March 2026
I had a conversation last week with a founder friend in Berlin. We were talking about what’s happening in the market right now, and I realized I’d been circling around a thesis for months without writing it down.
So here it is.
I think the AI landscape is consolidating into three defensible categories. Everything in between is getting squeezed.
The reason is structural. AI is collapsing the middle of the value chain. It can generate UIs, orchestrate workflows, and replace coordination labor. But it can’t generate proprietary data, physical atoms, or the trust that comes from actually delivering an outcome. So the value is migrating to the extremes: the infrastructure layer that everything runs on, the outcome layer where real work gets delivered, and the physical layer where software replication doesn’t apply. What’s left in the middle is the workflow and UI layer that used to justify SaaS pricing. That’s the part that’s dying.
Supabase. GitHub. Vercel. Postgres.
These are the companies that benefit most from the AI wave without being AI companies. Every developer using Claude Code or Cursor still needs a database, still pushes to GitHub, still deploys somewhere.
What makes this relevant to the thesis: AI isn’t replacing these tools, it’s driving more usage to them. Vercel reported that over 21% of its deployments now originate from AI agents, and customers using Claude represent just 1% of its user base but account for nearly 15% of all deployments. AI is generating more code, which means more repos, more builds, more database queries. The infrastructure layer grows as the application layer thins.
The catch: these markets are winner-take-all. GitHub crossed 180 million developers in 2025. PostgreSQL hit 55.6% adoption in the latest Stack Overflow survey. Supabase reached a $5B valuation with 55% of the latest YC batch running on it. If you’re not already the default, you’re not going to become it. Distribution compounds. Defaults are sticky.
This is where I think the most interesting opportunities live right now.
Sequoia calls it “service-as-a-software.” The inversion of SaaS. Instead of selling software to service providers, you become the service provider and replace the labor with AI. Pat Grady and Sonya Huang framed it as “attacking both software and services markets simultaneously, representing a profit pool at least an order of magnitude larger than previous technological transitions.”
Foundation Capital sized the opportunity at $4.6 trillion. That’s not the software market. That’s the total annual spend on labor and outsourced services that AI agents can absorb.
YC’s Jared Friedman put it more bluntly: “Suppose you think AI can automate legal work. You could build an AI product and sell it to law firms. Or, you start your own law firm, staff it with AI agents, and compete with existing firms. Instead of selling to the dinosaurs, you could make them extinct.”
Think about accounting. Our firm in the US is fully AI-native. They pull data from Stripe, Mercury, our bank accounts. They automate reconciliation, generate reports, validate with minimal human oversight. They don’t employ traditional accountants. They’re cheaper, faster, and more transparent than any firm I’ve worked with in Europe.
Are they going to be the global winner? Probably not. But the model is replicable across any compliance-heavy, regulation-specific vertical. Take Slovenian accounting. The market might be 100 million euros. Anthropic is never going to build workflows that comply with Slovenian tax regulations. But someone who understands the local legislation and builds an AI-native operation around it can probably capture a meaningful share. That’s a real business.
The defensibility isn’t technology. It’s operational knowledge, domain expertise, and the ability to sell outcomes rather than dashboards.
Sierra is probably the clearest proof point at scale. Bret Taylor’s company hit $100M ARR in seven quarters on pure outcome-based pricing. No per-seat fees. Brands pay for customer issues resolved. Taylor frames it as the natural evolution: boxed software became SaaS, and SaaS becomes paying for outcomes. “The atomic unit of AI productivity is a process, not a person.”
Hardware. Robotics. Physical infrastructure.
VC investment in robotics hit $22.2 billion in 2025, up 69% year-over-year. Twenty-seven physical AI startups raised $50M+ in Q1 2026 alone. The underlying logic: a humanoid robot that has learned fine motor coordination across thousands of hours cannot be copied with a single gradient update. Physical systems resist the easy replication that software faces from AI.
These categories also benefit from accelerating wealth concentration. Google DeepMind has offered top researchers up to $20M a year. AI-linked ventures minted over 50 new billionaires in 2025. When capital concentrates, demand for physical goods, premium experiences, and scarce assets grows. When software becomes commoditized, hardware differentiation matters more.
The challenge: longer development cycles, higher capital requirements, cofounders with deep domain expertise. But these companies are structurally protected from the “Claude can do that now” problem.
The most dangerous place to be right now is in the middle. Building a SaaS product that wraps a workflow in a UI.
Satya Nadella said it directly in early 2025: “SaaS as we know it is dead.” Garry Tan has been telling YC founders that “per-seat pricing is dying” because when one AI agent can do the work of ten people, you don’t need ten seats. And the data is showing up in real pricing decisions: Salesforce introduced $2/conversation pricing for Agentforce. Zendesk moved to per-resolution. Intercom did the same. Emergence Capital published a piece titled “AI’s Latest Victim: The Per-Seat Pricing Model.”
I’ll give you a concrete example from my own workflow: Apollo.
I pay for Apollo. I use Apollo. But the only thing I actually use is their data. I don’t touch their interface. I connect through their MCP, pull contact information into my own system, run my own enrichment, build my own workflows that map to my specific ICP, and interact with everything through a terminal.
I’m paying for one seat instead of twenty. I’m paying for credits, not software.
And here’s the counterintuitive part: Apollo has to enable this. They have to offer MCPs and APIs. Because if they don’t, I’ll switch to whoever does. I don’t want to click their buttons. Nobody does anymore.
This is a structural problem for every SaaS company that monetized workflow complexity. When the workflow lives in Claude Code and the data lives in APIs, the middle layer disappears. The UI becomes worthless. The per-seat model breaks.
The honest counterargument: Aaron Levie (Box) put it well: “AI doesn’t eliminate the need for software. It dramatically increases it. Every AI workflow needs guardrails, permissions, audit trails, and human-in-the-loop review.” And he’s not wrong. Enterprise systems of record in healthcare, financial services, legal, HR, anywhere with deep compliance requirements, proprietary data moats, and high switching costs, those are more durable than this thesis might suggest. Nobody is canceling Salesforce because of Claude. The total addressable “dying middle” is smaller than the loudest voices claim.
But the distinction still matters. Systems of record survive. Workflow wrappers don’t. If your product is “connect to Stripe, show a dashboard, send alerts,” an MCP-equipped LLM can replicate that today.
This is exactly how we think about what we’re building.
Creator marketplaces have been a graveyard for a decade. Dozens of platforms tried to connect brands with creators. Most hit a ceiling. Not because demand was lacking, but because the cost of coordinating humans at scale exceeded the value captured per transaction.
We’re not building a marketplace. We’re building service-as-a-software for creator operations.
When a brand needs 50 creators across 10 markets posting three times a week, the old model required an army of account managers doing manual sourcing, screening, briefing, content review, and payment processing. The cost scaled linearly with creator count.
We’ve compressed that. Intake is automated. Screening is standardized. Communication flows are templated. Performance signals are structured. Underperformers get cut automatically. Winners get scaled programmatically.
We don’t sell a dashboard. We sell creator operations as a service. Brands pay for outcomes, not access to a platform. That positions us squarely in category two: full-stack, AI-native, outcome-based. Defensible through operational knowledge and supply creation, not through UI or matching algorithms.
I don’t know exactly how things play out. Nobody does.
But the direction seems clear. Default infrastructure wins. Service-as-a-software wins. Physically defensible categories win.
The middle layer of SaaS, the workflow wrappers, the per-seat dashboard companies, is in trouble. Not because AI replaces them overnight, but because AI makes their coordination layer worthless. The value migrates to data on one end and outcomes on the other.
If you’re building right now, the question isn’t “how do I add AI to my product.”
It’s “which of these three categories am I in, and what happens if I’m in none of them.”