The Price List Is Not the Kitchen

In nearly every conversation about cloud, somebody asks to see the price list. The request arrives as if it were the final step before a decision: show me the rates, let me look them over, and I will choose. At a smaller company that is easy to satisfy, because the public pricing pages publish a rate for everything. At a very large company the list is a different document, negotiated rates covered by confidentiality, because the discount itself is competitive information. Either way the same assumption sits underneath: that a list of prices is the thing you need in order to decide.

A price list answers one question: what does each configuration cost. The question people bring to it is usually a different question: can I have this, in the volume I need, starting next month. The restaurant version of the gap is familiar to anyone who has ordered the lobster at a good restaurant. The printed card prices every dish. You order. And there is a real chance someone comes back and says: we ran out. Every price on the card was true. None of them told you what the kitchen had.

A cloud price list has the same gap, and in cloud the gap is wider, because the dishes that tempt enterprise buyers most, the top-end AI accelerators, are the ones most likely to be out.

What a price list is

A cloud catalog sells SKUs, one sellable configuration of a service: one size, one feature set, one price (chapter 3). Every line on a bill names a SKU and every line on a price list prices one. AWS publishes on-demand rates on public pricing pages and exposes the same data through a Price List API, so the rates are not just visible, they are queryable.1 What the list does not carry is any answer to the second question. Nothing on a rate line says whether the configuration is launchable in your subscription, whether your region has stock, or how long you would wait. The list prices the dishes; the kitchen is a separate department.

The enterprise list is a different document

At a large company, the list you are shown is rarely the public one. Large accounts negotiate discounts in exchange for committed spend, and the resulting rates are not publicly disclosed, by design: the discount is the deal.2 3 This list is real and worth having. Budgets get built on it. It is still a list of prices and nothing else. Two large companies will often be looking at two different prices for the same SKU, because each deal was negotiated separately. That part is normal. What neither list carries is whether the SKU can actually be launched.

What the providers’ documentation says

The gap is not a claim from outside. The providers say it themselves. Microsoft’s quota documentation states: “An assigned Quota does not reserve or guarantee capacity for Customer use, and capacity may not be available at the time of request.”4 Google’s states that allocation quotas are the maximum you can create “if those resources are available,” and that “quotas don’t guarantee that resources are always available.”5 The permission to order and the existence of the dish are separate grants. AWS adds a third layer: on-demand quotas for several GPU families default to zero vCPUs for a new account, so the rate is published, the permission is a request, and the stock is a separate thing again. 1

The reservation products are the honest exception, because availability is what they sell. AWS Capacity Blocks hold accelerator capacity for a future window, paid up front, and the pricing page states that “reservation prices are updated regularly based on trends in supply and demand,” with the next update scheduled for October 2026.6 The price of holding a dish has moved twice this year: about 15 percent in January, and around 20 percent more from July 1.7 8 That is the price of certainty moving with scarcity (chapter 20, chapter 10).

“We ran out of that today”

The restaurant’s apology has a cloud equivalent, and customers have been posting it on the providers’ own public Q&A sites, where a customer describes the problem in public and a support engineer answers in public.

On Microsoft’s Q&A site, an engineer running a production GPU workload describes opening quota requests for non-Spot H100 capacity in five Azure regions in late June 2026. Ninety days later: no approval, no denial with a reason, no estimate, no named owner, while a Spot H100 production VM was evicted five times, four of them within 48 hours. Microsoft’s reply states the separation plainly: “Azure quota and actual VM capacity are separate.”9 A second thread describes being unable to allocate a single H100 node on a dedicated Azure ML cluster in East US for more than three days, with one job queued about 45 hours before cancellation; support’s diagnosis was regional capacity unavailability with no published ETA, and the recommended workarounds were another region or another GPU.10 Google’s forums carry an earlier stage of the same story: a request to raise an A100 80GB quota from zero to one, one GPU, declined seconds after submission.11 The door closed before the order reached the kitchen.

The pattern across all three: the dishes that go missing are the rare ones. A general-purpose VM family is almost always there. The top-end accelerators concentrate demand, every large buyer wants the same GPUs at the same time, so those SKUs run short first, exactly the ones an enterprise price list makes look most attractive (chapter 24).

Price and availability are one signal

When demand outruns supply, the rate rises and the shelf empties together. They are two readings of the same condition. The reservation products make the connection explicit: you pay to hold the dish, and the holding price moves with scarcity. The rest of the catalog carries only the price half of the signal. The availability half stays hidden inside quota pages, region tables, and support tickets, which is why a price list can look complete and still omit the fact that decides the project. The book’s last chapter argues that what cloud capacity lacks is a shared signal, every rate negotiated privately and every discount kept confidential (chapter 24), chapter 24). Prices without availability are half a market. A price list will be the last document to tell you the kitchen ran out.

The availability picture also changes month to month. Providers rarely announce when a family becomes constrained. The change arrives in smaller ways: a quota request that used to be approved in days now sits unanswered, an allocation that used to land now queues, a region lists fewer sizes in the documentation than it did before. These are transitions you learn about because you have invested in the relationship with your providers. An account team that knows your roadmap can flag a constraint before a launch depends on the SKU (chapter 3). A company that only transacts finds out last, at the moment the launch fails. The relationship is not a courtesy; it is the channel the other half of the signal travels on.

What to ask for instead

So ask for the price list. Budgets need rates, and the request is fair. Then ask the questions the list cannot answer. Which SKUs can we actually launch: quota by family and region, for every subscription that will run the work, read as permission rather than stock. 4 5 How the rare ones get held: for top-end accelerators, ask what holds capacity for you, reservation, capacity block, allocation through the account team, and what the holding costs, because the holding is priced and repriced. 6 What happens when the dish is gone: a second region, a second SKU, a tested alternative (chapter 10). And who tells you when things change: who owns the account, how often you review consumption and upcoming needs together, and how you hear about a constraint before it reaches a launch.

A price list is the right document for the budget conversation and the wrong document for the availability conversation. The providers’ own documentation draws that line; the customer posts on the public Q&A forums show what the far side of it looks like. Ask for both halves, and remember which one you are holding. You are not dining at a fine restaurant. You are dining at a cloud provider.

Sources

1. Amazon Web Services, Purchasing On-Demand Instances for Amazon EC2, EC2 User Guide, checked October 8, 2026 — on-demand prices fixed and listed on the public pricing page, queryable via the AWS Price List API; per-family running on-demand quotas, with GPU families (G and VT, P, and others) defaulting to 0 vCPUs and adjustable by request.

2. Usage.ai FinOps Glossary, Negotiated Discount (Cloud) — custom pricing reductions exchanged for committed spend; the Enterprise Discount Program on AWS, the Microsoft Azure Consumption Commitment on Azure, and private pricing agreements on Google Cloud; exact discounts vary by deal and are not publicly disclosed.

3. CloudForecast, 2025: AWS EDP PPA 2025 Negotiation Guide and Cost Savings — structure of enterprise discount programs as industry context for negotiated, non-public enterprise rates.

4. Microsoft Learn, Quotas overview — Azure Quotas, last updated June 16, 2026, checked October 8, 2026 — “An assigned Quota does not reserve or guarantee capacity for Customer use, and capacity may not be available at the time of request”; approval of quota increases is subject to review of current usage, service availability, and capacity constraints.

5. Google Cloud documentation, Allocation quotas — Compute Engine, last updated August 26, 2026, checked October 8, 2026 — quotas are maxima “if those resources are available”; “Quotas don’t guarantee that resources are always available”; creation fails even with remaining quota when a region is depleted.

6. Amazon Web Services, Amazon EC2 Capacity Blocks for ML pricing, checked October 8, 2026 — “Reservation prices are updated regularly based on trends in supply and demand for EC2 Capacity Blocks”; current prices scheduled to be updated next in October 2026; the reservation fee is charged up front at the prevailing rate at purchase.

7. InfoQ, January 2026: AWS Hikes EC2 Capacity Block Rates by 15% in Uniform ML Pricing Adjustment — a 15 percent increase in EC2 Capacity Block rates for ML instances at the start of 2026.

8. Ben Wodecki, SDxCentral, 2026: AWS raises GPU Capacity Block prices again as AI demand outstrips supply — second increase in six months, around 20 percent effective July 1, 2026; an H200-based p5e.48xlarge rate moving from about $34.61 to about $39 per hour in most regions; AWS confirming prices are “updated periodically based on supply and demand.”

9. Microsoft Q&A, September 2026: 90 days, 5 support cases, zero answers: how Azure handles H100 capacity requests — quota requests for non-Spot H100 (ND/NC H100 v5) in five regions opened late June 2026; no approval, denial with reason, ETA, or named owner after 90 days; five Spot H100 evictions, four within 48 hours; Microsoft’s reply that quota and actual VM capacity are separate and deployment can fail under regional capacity constraints.

10. Microsoft Q&A, 2026: Persistent H100 (Standard_NC40ads_H100_v5) allocation failures on AML dedicated compute cluster in East US — 3+ days — no single H100 node allocatable for more than three days; one job queued about 45 hours before cancellation; support diagnosis of regional capacity unavailability with no published ETA; recommended workarounds of alternate region or alternate GPU SKU.

11. Google Cloud Community (discuss.google.dev), April 2024: Quota increase for NVIDIA A100 80GB GPUs is being declined — a request to raise an A100 80GB quota from 0 to 1 declined seconds after submission.