Where AI’s Money Is Moving
This week, one of the biggest companies selling AI models asks the public markets to put a price on it, and the analysts who watch this channel are already asking how much of its revenue will last.
Start with Anthropic. Channel Insider reports it plans to sell 825 million shares at 70 to 78 dollars each, with pricing Wednesday and trading expected Thursday on Nasdaq. That’s a fully diluted valuation of up to 1.4 trillion dollars. Its filing shows 33.2 billion dollars in revenue over the first nine months of this year, and Channel Insider puts its run rate at roughly 81 billion dollars by the end of October.
Now, Omdia’s read on that revenue. At last week’s Canalys Forum, chief analyst Jay McBain said, “We think that 86% of that revenue, which is going to be over $100 billion, isn’t durable.” That’s more than 86 billion dollars he doesn’t think Anthropic can count on keeping. Omdia’s Ishan Dutt presented data showing open models less than six months behind the frontier and, in his words, “way, way cheaper.” The reporting doesn’t say how Omdia reached 86 percent, so treat it as a forecast, not a finding.
Then look at the services side. Omdia’s Robin Ody reports AI managed services revenue in North America growing about 60 percent a year, against 10.2 percent for traditional managed services. Fifty-six percent of partners now say they’re ready to deliver AI services; seven months earlier, 82 percent said they weren’t. Yet only one in eight MSPs earns meaningful revenue from AI or automation. And 84 percent of end users told Omdia they trust their MSP to implement AI.
Hear what that adds up to. The company selling the model is about to be priced as one of the most valuable businesses in the world, while analysts question whether its revenue lasts. The services wrapped around AI are growing nearly six times faster than managed services overall. And clients are already trusting their MSP with the work, before most MSPs are earning from it.
In plain terms, the money is moving, and it isn’t moving toward the part with the famous name.
To see why, look at who the biggest names in AI are paying.
The Job AI’s Giants Can’t Staff
Underneath all of this is one dynamic. A model can do almost anything in general and nothing in particular. It only becomes worth paying for when someone learns how a specific business runs and builds it into the work. That someone is scarce, and the model makers know it.
The industry’s name for that person is the forward-deployed engineer. Strip away the title, and it’s an internal consultant who can build: someone who sits inside a business, learns its processes, and wires the model into them.
Start with who’s paying for them. Channelholic reports that OpenAI, Oracle, Microsoft and AWS are collectively spending billions of dollars on forward-deployed engineers, and Rich Freeman puts the rate firms like Accenture and Deloitte bill for this kind of work at a thousand dollars an hour. One 40-hour week at that rate is 40,000 dollars. Omdia’s McBain says the model makers don’t actually want the business: “None of these companies that are investing billions of dollars want to be direct services companies.” In his view, these programs buy time until partners are trained to take over.
Now look at the supply. TechCrunch reports research from executive search firm Christian & Timbers. It counts roughly 17,000 forward-deployed engineers in the U.S., and only about 2,000 with the experience to consistently deliver a return. As the study puts it, “Not 2,000 available. 2,000 total.” The firm projects demand up 2,100 percent by the end of this year. It also found that the share of companies planning to hire these engineers went from 5 to 10 percent at the start of the year to 70 percent by the end of June.
And the companies that can afford them want to own them. Chris Taylor, who runs Ode with Anthropic, says he’s hearing the phrase “internal forward-deployed engineers” more often. TechCrunch reports companies are building their own teams to keep their processes away from AI vendors.
That’s why the money moves. When the model gets cheaper and the person who makes it useful gets scarcer, the value goes to the person.
In plain terms, the model is the raw material. The scarce part is the work of fitting it to a business, and the biggest names in AI are paying a premium to do that work themselves until someone else can.
For the small businesses MSPs serve, that work has to come from somewhere.
Small Businesses Need the MSP
For the MSP, the consequence is that small businesses will need someone to do this work for them, and it won’t be the vendor’s engineer.
Start with the small businesses themselves. Earlier this year, Goldman Sachs surveyed 1,256 owners in its 10,000 Small Businesses program. Seventy-six percent were using AI. Only 14 percent had fully integrated it into their core operations, and 73 percent said they’d benefit from more training and resources to implement it. Those are Goldman’s own program participants, not every small business, but the gap is hard to miss.
Those businesses aren’t getting a forward-deployed engineer. Omdia’s McBain says MSPs focused on small business “don’t have the operating model, operating layers, to do that,” and their clients would never pay those rates. His colleague Bill Dannenmaier adds, “By the time it hits SMB, there’ll probably be another word for it.” Jimmy Hatzell, CEO of Hatz AI, which sells its AI platform through MSPs, has a candidate for that word. Writing in ChannelPro, he argues the role “already exists outside of the world of AI, just under another name. It’s the MSP.”
Then look at what that work looks like when an MSP does it. On this show, I covered Lyra Technology Group’s plan to put more than 50 forward-deployed engineers inside its MSPs by the end of 2027. Managed Services Journal now reports what those engineers actually do. Lyra’s Deniz Sagnaklar compares it to security: small and midsize clients couldn’t afford a full-time security engineer, so MSPs stepped in. In one case, utility invoice reconciliation that was done by hand now reads the invoices and meter readings and produces the import file for the finance system. And she says “the FDE title isn’t what moves a client.”
Put those together. Small businesses need the deployment work, the vendors won’t come down to them, and the MSP already has the relationship. What nobody has worked out yet is how to deliver it at a small-business price.
So here’s the choice. Build your own version of the forward-deployed engineer: someone on your team who spends scheduled time inside a set of clients, learns how each one works, and builds the workflows, sold as a recurring service. Or keep reselling the model, and let your clients’ AI depend on a price someone else sets and a vendor engineer they’ll never get.
That choice comes with an objection worth taking seriously.
The strongest objection is that the tools will make this job disappear, because every vendor is shipping AI that’s supposed to set itself up. But the Goldman numbers already answer it: three in four of the small businesses Goldman surveyed are using AI, and only one in seven has it built into how the business runs. The tools aren’t missing. What’s missing is someone who knows the business well enough to put the tool to work, and that’s the job the MSP already has.
What to Consider
- Train for the business, not just the tool. The person doing this work needs the skills of a consultant: running a discovery interview, mapping a process, and measuring a workflow before and after. Vendor training teaches the model, not your client’s business, and the model makers are paying heavily to train partners to take over from their own engineers, so the first course on offer will usually teach one vendor’s model. Budget for it as the cost of building a service line, not as continuing education.
- Start from the process, not the platform. Pick two or three clients and find the job someone still does by hand every week, then build that workflow first and record the hours before and after. Sell what comes next as a monthly allocation inside each client, so the work recurs instead of ending with a project invoice.
- Keep the work independent of the model. Document the prompts, steps and data connections for every workflow you build, so you can change the model underneath without rebuilding. The model’s price is about to answer to public markets, and open models are close behind, so the workflow should be yours even when the model isn’t.
If this trend continues: By the end of 2027, the MSPs earning real AI revenue will be the ones that can name the workflows they built and run inside each client, while the ones that resold model seats watch those seats get repriced or replaced by cheaper models.

