Gartner projects that most GenAI projects will blow their budgets by 2028, most home-built models will get abandoned, and the real cost is not where you think it is. Three numbers worth knowing before your next AI invoice lands.
Gartner's Generative AI Hype Cycle has three cost predictions that read less like forecasts and more like a warning label. Let me hand them to you straight, then tell you what to actually do about them.
One. More than half of GenAI projects will overrun their budgets by 2028. Gartner's stated cause is worth sitting with: it is not the model that blows the budget. It is poor architecture and thin operational know-how. The technology is not the problem. The way it gets built and run is.
Two. More than half of enterprises that built their own models from scratch will abandon them by 2028. Cost, complexity, and technical debt drag them under. All that effort, walked away from, because "build our own" sounded strategic in the pitch and turned into a maintenance sinkhole in reality.
Three. The money is not in training. It is in running the thing. Gartner projects that inference, the day-to-day cost of actually using a model once it is live, will be 70 percent or more of a model's total lifetime cost through 2028. Training is the down payment. Inference is the mortgage.
Why so many teams get this wrong
Because everyone budgets for the exciting part and underbudgets the boring part.
The exciting part is building the model and launching the pilot. It photographs well and it demos well. The boring part is running it, governing it, and maintaining it month after month, which is where the actual cost lives. This is the precise reason pilots look cheap and production looks brutal, and why so many AI initiatives stall out right after the celebration.
Here is the gap in one look:
| Where firms think the cost is | Where Gartner says it actually is |
| Building and training the model | Running it every day (inference) |
| The upfront project | The ongoing operations and governance |
| The tooling license | The architecture and the people to run it right |
How to not be in the 50 percent
Three moves, none of them exotic:
Design for the inference cost from day one, not after the invoice shocks you. Architecture decisions made early are what put you in the over-budget half or keep you out of it.
Be honest about buy versus build. Building your own model from scratch is the right call far less often than ego suggests, and Gartner is now projecting that most who try it quit. Unless a home-built model is a genuine competitive edge, you are usually buying yourself technical debt, not an advantage.
Bring operational discipline before scale, not after. "Lack of operational know-how" is the polite Gartner phrase for shipping something you were not set up to run. Governance, monitoring, and clear ownership are not overhead. They are what keeps the project alive.
The takeaway
The firms that win with GenAI are not the ones that spent the most or built the flashiest thing. They are the ones that architected for the real cost curve and had the discipline to run it properly. Everyone else is quietly funding the half of Gartner's forecast nobody wants to be in.
That is the unglamorous middle where Ayvede works: making sure AI gets adopted in a way that survives contact with an actual budget and an actual audit. Not the demo. The operation. Adoption with less risk, and without the surprise invoice.
Advise. Vet. Deliver. More at ayvede.com.
One thing to keep: in AI, the cheapest-looking project and the most expensive project are often the same project, twelve months apart. Budget for the mortgage, not just the down payment.