The Secret to Reducing Customers' AI Operating Costs: Apply Your Cloud Skills

Your clients' AI bills are about to become a problem. Not eventually. Soon. Maybe even now!

The five biggest hyperscalers are projected to spend $660 to $690 billion on AI infrastructure this year, according to Futurum Group's analysis of 2026 capital expenditure plans. Layer onto that what the model makers themselves are spending. Reuters reported in February that OpenAI is targeting roughly $600 billion in compute spend through 2030, against a projected $280 billion in revenue over that same stretch. So, AI getting cheaper for the customer certainly doesn’t seem to be on the cards.

It's already showing up in the number that matters most: gross profit margins. The same reporting, citing data from The Information, noted that OpenAI's own inference costs, what it costs the company to run the models that your clients use, quadrupled in 2025, pushing gross margin down from 40 percent to 33 percent.

You and your customers haven’t really felt this yet because AI investors have been subsidizing what you pay, keeping prices artificially low. But when the company selling the tokens is watching its own margin shrink, the subsidized pricing your clients enjoy today will not continue to be a permanent feature. It's a temporary one, funded by investor patience that's already wearing thin.

When that patience runs out, whether through a sharp correction or a slow squeeze, your clients' AI spend rises and rises fast. The MSPs who saw this coming will be the ones your clients call first. The ones who didn't will be explaining, after the fact, why nobody warned them.

You built a practice around that. Visibility dashboards, rightsizing, reserved instances, waste elimination. You turned cloud cost chaos into a retainer and, by so doing, turned bad news into tight controls.

AI cost management requires the same skills, just aimed at a different meter. If you can explain why a client's Azure spend spiked and where to trim it, you already understand the discipline this requires. The technology is different, but the management process isn't.

That's the next evolution this column suggests you lean into. You don't need to become an AI research shop. You need to apply the operational discipline that you already sell, which is cloud cost governance, to a line item that's about to matter a lot more to your clients than it does today. When the illusion gets stripped away, the remaining reality must be reasonable. That’s not going to happen unless you help your customers change much of what they do.

• Get visibility first. Famed management consultant Peter Drucker first taught us that you cannot manage something if you do not measure it and therefore cannot see it. Every client running any meaningful AI workload needs a usage breakdown: which model or models are used, which team owns the process, which tasks are being performed and what each one costs. Most clients have zero visibility into this today. That gap alone is your opening, and it’s a big one.
• Audit for model overkill. Most organizations default every task to the same expensive model, whether it's answering a two-line question or doing genuine complex reasoning. Map tasks to the cheapest model that the specific task requires. Learn to use any of the many routers that have emerged, which assign tasks even within the same process to different models depending on greatest need and lowest cost. This single move often cuts inference costs in half.
• Build a routing layer, not a single vendor dependency. If a client's entire AI workflow is hardcoded to one provider, they have no shock absorber when pricing shifts. Unless you include multiple models in the plan, they are at the mercy of that one provider’s pricing and policies. A routing layer that can send work to whichever model fits, cheap or premium, is the cost-control equivalent of not putting all your eggs in one proverbial basket, or cloud spend in one hyperscaler.
• Consolidate shadow AI. Just like shadow IT before it, individual employees and departments are signing up for the AI tools they like best on their own, outside any central visibility. Bringing that under one umbrella almost always reveals redundant tools and gives you leverage to negotiate better terms. Having unknown resources on your network can introduce exposures that can create almost unimaginable increased costs.
• Price this as a service line, not a favor. Avoid the use of that awful four-letter word that begins with “f”. Don't fold AI cost management into general advisory work for “free.” Baseline the client's current spend, show the waste in dollars, then price the ongoing management as a retainer or a percentage of savings. Clients will pay for this the same way they pay for cloud cost optimization, because the ROI shows up in the very next invoice.

Start this week. If you're not sure where to begin, start here, in this order:

1. Pull usage data on your three heaviest AI-using clients and find out what they're actually spending, by model, by team.
2. Identify one workflow per client that's running on an expensive model for a task that doesn't need it and swap it.
3. Ask each of those clients whether they know what tools their own employees have signed up for outside your management. They probably don't.
4. Draft one internal one-pager positioning "AI Cost Management" as a named service, not a bullet point buried in your AI offering.
5. Book fifteen minutes with your top client to show them the gap between what they're spending and what they should be spending. Let the number do the selling.

You already know how to do this work. You built the playbook during the cloud cost wars. The only thing new is the meter you're reading. The return this time will be far, far greater!

Posted by Howard M. Cohen on August 11, 2026