AI Services Grow, Prices Don’t Settle
AI services are growing fast, and almost nobody agrees on what they should cost.
Start with Omdia. In its September sentiment poll of channel partners, 44 percent expected third-quarter revenue to fall from a year earlier, and 41 percent expected it to grow. On this show, Omdia principal analyst Jessica Davis said the MSPs in that poll looked about the same as everyone else. Managed services overall are still growing. This isn’t a market shrinking. It’s a market splitting.
The fast growth is narrow. Omdia’s Robin Ody told partners in Barcelona that AI-related managed services are growing around 60 percent, against about 10 percent for traditional managed services. Davis said most of that comes from a small minority of organizations. She described what most MSPs are doing with AI as internal: more tickets resolved per technician, level-one work automated. What they sell to clients, she said, is still experimental. Her read on where the industry sits is “the 2008 part of it, not the 2012 part.” In plain terms, before the playbook.
Omdia’s Lisa Lawson put a size on that advanced group. She told Channel Dive that North American partners could be looking at nearly 150 billion dollars in AI services revenue by 2031. Only 20 percent of partners have the maturity to capture it today. Another 44 percent sit at a low level of maturity. On the buyer side, most enterprises report that fewer than one in five proof-of-concept projects has produced any real value.
Now look at what one MSP did with its price list. Manhattan Managed IT Services, in its own announcement, put “AI strategy and implementation” inside its Fully Managed plan. That plan starts at 150 dollars per user per month, alongside quarterly vCIO sessions. The AI work isn’t a separate line. It rides inside the flat fee.
So there are a few partners growing fast and most still experimenting. At least one has already folded the work into a fixed monthly number.
Nobody has settled what AI services should cost. The question is why it can’t be settled yet.
To see why, look at what sits underneath the price.
Why Today’s AI Price Is Borrowed
The price of AI isn’t a cost. It’s a decision the vendor makes, and the vendor can make it again.
Start with what a subscription actually covers. SemiAnalysis, which tests the usage limits on every major AI plan, calls subscription plans highly subsidized. In its rough model of Anthropic, subscriptions bring in about 10 percent of revenue but can take up more than 40 percent of the compute used to run the models. That gap is deliberate. The plans work as customer acquisition. And the limits move. By SemiAnalysis’s tracking, OpenAI cut the value of its 200-dollar plan in half and added a 500-dollar tier. While testing, SemiAnalysis also found one of three identical subscriptions with limits about 20 percent lower than the other two. The provider confirmed it was a small A/B test. SemiAnalysis’s conclusion is that providers can change subscription limits silently, at any time.
Usage-based pricing doesn’t escape this. On this show, Techaisle’s Anurag Agrawal explained what he calls token shock. The price per token is falling. But an agent doesn’t use one token. It uses thousands per task, because it loops, reasons and checks its own work. Consumption per task is rising faster than the price per token is falling, so the bill goes up. In Techaisle’s data, bill shock ranks as buyers’ top frustration with their vendors, at 42 percent.
So who pays the gap between price and cost? Investors. Gartner estimates AI providers will spend nearly a trillion dollars on infrastructure in 2026, and six trillion more from 2027 through 2030. To cover that, plus financing and a 12 percent return on invested capital, data center owners would need about 8.5 trillion dollars in revenue. Gartner forecasts no more than 3.6 trillion. For every dollar the buildout needs to earn back, Gartner sees about 42 cents. That’s why it puts more than 3.5 trillion dollars of that spending at risk of being written down. Gartner also expects more than 80 percent of the smaller AI cloud providers to stop operating independently by 2030. The price isn’t the only thing that can change. So can the vendor. The International Monetary Fund’s Global Financial Stability Report raised the same write-down risk from outside the industry. It warned that markets are exposed if the profits don’t arrive, and flagged circular deals where AI firms act as each other’s investors, financiers and customers.
In plain terms, today’s price is an introductory rate, paid by capital that expects a return it may not get, and nobody has said when the introduction ends.
Which means any price built on top of it is borrowed.
That borrowed price doesn’t stay with the vendor. It ends up in your contracts.
Price It Now, Write It to Reopen
For an MSP, the borrowed price lands in one place: the agreement you sign with a client.
On this show, Agrawal described why so many AI pilots never become production work. He listed four reasons, and he said the partner controls every one. The one that matters here is the last: no production price. The partner treats the pilot as a pre-sales cost and hopes the big deal shows up later. His line was that a pilot with no production price is a permanent free trial. His advice runs the other way. Sell data readiness as phase zero, and charge for it. Pick one narrow use case and commit to showing value within thirty days.
That cuts against the instinct to wait. When Jessica Davis was on this show, she said making a firm decision about AI pricing now would be “perhaps foolhardy.” She also said doing nothing at all is a mistake. Both can be true, and I’ve lived the version where they were.
Davis went back and read the columns I wrote for Channel Insider while I was still running an MSP. In 2011, I was worried that Microsoft would dictate the cloud pricing model, because it offered essentially one. Microsoft did set the first model. Then the market rewrote it. Part of the reason a client can buy Microsoft licensing on subscription through a partner is that the first answer didn’t hold. Anyone who had locked their business to that first model ended up renegotiating anyway, on someone else’s schedule.
The lesson isn’t that early prices are wrong. It’s that they’re temporary, and the agreement should say so.
So here’s the choice. Price every AI engagement now, and write the price to be reopened: a production price agreed before the pilot starts, a short term, and a stated trigger to revisit it when the vendor’s cost changes. Or keep giving the work away until somebody publishes the playbook. Or fold it into a flat fee that silently absorbs whatever the vendor decides next.
And how you write that choice decides who you get compared against.
Because the first AI quote you write becomes the anchor every renewal gets measured against. And the competitor it’s measured against is the shop that folded AI into a flat fee, which looks cheaper for exactly as long as the vendor’s subsidy lasts. So say the reopener out loud in the quote meeting: here’s the production price, here’s the term, and here’s what happens if the model vendor changes what it charges us. When the price moves, the flat-fee shop eats the margin or reopens a contract that never said it could. The separator isn’t the lowest AI price. It’s the agreement that survives the vendor’s next decision.
What to Consider
- Write the trigger, not just the term. Name the event that reopens the price: a published price change from the model vendor, a cut to the plan’s usage limits, or a client’s AI cost moving past a set percentage. A term alone tells the client when you’ll ask again. A trigger tells them why.
- Price phase zero. Charge for a data-readiness engagement before any pilot starts, and put the production price in the same document. The pilot then has a defined end, and the client has already seen the number the work becomes.
- Make the competitor’s contract the comparison. When a prospect holds up a bundled AI offer, ask what that provider’s agreement says happens when the vendor changes its price. If it says nothing, the client is buying the current subsidy, and that’s the comparison to put on the table.
If this trend continues: If Gartner’s math holds, AI data center owners face a gap of nearly five trillion dollars between the revenue they need and the revenue Gartner forecasts through 2030. The MSPs with a reopener in their AI agreements will reprice on their own schedule before then, and the ones on flat fees will reprice in a dispute.