The AI Savings Mirage: Why Buyers Are Underwriting a Cost Structure That Doesn't Exist Yet

Two news items from the past few weeks deserve more attention than they've gotten from most in the M&A world.

First, Microsoft. The company that has spent over $100 billion partnering with OpenAI just told the 12,000 engineers in one of its biggest divisions to stop using Anthropic's Claude Code by the end of June.¹ The polite reason was "tool consolidation." The other reason, less polite, was that the bill kept getting bigger.²

Second, Uber. The CTO admitted to The Information that the company's entire AI budget for 2026 was blown through within a few months of the year starting. Engineers were using AI coding tools so aggressively that average monthly spend ran $150 to $250 per engineer, with heavy users hitting $2,000. The CTO himself burned $1,200 in a two-hour demo. His exact quote: "I'm back to the drawing board."³

Two of the most sophisticated buyers of AI in the world just told you, in plain terms, that AI costs more than they thought it would. If two companies with that much money and that much technical depth are getting their AI bills wrong, what should that tell a buyer of a $5M EBITDA business about leaning on AI savings to make a deal work?

The pricing isn't real

The AI pricing you see today isn't what these tools actually cost to run.

OpenAI is projected to lose around $14 billion in 2026 on roughly $13 billion in revenue.⁴ That's about $1.69 spent for every dollar earned. Anthropic, in a March 2026 court filing, disclosed cumulative lifetime revenue exceeding $5 billion against more than $10 billion in inference and training costs.⁵ The company has since reported sharp growth in its run-rate revenue, but the historical gap is real and the implication is the same: These companies are pricing below cost to capture market share, and venture capital and hyperscaler subsidies are covering the difference.

The repricing has already started. GitHub announced that all Copilot plans are moving to usage-based billing on June 1, 2026. The company's Chief Product Officer explained the change directly: "Today, a short chat question and a multi-hour autonomous coding session can cost the user the same amount."⁶ Translation: GitHub was eating the difference, and that's ending. Expect the rest of the industry to follow.

And the energy bill is going up

The cost of AI doesn't end with the inference invoice. The data centers running these tools have to be plugged in.

Per the IEA, data center electricity demand surged 17% in 2025, far outpacing the 3% growth in global electricity demand. AI-focused data centers are growing even faster.⁸ PowerLines, a nonpartisan consumer group, reviewed capex plans from 51 investor-owned utilities and found $1.4 trillion in planned spending through 2030, a 21% jump from last year. Data center demand was the top driver cited. PowerLines estimates residential customers will absorb roughly $700 billion of that through rate hikes, on top of a 5.1% bump the EIA projects for 2026.⁹

Then layer in oil. After the Iran conflict erupted in late February and the Strait of Hormuz effectively closed, Brent has spent months trading between $100 and $114 a barrel. About 8 million barrels a day remain offline.¹⁰ Goldman raised its 2026 forecast to $85, and at least one veteran oil analyst told Bloomberg the path to $150-$200 is open if the situation drags on.¹¹ US gas prices crossed $4 a gallon in May.¹²

Energy is the input behind nearly every cost a small business pays. Trucking. Plastics. Packaging. The electricity bill. The data centers that run the AI tools you might one day deploy. The cost structure you're buying today is more expensive than the one the seller built the business on, and there's no obvious reason that reverses anytime soon.

So picture the actual problem. AI is priced below true cost and that's starting to correct. Energy is structurally more expensive and constrained by physics in a way money alone can't fix. Buyers are still penciling in deals that need both of those things to break the buyer's way. Unless someone invents cold fusion overnight, the math doesn't work the way most people are hoping it works.

How this actually shows up in deals

Most buyers in the lower middle market don't add a line called "AI synergies" to their LBO model. The risk shows up somewhere harder to see.

It shows up in how much debt service the buyer is willing to swallow at close. A search fund buyer signs an SBA loan with a 1.2x or 1.3x debt service coverage ratio in year one. The deal looks tight on paper, but the buyer tells himself it's fine. He'll cut a little here, improve margins a little there, use AI to handle some of the back-office work the seller was paying employees to do. By year three, coverage will be 1.6x and he'll be sleeping at night.

That's the assumption that doesn't always survive contact with reality. Margins don't improve on the timeline you need. The AI tool that was supposed to save $40K a year on customer service ends up costing $30K a year and not actually replacing the headcount. Or the AI vendor raises prices 40% mid-year because they were losing money on every account and finally had to fix it. Meanwhile, the SBA payment doesn't budge.

The fallacy this resembles

Think about people who bought houses they couldn't quite afford in 2021. The mortgage payment was uncomfortable on day one. But they told themselves the house would appreciate, or rates would come down, or their income would grow. They underwrote the purchase based on a future state, not on the state they were buying into. Some of them ended up fine. Others didn't.

Buying a business assuming AI will save you money you don't currently have is the same logic. You are paying today's price for a cost structure you hope will exist tomorrow. If it doesn't, you're stuck with the cash flows the business actually generates, servicing debt and operating costs that were never going to be subsidized for you.

The right framing isn't "what will my margins look like once I deploy AI." The right framing is "can I service this debt and live on this business if nothing changes." If the answer is yes, AI tools become real upside. If the answer is no, you bought a problem and you're hoping a magic wand fixes it.

What diligence should actually ask

A buy-side QoE doesn't usually have a line called "AI sensitivity," but in 2026 it should be doing something close to it.

The questions worth asking before you sign:

• Can you service your acquisition debt at the seller's actual cost structure, not a pro forma one?
• If energy stays where it is or moves higher, what does that do to gross margin?
• If your AI vendors raise prices in the next 18 months, do the productivity gains still cover the spend?
• Are you paying a multiple of trailing earnings, or a multiple of imagined earnings?

If the deal only works with AI savings, the deal doesn't work. That's underwriting, not pessimism.

The bottom line

The buyers who get burned over the next two years won't necessarily be the ones who avoided AI. They'll be the ones who paid acquisition prices that assumed AI was going to be cheap, energy was going to behave, and inference costs were going to keep falling forever. Some of those bets might pay off. Many won't.

Buy what's actually in front of you. Stress test the deal at the cost structure that exists today. Treat AI as upside in your value creation plan, not as the thing that makes a marginal deal pencil. The downside of being conservative is leaving some return on the table. The downside of being aggressive is paying too much for a business that doesn't generate the cash flows you assumed.

A QoE that treats AI cost takeouts as proven savings is doing the same thing the 2021 mortgage broker did when he told the buyer not to worry about the payment. Don't let it happen to your deal.