Buy vs Rent GPUs — Cost Comparison by Stage
Rental list prices and provider terms checked 19 September 2026. Rental prices change without notice; this page is reviewed quarterly, next review December 2026.
Rent GPUs while the workload is still changing. Buy only when three things are true at the same time: utilization is high and predictable enough to clear the break-even point at the rent the buyer would actually pay, a data center with power and cooling for the hardware is contracted, and the buyer can name the end user, the installation address and the intended use. The third condition is the one most cost comparisons leave out. A purchase of data-center GPUs goes through end-user and destination screening before a supplier commits hardware. A rental does not, because the hardware stays in the provider's facility and was screened when the provider bought it. The renter is still screened as a customer. The difference changes when a project can start, and whether it can start as a purchase at all, not only what it costs.
Should a company buy or rent GPUs at its stage?
The answer follows the shape of the workload more than the size of the budget.
| Stage | What the workload looks like | Usually | Why |
|---|---|---|---|
| Exploration and prototypes | Models, frameworks and GPU types change from week to week | Rent on demand | There is no stable utilization to pay off hardware, and usually no site or documented end use that a purchase request needs |
| Scheduled training runs and fine-tuning | Large, time-bound bursts with idle weeks between them | Rent reserved capacity for fixed dates | Owned hardware sitting idle between runs is the most expensive outcome of the whole decision |
| Steady production inference | Measurable, predictable load | Run the break-even below | From here the answer depends on real utilization, the rent actually available, and whether a site is ready |
| Sustained training or inference at scale | Continuous use of large clusters | Buy, or a long-term dedicated contract | Both are procurement projects. Self-service reservations have ceilings; beyond them, rental is a negotiated contract |
| Data that must stay in a jurisdiction or network | Any | Buy, or dedicated capacity in that jurisdiction | The constraint is legal or contractual, not cost; the trade-offs are set out in cloud vs private AI |
How far self-service rental goes is published. On AWS, Capacity Blocks for ML reserve GPU instances for 1 to 182 days, bookable up to 8 weeks ahead, with up to 64 instances in one block and up to 256 across blocks (AWS documentation, checked 19 September 2026). Instances in a block begin terminating 30 minutes before its end time. Continuous capacity past those limits is agreed with a provider, which is itself a procurement exercise on the provider's side.
What does buying GPUs require that renting does not?
A renter and a buyer are screened as different things. The renter is screened as a customer of a service. The buyer is screened as the end user of controlled hardware going to a specific place.
| Question | Renting | Buying |
|---|---|---|
| Who is reviewed | The account holder, through representations in the provider's terms and the provider's own checks | The end user and every party to the transaction, including ownership and the ultimate parent |
| Installation address | Not asked: the hardware stays in the provider's facility | Must be known, and the same in every document |
| Right to use a data center, and its capacity | The provider's concern | A colocation contract or letter of intent for capacity, plus the site's power, rack density, cooling and network |
| Intended use | Accepted under the provider's acceptable-use terms | Described in writing and checked against the quantity and the site |
| Signed end-user statement | Not normally | Normally requested by the supplier, OEM or distributor |
| Export review of the hardware | Done when the provider bought it, for the provider's site | Done for this order, to this destination |
| Moving later | Another region, within the provider's terms | A new review; moving hardware to another party or country needs authorisation where the rules require it |
On the rental side, the check is a representation by the customer. The AWS Customer Agreement, section 11.6, has the customer represent that neither it nor any party that owns or controls it is on a sanctions or restricted-party list, naming the Entity List among others, and makes the customer responsible for complying with export and sanctions laws in how it uses the service (AWS Customer Agreement, last updated 14 August 2026, checked 19 September 2026). On the purchase side, the full list of what is asked, and why each item is asked, is in what a GPU end-user statement contains.
Renting does not change who may use the compute. It moves where the screening happens, not what it allows. BIS's industry guidance of 13 May 2025 on preventing diversion of advanced computing chips treats as a red flag an infrastructure-as-a-service provider that "does not or cannot affirm" that its users are not headquartered in China, and providers apply sanctions and restricted-party screening to their customers. A party or a use that could not be supplied as a buyer cannot be supplied as a renter either. How access to compute is reviewed for GPU cloud operators is covered in what if the buyer is not the end user.
How does screening change when GPUs can be used?
Renting starts when the provider has capacity. On-demand instances start when capacity is free in the chosen region. Reserved blocks fix dates in advance; on AWS, up to 8 weeks ahead, per the documentation cited above.
Buying starts with a sequence, and each step is dated by a different party. The project is qualified and the end-user package assembled; the supplier or OEM channel reviews it; supply is confirmed; the hardware is delivered and installed; the site is energized and the cluster accepted. None of these dates is set by the buyer alone, and no honest source can give one figure for the total before the supplier channel has been checked. What moves the supply part of it is set out in what determines GPU lead times.
Without a site, the purchase path does not start at all. The minimum information for a preliminary assessment includes the data center or hosting provider and the country and address of installation (what is needed first). A team that has not yet contracted data-center capacity is not choosing a slower option by buying. It does not yet have a purchase option. For that team, renting is the only way to have GPUs this quarter, and the comparison of costs comes later.
Which projects are not ready to buy yet?
These are gaps to close before a purchase request, not reasons to rent instead of being reviewed. Each of them stops a purchase request until it is resolved:
- no contracted data-center capacity, or a site without the power and cooling for the configuration;
- an end user that isn't named yet, for example an integrator bidding before its client has signed;
- an intended use too vague to assess, such as "AI compute";
- a quantity that neither the workload nor the site explains;
- capacity meant to be resold to clients who are not yet known, which is among the hardest cases to clear.
Which kinds of project clear most easily, and which are hardest, is set out in which projects clear most easily.
How is the buy vs rent break-even calculated?
The break-even is the number of GPU-hours after which owning has cost less than renting the same hours. It can be written without any market price, and every input comes from the buyer's own quotes:
Break-even GPU-hours = (P − V) ÷ (r − o)
- P: all-in purchase cost per GPU: the server's share, plus its share of the network fabric, storage, installation and spares;
- V: resale value per GPU at the end of the period the buyer plans to keep it;
- r: the rent per GPU-hour the buyer would actually contract for the same GPU;
- o: the cost of running an owned GPU per hour: power including cooling overhead, colocation, support and operations staff.
Divide the break-even hours by the hours in the period to get the utilization the owned hardware has to sustain. The table shows how that threshold moves when the purchase cost, expressed as hours of rent, changes. It assumes a three-year period (26,280 hours), running costs of 20% of the rent, and no resale value. It is arithmetic, not market data.
| Purchase cost per GPU, in hours of rent (P ÷ r) | Break-even GPU-hours | Utilization needed over three years |
|---|---|---|
| 5,000 | 6,250 | 24% |
| 10,000 | 12,500 | 48% |
| 20,000 | 25,000 | 95% |
The rent used for r decides the answer. The same GPU rents at very different rates on the same day, depending on the provider and the commitment:
| List price per GPU-hour, 19 September 2026 | H100 | B200 |
|---|---|---|
| AWS on demand, 8-GPU instance, Linux, us-east-1 | $6.88 (p5.48xlarge, $55.04 per hour) | $14.24 (p6-b200.48xlarge, $113.93 per hour) |
| AWS 1-year reserved, same instance | $2.97 ($23.78 per hour) | not listed |
| AWS 3-year reserved, same instance | $2.60 ($20.78 per hour) | not listed |
| Lambda on demand | $3.99 | $6.69 |
AWS rates are list prices as tracked by instances.vantage.sh (data updated 19 September 2026), divided by eight GPUs per instance; Lambda rates are from lambda.ai/pricing, checked the same day. For H100, the highest rate in the table is about 2.6 times the lowest. Comparing a purchase with on-demand hyperscaler rates, when the realistic alternative is a committed rate or a specialist provider, makes buying look far better than it is. AWS also raised its published Capacity Block rates by 15% in January 2026 (InfoQ, 15 January 2026), a reminder that r is not fixed over a three-year period either.
Power is the running cost buyers most often underestimate. NVIDIA lists the DGX B200, with eight GPUs, at about 14.3 kW maximum system power (NVIDIA, checked 19 September 2026), or about 1.8 kW per GPU at the wall before cooling overhead. Multiply by the site's PUE and electricity tariff to get the power part of o.
Resale value is the input with the widest error. It depends on how many newer generations have shipped by the time of sale. A calculation that only works with an optimistic V is a calculation that doesn't work.
For current server cost ranges, running cost by region and the same comparison worked through for H200, B200 and B300, see AI server cost in 2026.
When is renting the better choice?
Renting is the better choice, and no procurement or allocation work is needed, when any of these is true:
- utilization is below the break-even at the rent actually available, or can't be predicted yet;
- the model, framework or GPU generation is likely to change within the period the hardware would be kept;
- no data-center capacity with the required power and cooling is contracted;
- the project has an end date;
- the end user, the site or the use can't be stated yet;
- GPUs are needed sooner than a purchased cluster could be supplied, installed and energized;
- there is no team to operate hardware.
When is buying the better choice?
- sustained utilization clears the break-even at the rent the buyer could really contract, not at on-demand rates;
- a site is contracted and its energization date is known;
- the end user, the installation address and the intended use are clear and will not change;
- data or control requirements rule out shared infrastructure;
- the scale needed is beyond what can be rented on acceptable terms.
Leasing or financing the hardware does not change the screening. A lessor or lender may hold title, but the operator is still the end user, and the purchase path applies in full.
Why rent first even when buying is the plan?
A rental period produces the inputs that a purchase decision and a purchase request both need:
- measured utilization, which replaces the guess in the break-even;
- a precise workload description, which is what the intended-use field of an end-user package asks for and what makes a requested quantity explainable;
- a proven configuration, so the quantity ordered matches the workload and the site;
- capacity in the meantime, while the purchased hardware is reviewed, supplied, delivered and energized.
Rental history is not a document suppliers require. Its value is that the purchase request built on it is specific, and specific requests are the ones that can be assessed.
Frequently asked questions
Is it cheaper to buy or rent GPUs?
It depends on sustained utilization. Owning is cheaper once the hours used pass (P − V) ÷ (r − o): the purchase cost less resale value, divided by the difference between the rent and the running cost of owned hardware. Use the rent that could actually be contracted; on 19 September 2026 list prices for the same H100 ranged from about $2.60 to $6.88 per GPU-hour depending on provider and commitment.
Do you need an end-user certificate to rent GPUs in the cloud?
Not normally. A cloud provider screens its customers under its own terms, for example through sanctions and restricted-party representations. The end-user statement belongs to a hardware purchase: it ties the end user to an installation address and an intended use.
Can renting GPUs be used to get around export restrictions?
No. Renting moves where the screening happens, not what it permits. Providers screen customers, BIS guidance treats as a red flag a provider that cannot affirm its users are not headquartered in China, and a party or use that could not be supplied as a buyer cannot be supplied as a renter.
How long before rented or purchased GPUs can be used?
Rented GPUs can be used when the provider has capacity; reserved blocks on AWS can be booked up to 8 weeks ahead. For purchased GPUs there is no single figure before the supplier channel is checked: the total covers qualification, supplier review, supply, delivery and site energization.
Does leasing GPUs count as buying for screening?
Yes. The lessor or lender may hold title, but the operator is the end user, and the end-user, site and use questions apply as for a purchase.
What decides whether a GPU purchase request is supplied, and in what order it is reviewed: How NVIDIA GPU allocation works →
Final allocation and hardware availability remain subject to manufacturer/OEM/supplier approval, applicable compliance requirements and supply availability.
Rental rates on this page are third-party public list prices at the date shown, not Haink prices, and they change without notice. The break-even table is arithmetic on stated assumptions, not market data. This page is not financial or legal advice. Sources checked 19 September 2026: AWS Capacity Blocks documentation, AWS Customer Agreement (last updated 14 August 2026), instances.vantage.sh, lambda.ai/pricing, NVIDIA DGX B200 specifications, BIS counter-diversion guidance of 13 May 2025.