Enterprises are buying AI infrastructure faster than they can measure what it costs them. That’s the core finding from a survey of 107 enterprises detailed in VentureBeat AI, and it describes a gap that’s about to get expensive for a lot of companies. Spending is accelerating. The ability to see and steer the economics behind that spending isn’t keeping pace.
What the numbers actually say
Most of the organizations surveyed run their AI on a familiar base: hyperscaler cloud plus model-provider APIs. No surprise there. That’s the default stack for anyone who started building in the last two years.
What stands out is where the next dollar is headed. According to VentureBeat AI, the next wave of investment is aimed at specialized compute that almost none of these companies use today.
So you’ve got two problems stacking on top of each other:
- Companies can’t fully account for the economics of the infrastructure they already run
- They’re budgeting for infrastructure they have zero operational history with
That’s not a procurement problem. That’s a visibility problem wearing a procurement costume.
Why the gap opened
AI spending grew up outside normal IT governance. It started as experiments, pilots, a team here running a proof of concept, a team there burning API credits on a chatbot. Nobody built cost attribution for that, because at pilot scale it didn’t matter.
Now it matters. Inference costs scale with usage, not with headcount or seats. A feature that works well gets used more, which costs more, which means your best product wins can quietly become your worst margin lines. Traditional cloud FinOps tooling wasn’t designed for per-token, per-request, per-model economics.
And the shift toward specialized compute makes it harder, not easier. Different accelerators, different pricing models, different utilization curves. You can’t measure what you’ve never run.
The three-year picture
Here’s where this goes if the current trajectory holds.
Year one: CFOs start asking harder questions. AI budgets that got approved on enthusiasm get re-approved on evidence, or they don’t get re-approved. Expect a wave of quiet pilot shutdowns not because the tech failed, but because nobody could prove the unit economics.
Year two: AI cost attribution becomes a standard line item, the way cloud cost management did around 2018. Tooling consolidates. Some of it comes from the hyperscalers, some from startups that are being funded right now on exactly this thesis.
Year three: Compute strategy becomes a board-level competency at any company where AI touches the product. The winners won’t be the ones who bought the most GPUs. They’ll be the ones who knew what each one earned.
What to do about it
If you’re running AI infrastructure inside a business, a few moves are worth making before the next budget cycle:
- Tag everything now. Every API call, every workload, every model, mapped to a team and a product feature. Retrofitting attribution is painful. Doing it before you scale is cheap.
- Measure cost per outcome, not cost per token. What does one resolved support ticket cost? One generated report? That’s the number your CFO can actually use.
- Pilot specialized compute small. If your next dollar is going toward hardware you’ve never run, run it on one workload first. Get real utilization data before you commit real budget.
- Set a kill threshold. Decide in advance what cost per outcome makes a workload not worth running. Then actually enforce it.
- Put someone in charge. Not a committee. One person who owns AI compute economics and reports numbers monthly.
The honest read
This survey isn’t a prediction from a single analyst with a track record to check. It’s a snapshot of 107 companies describing their own behavior, which makes it more useful than most forecasts. Self-reported spending plans tend to be directionally accurate even when the specifics drift.
What it tells us is straightforward: the industry is in a buying phase, and the measuring phase hasn’t started yet. Historically, that ordering doesn’t end well for the companies that go deepest fastest without instrumentation.
The fix isn’t spending less. It’s knowing what you’re spending on. Full details are in the original VentureBeat AI report.