When Should a Company Stop Renting AI and Start Owning It?
Buying AI by the request works fine for experiments. Once it becomes a daily business operation, the maths can flip completely.
Key points
- Deloitte's 2026 State of AI in the Enterprise report found the share of companies with at least 40 percent of their AI projects in production is expected to double within six months.
- Worker access to AI rose 5 percent in 2025, according to the same report, signalling a shift from pilot programmes to always-on use.
- The crossover point, the level of sustained use at which owning AI infrastructure can cost less than paying per request, varies by workload and has no universal figure.
- Cloud providers have cut token prices repeatedly, meaning the buy-versus-rent calculation is a moving target rather than a one-time decision.
Sponsored content published by MIT Technology Review, paid for by Hewlett Packard Enterprise (HPE), raises a question many business owners haven't reached yet: when does renting AI by the request become more expensive than building your own capacity?
It's a fair question. The answer is more complicated than a simple cloud-versus-data-centre debate.
What does "renting versus owning" AI actually mean?
Right now, most companies pay for AI the way they pay for metered electricity: send a request to a cloud service, it processes the request, and the bill reflects how many "tokens" (chunks of text or data) were handled. Flexible and low-commitment, that model works well when usage is small and unpredictable.
Owning AI capacity means buying or leasing physical hardware, GPUs (specialised chips that handle the heavy number-crunching AI requires) and servers, then running AI models on them directly or through a managed service. Fixed cost up front, but potentially cheaper per unit of work once volume is high enough.
The HPE-commissioned piece puts the crossover plainly: when usage becomes steady enough to keep capacity busy most of the time, ownership can beat per-request billing on both cost and predictability. When usage is still lumpy, renting wins.
Who is actually at this crossover point?
More companies than you'd expect. The Deloitte figure is striking: the share of businesses with at least 40 percent of their AI projects in full production could double in roughly six months. Customer-service agents, internal research tools, automated business-process systems running continuously. That's not tinkering.
AI agents, software that carries out multi-step tasks on its own, are a big part of this shift. A single automated workflow might call an AI model dozens of times to complete one task. Multiply that across thousands of employees and the per-request bill compounds quickly.
Our coverage of Oracle's AI cloud results on 17 September found the company had taken on over $100 billion in debt to build data-centre capacity, a signal of how seriously enterprises are pricing this bet. Infrastructure spending is already moving.
Three questions every business should ask before committing
The HPE piece offers a practical framework, and it's genuinely useful regardless of who paid for it:
| Question | What you are really asking |
|---|---|
| Is demand steady and predictable? | Can you forecast usage 12 to 18 months out? |
| At what usage level does ownership pay off? | Have you modelled your actual workloads, not generic benchmarks? |
| Can you keep the capacity busy? | Do you have the governance and adoption plan to fill it? |
The third question is the one most planning documents skip. Hardware sitting underused is just expensive hardware. The economics only work if someone is responsible for finding new workloads to run on it.
Should you trust a framework a hardware vendor paid for?
Honestly, the framework holds up. But the piece skips something important: cloud providers have cut token prices repeatedly and keep doing so. Every time a major provider drops prices, the ownership calculation shifts. That's not a reason to ignore the framework. It's a reason to rerun the numbers annually rather than treating a one-time model as settled.
If AI has moved from experiments into daily operations at your company, pull the last three months of AI spend and check whether usage was consistent or spiky. Consistent usage at meaningful scale is your signal to at least run the ownership numbers. Spiky and still changing? Keep renting for now.
