Yevgeniy Matsulevych posted a line on LinkedIn last week that crystallized a conversation a lot of software-services founders are having quietly: “AI shouldn’t make your product cheaper. It should change your margin.” Most founders are doing the opposite math on a napkin. If a $30K MVP can now be built by one operator using Claude Code in three weeks, the client will demand the same work for $8K. Time to drop the price or shut the doors.

That math is wrong. Three numbers published this week explain why.

The moat moves up the stack again

Your client never bought your hours. They bought an outcome. Maybe the MVP they show to users, or the internal tool that puts an end to the copy-paste between five tabs. Sometimes it’s the dashboard that ends every Monday’s argument over which number is the real one. If that outcome was worth $30K last quarter, it’s still worth $30K. Their problem didn’t get smaller because your pipeline got faster.

What changed is your side. The cost of delivering it. You used to need a team of four, three months, calls, handoffs, all the friction of running a services shop. Now one strong operator with a well-built agent stack and backend help where needed gets surprisingly far in weeks. Not everywhere (deep backend is still deep backend) but for MVPs, internal tools, and workflow software, the math has flipped.

The moat is moving up the stack again. In the nineties, the relational database became a commodity and the moat moved to the ORM. In the 2000s, the server became a commodity with the cloud, and the moat moved to deployment and product. I’ve been in computing for 36 years, since 1990, age 15, on a Commodore 64, and I lived another version of this exact move that maps directly to what’s happening this quarter. When offshore software development opened up in the early 2000s, US and European shops did the same napkin math: if a developer on another continent costs five times less, how do we compete? Half closed. The half that survived stopped pricing the hour and started pricing the system: architecture, operations, product judgment. The hour rate dropped. The margin went up. The napkin was wrong then too.

Today is another version of the same move. Faster, because the layer being compressed is general reasoning, not junior labor. The friction that’s collapsing now is the provider’s internal coordination cost, not the wage of an entry-level coder.

Three numbers this week that confirm where the margin went

The first number came from OpenAI on Monday, May 11. They launched the OpenAI Deployment Company: a majority-owned OpenAI entity backed by $4 billion, with 19 investment, consulting, and systems-integration partners. Same day they announced the acquisition of Tomoro, which brings 150 Forward Deployed Engineers and Deployment Specialists on day one. This is not a new tool. This is a services-shaped organization, modeled on Palantir, attached to the model. If the thesis were “the model is the product and it sells itself,” this company would not exist. It exists because deploying AI inside a real enterprise is services work, and the lab figured that out.

Aaron Levie at Box extended the read the same day: every tech wave produces a larger services wave behind it, and the agents wave will be the biggest because agents rewire the underlying business process, not just the medium of software. Moving an on-prem CRM to a cloud CRM changed the delivery, not the flow. Moving to a sales agent changes who does each step of the flow. That difference doesn’t ship in a license. It ships with people inside the client, mapping the real process, designing the agent stack, measuring whether the agent got it right. That’s services. Not product.

The second number came from Anthropic. Linas Beliūnas summarized the curve: $10M ARR in December 2022, $1B by January 2025, $14B by February 2026, $45B by May 2026. Claude Code alone is at a $2.5B run rate after less than a year. Enterprise customers spending more than $1M a year doubled from 500 to 1,000 in two months. 80% of Anthropic’s revenue comes from enterprise, not consumer. The point of this number is not the size. It’s the composition. Enterprises are paying more for embedded workflow, not less. If AI were lowering the price of the outcome, this curve would slope the other way.

The third number came from Stripe. Linas’s Substack reported that Stripe is embedding one AI practitioner per 20 people across its marketing org. Not a pilot. Not a ChatGPT seat per person. A new full-time role, dedicated to redesigning how each individual works around a concrete agent. Aaron Levie and the Box job posting I covered yesterday in the harness as moat post point at the same role. The internal lab is becoming permanent because what it produces is something no SaaS ships in a box.

Read the three numbers together. The model lab enters the services layer. Enterprise revenue runs toward deep integration. The frontier company hires an embedded role per 20 people. This is not a story about prices falling. It’s a story about the delivery stack getting redesigned, with higher margin for whoever operates it well.

What this means if you run a B2B services shop

There are two threats and one opening. The obvious threat is OpenAI Deployment Co and the Anthropic consulting arm. They have the model, the capital, and now the services bench. The less obvious threat is the VC who passed last week on an AI-built startup with this note: the codebase was missing real architecture, lacked a security layer and offered no plan for scale. What looks fine in a demo and what ships to an enterprise procurement committee are two different products. Large companies are going to stand up dedicated audit practices for AI-generated code before letting it inside their systems. Software that doesn’t survive a technical review doesn’t get integrated.

The opening for a regional services firm sits between the two threats. You will not outprice OpenAI Deployment Co on the hour. You will not match its global scale. But you also won’t win selling weekend vibe-coded MVPs to enterprise buyers, because the buyer is learning the smell. The defensible band is operators with judgment, real context, model independence (Claude for one workflow, GPT-5 for another, Gemini for a third, and something new next quarter), shipping outcomes that survive an enterprise audit. Same time zone as the client. Real legal jurisdiction. Knows the local procurement cycle. The client isn’t buying the model. They’re buying the outcome delivered by someone who will be there in six months and answers the phone on a Tuesday at 3pm.

At IQ Source, we operate in exactly that band across Latin America. AI Maestro is our two-month discovery program where we map how the company actually operates (not how the handbook describes it), train the team for AI fluency, and deliver three artifacts: a real-process map, an AI opportunity score, and a prioritized recommendation. End of month two, there’s a Go/No-Go gate. What the client buys is not the hours inside the program. It’s the diagnostic outcome and the right to a second stage if the gate clears. If it doesn’t clear, the artifacts are theirs to keep. Technology Partner is the variant where we also build the product inside an already-redesigned delivery flow: the harness operates inside the development process, not next to it. I covered that layer in the runtime commodity, workflow moat post in April. This week’s post is the commercial side of that same technical thesis.

Two tests for your delivery model this week

The first is the pricing test. If you cut your quote by 60% because you saw Claude Code build what you were quoting, does your company still operate next quarter? If so, you’re not lowering the price. You’re correcting a previous overcharge. If not, you’re subsidizing the client out of your profit, and the person who pays that is the team you’ll have to let go in six months. The exit isn’t dropping the price. It’s dropping the cost of delivery and keeping the spread.

The second is the audit test. When the enterprise buyer at your next client opens the repo or the system you delivered, does what they see survive a serious technical review? Legible architecture, security layer, scaling plan, audit trail. If your delivery model only produces code that passes a demo, you’re in the band that VC was rejecting. If your stack produces something a technical committee signs off on in quarter one, you’re in the band that captures the enterprise revenue Anthropic is measuring.

The shop that wins the AI-native delivery model before yours does, charges the same price you charge today at half your cost. The shop that thinks the adjustment is to drop the price slowly eats its margin while waiting for the prior cycle to return. It is not returning. If you want to review where your real cost of delivery sits and which of the two tests you’re not passing yet, get in touch and we’ll walk it together in a one-hour intro call.