Home โ€บ Blog โ€บ The economics
AI Factory as a Service ยท Part 6 ยท The economics and the decision

Build, Install, Buy or Rent? The Economics of Getting an AI Factory

The price of a GPU-hour is the wrong number to compare. What matters is the cost of each hour a GPU does useful work, and that depends on how busy you keep it, how long it stays useful and what you pay for power, space and people. This final part puts current prices next to the costs of owning, lets you build your own break-even model, and turns the series into a buyer's scorecard.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Rental prices come from GPU price trackers and comparisons (CloudZero, Spheron, Ornn, Mercatus, DataStorage, Shattered), which are mostly aggregators and vendor blogs and disagree with each other, especially on B200. Colocation and electricity figures come from advisory firms and analyst reports. We found no independent study of where owning beats renting, so the calculator below is our own simplified model: every assumption is shown and editable. Not financial advice. See Reading the numbers.

1. The right question

A GPU you own costs roughly the same every hour whether it is busy or idle: the hardware is depreciating, the building and staff are paid for. A GPU you rent on demand costs nothing when you are not using it. So the comparison is not "owned price per hour" against "rental price per hour", but cost per useful GPU-hour: total cost divided by the hours that actually do work. Utilisation decides almost everything, which is why it came up in every part of this series and in CoreWeave's loan arithmetic in Case Study 7.

2. What renting costs

The price ladder

Indicative per-GPU-hour prices from trackers in 2026. Toggle the chip. Ranges are wide and change monthly.

The same H100 can rent for anywhere from about $1.49 to $6.98 an hour, a 4.7-times spread, with market medians around $2.95โ€“3.46. AWS's eight-GPU H100 instance works out to about $6.88 per GPU on demand but nearer $4 with commitments. B200 prices are more volatile: trackers in late 2026 range from about $5 to $8.40, and one index reported a 79% rise in three months as demand for the newest chips outran supply.

3. What owning costs

Where an owned GPU's money goes

Annual cost per GPU under the calculator's default assumptions (section 4). Hardware dominates.

4. The calculator

Cost per useful GPU-hour, four ways

Our simplified model, with every assumption editable. "Buy & run" = own hardware in colocation with your own team. "Install" = a cloud runs its hardware in your building (modelled as owning cost plus a management premium, minus most of your staff). "Reserved" = rent capacity for the full term. "On-demand" = pay only for hours used.

5. The break-even curve

Cost per useful hour vs utilisation

Uses the calculator's inputs. Owning gets cheaper per useful hour the busier you are; on-demand stays flat. Where the lines cross is your break-even utilisation.

6. Beyond cost

Most of the customers in Part 5 did not decide on cost alone. A ministry that must keep data in the country will pay more to do so. A hospital needs answers next to its scanners. A startup without a GPU team should not hire one to save money on paper. A telco may accept a long payback for a strategic business. And a cheaper route that locks you into one provider for five years may cost more later.

Sovereignty & controlOwning or installing keeps data and models on your premises; renting usually does not.
Skills & speedRenting starts in days and needs no cluster team; building takes years and a team.
Risk & lock-inOwners carry obsolescence risk; renters and installers carry lock-in and price risk.

7. Hybrid is normal

IDC's survey of more than 1,000 decision-makers (Part 1) found that most organisations expect to need both on-premises and public cloud for AI. In practice the split tends to follow the workload.

Which workload goes where

A common pattern, not a rule. Tap a workload.

Tap a workload.

8. Buyer's scorecard

The four routes, re-weighted your way

Scores 1โ€“5 are our editorial summary of Parts 1โ€“5. Set how much each criterion matters to you.

9. The series

Verdict: how to get an AI factoryIf your GPUs will be busy most of the time, your data must stay with you and you can hire the team, owning wins, through building, buying from a vendor or having a cloud install it. If demand is spiky, speed matters and you lack the skills, rent. Most organisations will do both. The decisive numbers are not the headline price of a GPU-hour but utilisation, useful life and the non-cost constraints: sovereignty, skills and lock-in. The same arithmetic runs from a three-rack Oracle region to Colossus; only the scale changes. For the giants, see AI Factories at a Glance.

10. Reading the numbers

Rental prices
H100: $1.49โ€“6.98 range, medians ~$2.95โ€“3.46; AWS ~$6.88 on demand, ~$4 committed (CloudZero, Spheron, DataStorage). B200: ~$5โ€“8.41 across trackers; +79% in three months per one index (Shattered citing SemiAnalysis); Mercatus average $6.39. Reserved discounts "up to 60%" are a vendor-blog claim.
Owning inputs
Colocation $200โ€“400/kW-month retail primary, $80โ€“157 wholesale (Encor and others); electricity ~$0.12โ€“0.14/kWh; ~1.76 MW per 1,000 H100s (Spheron). Hardware cost, useful life, staffing ratio, software cost and the install premium are our assumptions.
Model
Straight-line depreciation, no financing cost, tax, resale value or price changes over time. On-demand assumes capacity is always available when needed, which is not guaranteed for the newest chips.
Scorecard
Editorial.