> ## Content Index
> Fetch the complete content index at: https://www.sotkn.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# On-Premise vs Cloud for AI: Who Pays for the GPUs, and Where They Fit
- URL: https://www.sotkn.com/explained/on-premise-vs-cloud/
- Published: 2026-11-10T10:36:34.000Z
- Updated: 2026-11-10T10:36:34.000Z
- Description: On-premise means a company buys and runs its own servers; cloud means it rents them from a provider. For AI the choice is less about software than about who carries the cost of the GPUs and whether there is a room with enough power to plug them in.
- Author: Sean
- Tags: Versus, #desk-explained

EXPLAINED · Versus · Last reviewed: 10 Nov 2026 · Next review: after Dell’s and HPE’s next quarterly results and Synergy’s fourth-quarter 2026 cloud tally

On-premise and cloud are the two places a company can run its computing. On-premise means it buys the servers and runs them in a room it controls; cloud means it rents capacity from a provider such as AWS, Microsoft Azure or Google Cloud and pays as it uses it.

**In short.** On-premise means a company owns its servers; cloud means it rents them. For AI, the cloud is still where most new spending goes: cloud services grew 55% to $167.7 billion in July to September 2026, and for the first time more corporate work runs in third-party venues than in companies’ own data centres. But buying is back too. Dell booked $60.9 billion of AI server orders in one quarter, and HPE said enterprise AI demand more than doubled. Our read: the server makers and chip suppliers are paid either way. The stuck step is the room. A modern AI rack can draw 142 kilowatts, against about 11 in the typical company data hall, so much of the AI that is bought rather than rented still ends up in someone else’s building.

Why an investor cares: every AI server is paid for by someone. In the cloud, the provider pays up front and earns it back by the hour; on premises, the customer pays and carries the risk of the hardware sitting idle. Which way the market leans decides whose revenue grows first, the clouds’ or the server makers’. Our [guide to what a hyperscaler is](https://www.sotkn.com/explained/hyperscaler-explained/) covers the companies that rent; this page covers the choice between renting and owning.

## What is the difference between on-premise and cloud?

On-premise, often shortened to on-prem, means the company owns the hardware. It buys servers, storage and networking, installs them in a data centre or server room it controls, and pays its own staff, power and cooling bills. The cost is mostly up front and is then spread over the hardware’s life as depreciation.

Cloud means the company rents. A provider such as Amazon Web Services, Microsoft Azure or Google Cloud owns the servers and sells time on them, by the hour, by the amount of data, or, for AI models, by the token. The provider carries the cost of the hardware and the risk that it sits unused. Our [GPU-as-a-service Explained](https://www.sotkn.com/explained/gpu-as-a-service-explained/) covers the rental model for AI chips in detail.

|                           | On-premise                                            | Cloud                                                                            |
| ------------------------- | ----------------------------------------------------- | -------------------------------------------------------------------------------- |
| **Who owns the GPUs**     | The company                                           | The provider: AWS, Azure, Google Cloud, Oracle or a neocloud                     |
| **How you pay**           | Up front, then depreciation over several years        | By the hour, by the token, or on a reserved contract                             |
| **Who carries idle time** | The company                                           | The provider, unless capacity is reserved                                        |
| **Where the rack sits**   | The company’s room, or a colocation hall it rents     | The provider’s data centres, in regions you choose                               |
| **Speed to start**        | Months: order, delivery, power, cooling, install      | Minutes to weeks, if capacity is free                                            |
| **Data control**          | Full; data stays inside                               | Shared; set by contract and region                                               |
| **Paid first**            | Server maker, chip maker, power and cooling suppliers | The same suppliers, paid by the provider; the provider is paid back hour by hour |

Simplified. Many companies use both, and colocation sits in between: the company owns the servers but rents the room.

There is a middle option. In colocation the company owns its servers but rents space, power and cooling in a third-party building. That matters more for AI than for ordinary computing, for reasons the stuck-step section below explains. Our [hyperscale vs colocation comparison](https://www.sotkn.com/explained/hyperscale-vs-colocation-data-center/) covers the building side.

## Which is growing faster, on-premise or cloud?

Cloud, by a wide margin, and AI has sped it up. Synergy Research Group estimated that companies spent $167.7 billion on cloud infrastructure services in July to September 2026, 55% more than a year earlier, and $558 billion over twelve months. It said AI-specific services were about a third of the market, which is roughly $56 billion in the quarter by our arithmetic. Amazon held 26% of the market, Microsoft 19% and Google 16%.

The shift shows up in where companies say their work runs. In Uptime Institute’s 2026 survey of data centre operators and IT staff, third-party venues, meaning public cloud, colocation, hosting and software as a service, held 46% of corporate IT workloads against 44% in companies’ own facilities. Uptime said it was the first time third-party venues had the larger share. Corporate data centres’ own share fell from 58% in 2020 to 38% in 2026, a drop of 20 percentage points.

Uptime also found the move has limits. Respondents expect third-party venues to reach only 48% by 2028, and Uptime noted that companies have tended to overestimate how much they will move. Few large organisations, it said, are comfortable moving everything out.

## Is on-premise AI coming back?

Buying AI hardware is booming, but much of it is not on-premise in the usual sense. Dell said it booked $60.9 billion of AI server orders in the quarter to 31 July 2026, recognised $16.4 billion of AI server revenue and ended with a $95 billion AI backlog. That is about 3.7 dollars of new orders for every dollar shipped, and almost six quarters of shipments waiting, by our arithmetic. Dell raised its full-year AI server revenue guidance to $74 billion from $60 billion.

The buyers matter. On the call, an analyst noted that most of Dell’s AI server business had come from neoclouds and smaller cloud providers, and Dell’s chief operating officer said the mix had not necessarily changed, because it was still winning sovereign and large neocloud deals. In other words, a large part of the hardware boom is cloud capacity being built by someone other than the big three.

The enterprise signal is in the customer count. Dell said more than 6,500 customers now buy its AI servers, and 3,300 of them were added in the last three quarters, after eight quarters to reach the first 3,200\. That is about 1,100 new customers a quarter against about 400 before, by our arithmetic, and Dell said the acceleration was enterprise.

HPE told a similar story at smaller scale. It booked $2.4 billion of AI systems orders in its quarter to 31 July 2026, up more than 30% on the previous quarter, against almost $1.6 billion of AI systems revenue, and said enterprise demand more than doubled. Its chief executive said customers were running AI agents on premises to lower what they pay the public cloud for tokens. Both companies are also selling to cloud builders: HPE said that after the quarter it won a multi-billion-dollar inference server deal with a hyperscaler.

| Signal                            | Figure                                         | Period                  | What it says                                                          |
| --------------------------------- | ---------------------------------------------- | ----------------------- | --------------------------------------------------------------------- |
| **Cloud infrastructure services** | $167.7bn, up 55%                               | July–Sept 2026          | Renting is still growing fastest; AI services are about a third of it |
| **Dell AI server orders**         | $60.9bn; backlog $95bn                         | Quarter to 31 July 2026 | Buying is booming, but mostly by neoclouds and sovereigns             |
| **Dell AI customers**             | More than 6,500; 3,300 added in three quarters | To 31 July 2026         | The enterprise count is what is growing fastest                       |
| **HPE AI systems orders**         | $2.4bn, up more than 30% on the quarter        | Quarter to 31 July 2026 | Enterprise AI demand more than doubled                                |
| **Work in third-party venues**    | 46%, vs 44% in companies’ own sites            | 2026 survey             | First time cloud and other third-party venues lead                    |

From Synergy Research, Dell, HPE and Uptime Institute. Hardware orders and cloud revenue measure different things and are not added together here.

## Why can’t most company data centers host AI racks?

Because the rooms were built for a different load. Uptime found that the most common rack in its 2026 sample drew a little over 11 kilowatts, up from 9 a year earlier, and 7.8 kilowatts if a few very dense new sites are left out. Only 24% of respondents had any racks at all of 30 kilowatts or more, the usual lower bound of high density.

A current AI rack is in another class. Vertiv’s reference design for NVIDIA’s GB300 NVL72, a rack of 72 GPUs that works as one system, supports up to 142 kilowatts per rack and is built for liquid cooling. One such rack draws about as much as 13 typical racks, by our arithmetic, or about 18 if the dense outliers are excluded.

![On-premise vs cloud: an NVIDIA GB300 NVL72 rack at up to 142 kW against 30 kW high density, 11 kW and 7.8 kW typical enterprise racks](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/on-premise-vs-cloud-rack-power.png)

An AI rack against a typical room. Vertiv and Uptime Institute figures; ratios are Second Order arithmetic.

Fitting that into an existing room usually means new power feeds, new switchgear and new cooling, and sometimes a new building. That is why much of the AI that companies buy rather than rent still ends up in a colocation hall or a cloud region. Our [rack-scale AI Explained](https://www.sotkn.com/explained/rack-scale-ai-explained/) covers what is inside these racks, and our [guide to AI data center power](https://www.sotkn.com/explained/how-much-power-does-an-ai-data-center-use/) covers the load at campus scale.

## When does each one make sense?

- **Cloud fits work that is new, spiky or uncertain.** Training a model once, testing a product, or serving demand that may not last. The provider carries the idle time, and the company can stop paying when it stops using.
- **On-premise fits work that is steady, heavy and sensitive.** A model that answers the same kind of question all day, on data a bank, hospital or government cannot send out. Uptime’s 2026 survey found 68% of respondents named data security, and 55% regulation, as barriers to cloud, up from 60% and 44% in 2024.
- **Colocation fits companies that want to own the servers but lack the room.** They keep control of the hardware and data, and rent the power and cooling that their own buildings cannot supply.
- **Most large companies do all three.** Uptime calls this hybrid IT: each workload goes where its rules, cost and speed fit best.

Price is the hardest part to compare, because it depends on how busy the hardware is. A rented GPU that is used a quarter of the time is cheap; an owned one used a quarter of the time is expensive. Our [GPU rental price Explained](https://www.sotkn.com/explained/gpu-rental-price-explained/) and [reserved vs on-demand comparison](https://www.sotkn.com/explained/reserved-vs-on-demand-gpu/) cover the rental side, and our [capex vs opex comparison](https://www.sotkn.com/explained/capex-vs-opex/) covers how each shows up in the accounts.

## Who gets paid?

Our read: the suppliers are paid whichever way the customer chooses, and the choice mainly decides who carries the risk.

- **Chip and server makers.** NVIDIA, and server makers such as Dell and HPE, sell to clouds, neoclouds, governments and companies alike. Dell’s and HPE’s figures show both groups buying at once, and both ended July with record backlogs.
- **Cloud providers.** AWS, Azure, Google Cloud, Oracle and the neoclouds are paid back over time, by the hour or the token, after they have paid for the hardware. Their risk is capacity that sits idle. Our [AWS vs Azure vs Google Cloud comparison](https://www.sotkn.com/explained/aws-vs-azure-vs-google-cloud/) covers the big three’s numbers.
- **Landlords and power and cooling suppliers.** Whether the rack belongs to a cloud or a bank, it needs a powered, cooled hall. Colocation operators, and makers of [transformers, switchgear and cooling](https://www.sotkn.com/explained/who-makes-data-center-power-equipment/), are paid when that hall is built.
- **The customer’s own balance sheet.** Owning moves the cost from a monthly bill to an up-front purchase that is depreciated over years. The company gains control and, if the hardware stays busy, a lower cost per unit of work; it also takes on the risk that the hardware ages before it pays back.

![On-premise vs cloud explained: who pays for an AI server and when, from picking the venue, renting or buying, finding a powered room as the stuck step, to paying it back](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/11/on-premise-vs-cloud-who-pays.png)

Who pays for an AI server, and when. Second Order.

## Where is the stuck step?

Our read: a room with enough power and cooling. Money and software are not what stop most companies running AI on premises; the building is. A company that wants to own a few dense AI racks often has to find the power first, and the fastest place to find it is usually someone else’s data centre.

Supply is a second bind. Dell said demand for its traditional servers outstripped what it could supply, and HPE said supply constraints limited how much of its higher demand it could turn into revenue. Both ended the quarter with record backlogs. A purchase order for servers is not a running cluster until the hardware arrives and a powered rack position is ready for it.

The clock to watch is how fast company rooms are upgraded. Uptime’s share of respondents with any racks above 30 kilowatts rose from 19% in 2025 to 24% in 2026\. If that share keeps climbing, more AI can run where the data already sits; if it stalls, the work keeps flowing to colocation and the cloud.

## What would change the view?

- **Enterprise rooms upgrade faster.** If many more companies add liquid cooling and high-density power, owning becomes easier and more AI moves in-house.
- **Cloud prices fall.** If GPU and token prices drop as new capacity arrives, renting becomes cheaper and the cost case for owning weakens.
- **Rules tighten.** More data residency and sovereignty rules would push sensitive AI toward company rooms and in-country providers.
- **Hardware ages faster.** Each new GPU generation shortens the useful life of the last one. That hurts the owner most, whether it is a cloud or a company.

## Common misconceptions

- **People assume the server boom means companies are leaving the cloud.** Actually a large share of AI servers are bought by neoclouds and other cloud builders, and cloud services are still growing faster.
- **People assume on-premise is always cheaper.** Actually it is cheaper only if the hardware stays busy; idle owned GPUs still cost money.
- **People assume any data centre can take an AI rack.** Actually the typical rack position was built for about 11 kilowatts; a GB300 NVL72 rack can need 142.
- **People assume colocation is the same as cloud.** Actually in colocation the customer still owns the servers; it rents only the room, power and cooling.

## FAQ

**What is the difference between on-premise and cloud?**  
On-premise means a company buys its own servers and runs them in a room it controls. Cloud means it rents computing from a provider and pays as it uses it. The work can be the same; who owns the hardware and who carries its cost is different.

**Is on-premise cheaper than cloud for AI?**  
It can be for steady, heavy use, because the company pays for the hardware once instead of paying a provider’s margin every hour. Dell and HPE both told investors in September 2026 that running AI on premises can win on cost per token for the right workloads. For spiky or experimental work, renting is usually cheaper because idle hardware still costs money.

**Why do companies still keep AI on premises?**  
Control. In Uptime Institute’s 2026 survey, 68% of respondents named data security and 55% named regulation as barriers to using the cloud, up from 60% and 44% in 2024.

**Can a normal company data center run AI servers?**  
Often not without work. Uptime found the most common rack in its 2026 survey drew about 11 kilowatts, while a reference design for an NVIDIA GB300 NVL72 rack allows up to 142 kilowatts and liquid cooling. Only 24% of respondents had any racks at 30 kilowatts or more.

**What is hybrid cloud?**  
Running some work on premises and some in the cloud, and moving it between them as needed. Uptime describes this as the normal pattern now: companies place each workload where its rules, cost and speed fit best.

**Is colocation on-premise or cloud?**  
Neither, strictly. In colocation a company owns the servers but rents space, power and cooling in someone else’s building. Uptime counts it with the cloud as a third-party venue; the company still carries the hardware cost, as it would on premises.

**Is this a list of stocks to buy or sell?**  
No. This page compares two ways of running computing and where the money goes. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research.

Related: [what a hyperscaler is](https://www.sotkn.com/explained/hyperscaler-explained/); [cloud regions](https://www.sotkn.com/explained/cloud-region-explained/); [GPU-as-a-service](https://www.sotkn.com/explained/gpu-as-a-service-explained/); [neoclouds](https://www.sotkn.com/explained/neocloud-explained/); [sovereign AI](https://www.sotkn.com/explained/sovereign-ai-explained/); [GPU depreciation](https://www.sotkn.com/explained/gpu-depreciation-explained/).

## Stack doors

Layer: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/). Also: [Neoclouds](https://www.sotkn.com/stack/neoclouds/) and [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Markets: [US](https://www.sotkn.com/markets/us/). Analysis: [where does the stuck step move](https://www.sotkn.com/analysis/where-does-the-stuck-step-move/).

[Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes on where AI computing gets built and who pays for it.

Not investment advice. Do your own research.

**How we check this**

Rack densities, workload shares and cloud barriers come from Uptime Institute’s Global Data Center Survey 2026, keynote report 209, published in July 2026; they are survey answers from data centre operators and enterprise IT staff, not a census. Cloud market size, growth and shares come from Synergy Research Group’s release of 29 October 2026, which is an estimate built from company reports. Dell’s orders, revenue, backlog and guidance come from its results release of 1 September 2026, and its customer counts and customer-mix remarks from the call the same day, as transcribed. HPE’s revenue comes from its release of 2 September 2026, and its AI orders, AI revenue and enterprise remarks from the call the same day, as published by HPE. The 142 kilowatt rack figure is the ceiling of Vertiv’s reference design for the NVIDIA GB300 NVL72, announced on 11 June 2025; real racks vary with configuration. Ratios, such as how many typical racks one AI rack equals, and the rough size of AI cloud services, are our arithmetic, checked in code and labelled as ours on the page. Who gets paid and the stuck step are our read. Hardware orders and cloud revenue measure different things, and we do not add them together. The featured photograph is credited in its caption. This page compares two ways of running computing. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research.

**Last reviewed:** 10 Nov 2026 · **Next review:** after Dell’s and HPE’s next quarterly results and Synergy’s fourth-quarter 2026 cloud tally

Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here.

[Our editorial standards →](https://www.sotkn.com/editorial-standards/)

## References

1. [Uptime Institute — Global Data Center Survey 2026, keynote report 209 (July 2026)](https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/UptimeInstitute.GlobalDataCenterSurvey.2026.pdf?ref=sotkn.com) — Modal rack density above 11 kW, 7.8 kW without a few super-dense sites; 24% of respondents have any racks of 30 kW or more; third-party venues 46% of enterprise IT workloads against 44% in enterprise-owned sites, the first time ahead; corporate data centres 58% of workloads in 2020, 38% in 2026; data security cited as a cloud barrier by 68%, regulation by 55%; 48% third-party expected by 2028
2. [Synergy Research Group — Cloud market growth rate surges to well over 50% per year; Q3 revenues reach $168 billion (29 Oct 2026)](https://www.srgresearch.com/articles/cloud-market-growth-rate-surges-to-well-over-50-per-year-q3-revenues-reach-168-billion?ref=sotkn.com) — Cloud infrastructure services $167.7 billion in Q3 2026, up 55%; trailing 12 months $558 billion; AI-specific services about a third of the market; Amazon 26%, Microsoft 19%, Google 16%; neoclouds among the fastest-growing tier-two providers
3. [Dell Technologies — Second quarter fiscal 2027 financial results (1 Sep 2026)](https://investors.delltechnologies.com/news-events/press-releases/detail/1653/dell-technologies-delivers-second-quarter-fiscal-2027-financial-results?ref=sotkn.com) — AI server orders $60.9 billion, AI-optimised server revenue $16.4 billion, AI backlog $95 billion; traditional servers and networking $10.5 billion, up 122%; full-year AI server revenue guidance raised to $74 billion from $60 billion
4. [Dell Technologies — second quarter fiscal 2027 earnings call transcript (1 Sep 2026)](https://www.fool.com/earnings/call-transcripts/2026/09/08/dell-dell-q2-2027-earnings-call-transcript/?ref=sotkn.com) — More than 6,500 AI customers, 3,300 added in the last three quarters; demand broadening across neoclouds, sovereigns and enterprises; AI server mix still led by sovereign and large neocloud wins; on-premises called attractive on token economics for the right workloads; supply constrained
5. [HPE — HPE reports fiscal 2026 third quarter results (2 Sep 2026)](https://www.hpe.com/us/en/newsroom/press-release/2026/09/hpe-reports-fiscal-2026-third-quarter-results.html?ref=sotkn.com) — Revenue $12.2 billion, up 34%; server revenue $6.8 billion, up 35.3%; record orders and backlog; full-year and fiscal 2027 outlook raised
6. [HPE — third quarter fiscal 2026 earnings call transcript (2 Sep 2026)](https://investors.hpe.com/~/media/Files/H/HP-Enterprise-IR/documents/q3-2026/q3-2026-transcript.pdf?ref=sotkn.com) — AI systems orders $2.4 billion, up more than 30% sequentially; enterprise AI demand more than doubled; AI systems revenue almost $1.6 billion; backlog up 14% to a record; customers running agentic AI on premises to lower public cloud token costs; multi-billion-dollar inference server award from a hyperscaler after quarter end
7. [Vertiv — Cooling and power reference architecture for the NVIDIA GB300 NVL72 platform (11 Jun 2025)](https://www.vertiv.com/en-us/about/news-and-events/corporate-news/vertiv-develops-energy-efficient-cooling-and-power-reference-architecture-for-the-nvidia-gb300-nvl72/?ref=sotkn.com) — Reference design for rack densities up to 142 kW for GB300 NVL72; liquid cooling-ready; work with NVIDIA on 800 VDC power for 1 MW racks from 2026

S

**[Sean](https://www.sotkn.com/author/sean/)**

Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck.

Covers AI infrastructure, semiconductors, and second-order stack reads.

[Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)