The Moat Doesn't Move. It Gets Taken.

Everyone's predicting the wrong consolidation. The axis isn't horizontal... and the winner isn't a software company.

David H. Friedel Jr./ 2026-06-16
Subscribe
MarketsAI
Listen to this post

The standard forecast goes like this. AI makes software cheap to build. Commodity SaaS gets crowded. Prices fall, the weak get bought, and the survivors expand sideways into adjacent categories until a few broad platforms own everything.

  • Shopify swallows payments and fulfillment.
  • HubSpot swallows sales and service.
  • Notion swallows your inbox. Horizontal consolidation.

We’ve all seen the chart. It’s the wrong axis.

The cheapening is real. But it doesn’t flow where the forecast assumes. It doesn’t flow outward to customers as lower prices, and it doesn’t reward the SaaS company that bolts on three more modules.

It flows upward… to whoever owns the engine that makes code cheap, the scarce talent that still matters, and the channel the buyer now runs on. The consolidation is coming. It’s just vertical. And the company at the top of it doesn’t sell software.

Here’s the argument the way it actually runs.

The dividend never reaches the customer

Start with the part everyone gets right: code gets cheap. The cost to train a frontier-class model has fallen from roughly $100M to single-digit millions to, in one 2025 academic reproduction, about thirty dollars. Building has collapsed in price, and the build-versus-buy decision is tilting toward build. An investor at One Way Ventures put it to TechCrunch that the barriers to creating software are now so low that companies increasingly build rather than buy. A team of three ships what used to take ten.

That’s the boring part, and it’s the part everyone is watching.

Here’s what they miss. The 10x isn’t free. Productivity that large doesn’t come from the tool; the tool is available to everyone, which is exactly why it commoditizes the output. It comes from the thing the tool can’t supply: taste. Judgment about what not to build. The ability to prune a thousand plausible features down to the three that matter.

That doesn’t get cheaper. It gets bid up… violently.

Look at where the money is going. Median engineer compensation at the top labs now runs $470K–$630K, with some packages past $900K, and elite hires reportedly into nine figures; OpenAI’s median total comp sits around $555K, and it has handed out retention bonuses as high as $1.5M to keep people from leaving. Meanwhile, the floor fell out from under everyone else: entry-level programmer employment dropped 27.5%, and new-graduate hiring at top firms fell more than 50%.

That is the dividend concentrating in real time. The people who carry judgment get NBA money. The juniors who would have been the next thirty hires never get hired.

And the winner doesn’t pass the savings on. They go the other way. Vendors are bundling AI into existing products and raising prices at renewal. The procurement firm Tropic calls it the “AI tax,” a 20–37% uplift on the way through. Salesforce moved enterprise customers onto a flat-fee agent bundle rather than cutting per-unit price. At the top of the market, prices are going up.

The compression is real… but it’s happening on the commodity floor, on a single measurable unit.

Customer-service resolution is the clean example. Zendesk launched outcome pricing at $1.50 per automated resolution in 2024; Intercom undercut it at $0.99 per resolved conversation; HubSpot dropped its customer agent to $0.50 per resolved conversation in April 2026.

The whole industry converged on one unit, the resolved ticket, inside roughly eighteen months, and is now racing it toward zero. That is what commodity compression looks like up close. It is not a gentle 30% glide. It’s a knife fight on the one layer where the work is interchangeable.

Underneath all of it, the pricing model itself is being rebuilt. Bloomberg estimates subscription-based pricing falling from about 60% of software toward 30% over the decade, with outcome-based pricing rising from 10% toward 60%.

The question stops being how much per seat and becomes how much per result, which sets up the part nobody has priced in.

The customer stops being a person

The buyer changes.

This is no longer a forecast. Gartner projects that by the end of 2026, roughly a quarter of enterprise software purchases will involve some form of AI-agent mediation. Enterprise agents already renew SaaS subscriptions, reorder supplies, and pay invoices on a company’s behalf. McKinsey puts agent-orchestrated spend at $3–5 trillion by 2030.

The buyer-side agent isn’t a thought experiment; it’s a line item.

For thirty years, SaaS moats were built for human psychology. Brand. Trust. Habit. The familiar dashboard. The switching cost that’s really just inertia… a person who doesn’t want to learn a new tool. The whole per-seat model rested on a human equation: more employees, more licenses. That equation is now wrong. Agents don’t need seats. They don’t log in with credentials, don’t accumulate a profile, and don’t show up in an admin’s license dashboard.

Replace the buyer with an agent, and every soft moat evaporates.

An agent feels no loyalty and has no habits. In agentic commerce, the buyer is a machine that reads structured data, not branding or imagery, and defaults to whichever offer is easiest to execute against clean, comparable terms. It will switch providers for a 1% advantage a human would never bother to chase, and Forrester already expects sellers to start fielding their own counteroffer agents in response, machine negotiating against machine.

“Build me a workflow for my HVAC business” doesn’t route to the brand the owner recognizes. It routes to whatever the agent scores marginally better this morning, and re-scores tomorrow. The entire emotional infrastructure of B2B software is worth nothing to a buyer with no emotions.

What survives an agent? Only the hard kind of moat. Data it genuinely cannot get anywhere else. A performance edge it can actually measure. Compliance and regulatory access it isn’t allowed to route around. An integration it physically cannot replicate. Sort your moats into “works on a human” and “survives an agent,” and watch how short the second list gets.

One honest caveat, stated once. The direction is now measurable; the endpoint is still a bet. Roughly 70% of consumers say they’re comfortable with an agent buying on their behalf, but only about 13% have actually completed an AI-referred purchase. The gap between comfort and completion is the whole timeline risk.

I think it closes, and faster on the enterprise side than the consumer side, because reordering supplies and renewing SaaS is exactly the low-ambiguity, repeat-purchase work agents are already good at. I’m making the bet. Everything below depends on it.

The friction doesn’t vanish. It moves up

The obvious objection: agents won’t really switch for 1%, because switching has costs… verification, integration risk, trusting an unknown vendor. True. Friction doesn’t go to zero.

But look at where it goes.

It moves off the vendor and onto the orchestrator… the layer that decides which API even gets called. And that layer is being built and owned by the model companies right now, in the open. Stripe and OpenAI shipped the Agentic Commerce Protocol; Google launched its Universal Commerce Protocol; Anthropic’s Model Context Protocol is the connectivity layer that lets agents reach live inventory and pricing.

The rails the buyer’s agent runs on are owned one level above every vendor on them.

And the vendor goes blind. When the purchase happens inside the chat, there’s no impression, no click, no session, no add-to-cart event… you can sell through the agent but you can’t see through it. Zero-click means the platform decides which products are even surfaced. The residual trust, the verification, the routing decision: all of it now lives with whoever owns the agent. And that owner does not route neutrally. It routes toward what it owns, what performs, what pays.

The friction that used to protect you is now the friction the orchestrator controls.

Follow the talent

Want to know where the value is going? Watch where the best people are going.

In July 2025, Meta stood up Superintelligence Labs and went hunting, packages reported as high as ~$300M over four years, hiring Scale AI’s Alexandr Wang and poaching a ChatGPT co-creator, Shengjia Zhao, out of OpenAI to run the science. The rivals poached back. The point isn’t the dollar figures. The lazy reading is that talent always chases the frontier, it does, and always has.

The signal is which layer they’re joining.

They’re not joining a SaaS company expanding horizontally. They’re joining the one entity that already owns the engine that makes code cheap and is building the commerce rails the buyer’s agent will run on.

Production and distribution, assembling the scarce judgment to aim both.

That’s the answer to the question every skeptic asks. We’ve heard “code is cheap now” before — open source, the cloud, no-code — and each time the number of software companies exploded instead of collapsing.

Why is this different? Because last time, the cheapening was a tool that diffused outward and enabled more companies. This time, the entity that owns the cheapening also owns the talent and sits on the channel. It isn’t handing you a tool and walking away.

It’s integrating the whole stack — production, judgment, and the buyer’s agent — into one layer.

The convergence

Three vectors. They point at the same actor.

  • Production: the model.
  • Talent: the hiring.
  • Distribution: the agent runs on ACP, UCP, and MCP, rails the model companies’ own, not you.

The company that holds all three doesn’t have to compete with you on features. It doesn’t have to expand sideways into your category. It sits above the buyer and decides whether your category gets called at all. The horizontal consolidation everyone’s drawing, platforms eating adjacent platforms, is the visible, slower, less important half.

The real consolidation is vertical, and it ends with the substrate itself, owned one layer up from anyone selling software.

What survives capture

Not nothing. But less than you were promised.

Below the orchestrator, there’s still a living: unique data, compliance, a genuine measurable edge. Agents already favor exactly these in the categories they’ve taken first… replenishment and repeat purchases decided on structured, comparable data.

But you hold those as a supplier to an agent, not as a destination a customer chooses. The agent calls you because you’re momentarily the best input… and drops you the morning you’re not. You keep the revenue. You lose the pricing power, the relationship, the loyalty, the brand equity, and, because the transaction happens inside someone else’s chat, even the ability to see your own customer.

Everything that lets a SaaS company compound now belongs to the layer above you.

The advice we’ve all been repeating — own the moat, own the substrate, own the customer relationship — assumed you’d be selling to a human who chooses you and stays. When the buyer is an agent and the substrate is being assembled one layer up, the moat isn’t a position you defend.

It’s a position that’s being taken.

Build for the layer that survives, or build to be the supplier of the thing above you can’t replace. Those are the two honest options. The third, keep shipping features to a human who’s already handed their buying decision to an agent, isn’t a strategy.

It’s a countdown.

Sources

Headcount and the labor effect

ThinkTree Substack

AI and the 2025 Hiring Freeze

2025 Hiring Freeze…

Read more

a year ago · ThinkTree

Pricing

The AI & Tech Society - Danar

AI Agents Are Breaking SaaS

How to price AI Agents…

Read more

2 months ago · 2 likes · 1 comment · Danar

Agent-as-buyer

The orchestrator layer

Talent flow

Back to the Journal