB200 NVL72 rack - traces back to an HSBC analyst estimate, not to a vendor catalog.
So the rule is: budget by quote, not by catalog. If a business case contains a hardware price you cannot trace to a quotation with your organization's name on it, that price is an estimate wearing the costume of a fact.
What you can plan against, because it is physical and falsifiable:
- A DGX B300 draws roughly 14 kW in 10U. Power and cooling are the binding constraint. The facility runs out before the budget does, and facility work has a longer lead time than procurement.
- Storage bandwidth scales linearly at about 12.5 Gb/s per GPU. Sixteen GPUs need roughly 200 Gb/s. That single rule turns a vendor conversation into arithmetic you can do on the back of an agenda.
Note what those two have in common: they are engineering constraints, not commercial claims, which is exactly why they are the numbers that survive contact with a procurement cycle.
LiveThe denominator trap: check what a price is per3 min▶
One more reading error, and this one is worth eight times your money. CoreWeave publishes per-instance prices, not per-GPU prices. Divide by eight or your model is wrong by a factor of eight, and their HGX shapes are sold as whole eight-GPU nodes anyway, with no single-GPU option.
Once you do divide correctly, a piece of received wisdom falls over. CoreWeave's H100 at $6.155 per GPU-hour against AWS at $6.88 is an 11% discount, not the two-to-three times that the specialist-cloud narrative implies (both fetched 2026-08-07). The genuine discounters on list are RunPod at $2.69 and Lambda at $3.99. What CoreWeave actually sells is committed contracts and InfiniBand fabric, which may well be worth buying - just not for the reason the market repeats.
So the habit for every price you are shown: ask what the denominator is. Per GPU or per instance. Per used unit or per installed unit. Per hour of runtime or per hour of reservation. Three of the four leaks in this session are denominator errors rather than price errors.
Self-study"AI Factory" and other unfalsifiable words3 min read▶
"AI Factory" is branding, not a technical standard. There is no independent certifying body, so no proposal can be disqualified for failing to be one. When a term cannot be failed, it cannot be used in an evaluation.
Underneath the branding there are falsifiable things, and those are what belong in a requirements document: validated node counts, certified-partner lists, and the storage sizing rule above. One sharp detail on certification, because it costs money to get wrong: the DGX SuperPOD certified-storage list is a subset of the general NVIDIA-Certified Storage list. A vendor being "NVIDIA-certified" does not qualify them for a SuperPOD purchase, and that distinction has been discovered late by more than one buyer.
Storage economics, run live ★ 20 min · everyone runs it
This is the only calculator in the leader track, and it does one thing: prices the same 50 TB of data against three published rate cards fetched 2026-08-06/07. No projections, no assumptions about your organization. Just three vendors' own numbers applied to one volume, so you can see how much of a storage bill is not storage.
Read the default first, before touching anything: 50 TB stored, 20 TB egressed per month. Wasabi $400/month, Cloudflare R2 $768/month, AWS S3 $2,961/month - and $1,784 of that S3 figure is egress, not storage.
Say the gap out loud, because it is the number that lands in a steering committee: S3 costs $2,193 a month more than R2 - $26,317 a year - and almost all of it is egress. The bytes are identical. The difference is what it costs to read them back.
Now note what this does not say. It does not say leave AWS. Egress is one line in a platform bill, and proximity to the rest of your estate has real value. It says: know which line you are actually paying, because most organizations believe they are buying storage.
Push egress above stored volume and watch the Wasabi row flip to a warning. You have left the 1:1 free-egress policy. Read carefully what that warning is: not a metered overage with a published penalty rate, but a terms-of-service risk. Wasabi's own words: "we reserve the right to limit or suspend your service."
That distinction is the leadership point of the whole demo. A billed overage is a number you can forecast and approve. A suspension right is an operational risk with no price attached, and it belongs in a risk register rather than a budget line. Two vendors can both say "free egress" and mean materially different things.
Before you sign any infrastructure contract ◐ 6 min read
Four questions. They are deliberately dull, they take a supplier about ten minutes to answer, and between them they cover every leak in this session. Ask them before signature, because after signature they become negotiations rather than questions.
Self-study1 · What is the egress cost of leaving?2 min read▶
Not the monthly egress cost of operating - the one-off cost of moving everything out. At S3's published tiers that is a large, calculable number, and it is the honest measure of switching cost for the storage layer.
Ask specifically whether any exit waiver applies and read its conditions. AWS's free-exit-egress programme, announced 2024-03-05 and still live, waives transfer-out charges for customers migrating off AWS - but it is request-based, approval-gated, credited after the fact, and time-boxed to 90 days. It was driven by regulation, principally the EU Data Act, rather than by a rate-card change. A waiver you must apply for and complete inside a quarter is a real benefit and a poor foundation for a migration plan.
Self-study2 · What is the minimum retention?2 min read▶
Minimum storage durations turn a cheap headline rate into an expensive one for high-churn data. Wasabi's own terms: "if stored objects are deleted before they have been stored with Wasabi for a certain number of days, a Timed Deleted Storage charge equal to the storage charge for the remaining days will be applied." That is 90 days on Pay-Go and 30 on Reserved Capacity, plus a 1 TB minimum billing (10 TB for Cloud NAS).
The practical reading: the cheapest per-terabyte rate on the market can be the wrong choice for short-lived, high-churn data, and nothing on the pricing page tells you that. Match the retention term to the actual lifecycle of the data, not to the headline rate.
Self-study3 · Is licensing per used unit or per installed unit?1 min read▶
Straight from Leak 3, and it generalises well beyond GPUs. Per-installed-unit licensing decouples your software bill from your utilization entirely, which means efficiency work delivers nothing on that line. Ask for the denominator in writing, ask what happens to it when a node is idle, and ask what happens when a card carries multiple physical processors.
Self-study4 · What does the reserved rate actually look like?1 min read▶
Ask for the committed rate from the vendor's own feed, with the term, the upfront structure, and the date on it. Do not accept an aggregator screenshot - that is precisely the 32% error described in Part 2. And note the trap running the other way: Lambda's committed clusters are inverted, with a 16-GPU H100 cluster at $6.16 per GPU-hour, 54% more than its eight-GPU on-demand rate of $3.99, because what you are buying is fabric and guaranteed capacity rather than raw hours. Committed does not automatically mean cheaper.
Try it yourself - this week ◐ 30 min total
- Pull your last three months of object-storage billing and split it into storage, egress and requests. Most people are surprised by the proportions, and the surprise is the point.
- Find every per-unit licence in your platform estate and write down the denominator for each: per used unit or per installed unit. Anything in the second category is immune to efficiency work.
- Take one GPU price currently in a business case and trace it to a first-party feed with a date on it. If it came from an aggregator or a slide, replace it.
- Ask your incumbent storage vendor the four contract questions from Part 3 in writing. Keep the answers - they become the switching-cost baseline for a5 and a6.
- Put a re-price date in the calendar for every GPU line in your budget. AWS moved P5 by 44% in a single day; an annual refresh is not fast enough.
Official sources covered
Every price on this page was pulled from a vendor's own machine-readable feed or published rate card and carries its fetch date of 2026-08-06/07. Nothing here comes from a comparison site, and Part 2 explains why that rule exists. This page covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · On AWS S3, how does the cost of reading a byte out compare to the cost of storing it?
That ratio drives architecture. At 50 TB stored and 20 TB egressed a month, S3 comes to $2,961 against R2's $768 - $2,193 a month, or $26,317 a year, almost all of it egress. Egressing 100 TB in one month is $0 on R2, $0 on Wasabi only if you also store at least 100 TB, and around $8,563 on AWS.
2 · NVIDIA AI Enterprise is $4,500 per GPU per year self-managed. Why does an idle eight-GPU DGX B300 still cost $36,000 a year?
Per-installed-unit licensing decouples your software bill from utilization entirely, so efficiency work returns nothing on that line. It is the item first-pass budgets miss most often. The cloud alternative is $1/GPU-hour, which works out to $8,760/GPU-year and crosses the self-managed price at roughly 51% annual utilization. Separately: the five-year term is $18,000 and so is the four-year term, so never buy the four-year.
3 · The most-linked GPU pricing aggregator labels AWS's 3yr-no-upfront rate as "1-Year Reserved." What is the lesson?
AWS's real one-year rates are $32-35/hr. Nobody was dishonest; a label was mapped wrongly and then linked everywhere. The same discipline applies to hardware: every circulating DGX figure traces to an HSBC analyst estimate. What you can plan against is physical - a DGX B300 is about 14 kW in 10U, and storage sizing runs ~12.5 Gb/s per GPU, so 16 GPUs need roughly 200 Gb/s.