6-minute read

SaaS and vertical AI are getting squeezed.

Public market SaaS growth rates are slowing and the AI application layer is facing both margin pressure and competitive pressures from their AI suppliers and even their own customers.

As Ethan Mollick asked in May, with 5-9 year startup exit horizons, what paths remain for startups that won’t be disrupted by AI or eaten by the labs?

In this blog I’ve centered around a few ideas

While Smart Squeeze and Dumb Money were largely theoretic, it’s become empirically clear that AI is changing the physics of how and where value accrues to firms.

So if opportunity isn’t in the AI application layer, what is a durable strategy for company building?

One thing that’s clear: AI is raising the floor of what’s possible. While that creates challenges for past opportunities, it clearly unlocks new ones: namely those of greater complexity.

The SaaS era was easy mode.

In the AI era, durable businesses will be built on Hard Mode.

SaaS to zero Link to heading

SaaS is going to zero.

With software in its fast-fashion era –– e.g., Claude 4.5 Sonnet generating the entire Claude web app in 30 hours from just a prompt –– it’s hard to see how SaaS would produce durable rents.

It’s to the point where some are proposing a SaaS “nuclear option” where management teams reduce costs and load up on debt. Then they strip the org to a skeleton, use captured revenues to service the debt, and pray distribution sticks.

New York’s recent go with fast fashion didn’t last long – it seems unlikely that software’s turn will be here to stay either.

Vertical AI to zero Link to heading

I previously wrote how the build v buy decision of the app-layer customer squeezes the vertical AI application layer

  • from below: the app layer can’t charge less than what cognition costs
    • cognition prices are public making app-layer margins easy to calculate
    • there are no reports of wholesale pricing to app-layer companies
  • from above: the app layer can’t charge more than what’d cause their customers to build over buy

This squeeze worsens as

  • tokens get cheaper
  • models get smarter
  • coding tools get better

all of which are expected to happen. [You can see the Smart Squeeze model dynamics in action here.]

Apps integrating infra Link to heading

Worse, the market already expects disintermediation dynamics to play out in the B2B infra layer.

Supabase posted big gains from Lovable’s explosive use of its database. And while the forthcoming Lovable Auth is built with Supabase, if Lovable Cloud becomes big, will Lovable shareholders really choose to sacrifice its own COGS for a handshake with Supabase? TextQL founder Ethan Ding is skeptical.

The same dynamic is beginning to play out in the app layer, but slower, with e.g., law firms waking up to the observation

Harvey isn’t some breakthrough in legal AI—it’s ChatGPT with a law costume.

The labs are coming Link to heading

You also wouldn’t blame the labs for marching up the stack into the application layer. Chatbase’s Yasser tweeting

I am strongly suspecting that OpenAI is coming for the b2b application layer

If app layer companies are charging huge margins on lab tokens (from software cheaply generated or maintained by Codex or Claude Code “pls fix” statements), you’d forgive labs for wanting (or needing) a piece of the biggest app layer markets.

We’ve already seen labs move up the stack in enterprise search – it now seems OpenAI is coming for CRMs, with Monday.com and Hubspot down 15% over the past two days.

The once attractive properties of SaaS – its intelligibility and high margins – are now its biggest problems. SaaS is so easy to pull together that competition is coming from everywhere (even from indie players with no funding!) and distribution costs are exploding.

Those days of venture on easy mode are over.

Hard mode Link to heading

With AI squeezing easy software plays, maybe the path is to do something hard.

But what is hard in this AI era?

AI is clearly collapsing the talent stack - we’re all able to do much more with less.

One path that remains hard: real world complexity. For instance, follow Rabois

Formula for startup success: Find large highly fragmented industry w low NPS; vertically integrate a solution to simplify value product.

Keith Rabois’ now nearly 10-year old pinned tweet still feels expansive and durable. It describes the strategy of his DoorDash, Rippling, Airbnb and OpenDoor.

In each of these cases, the startup provides a unified experience

  • booking a stay
  • ordering food
  • buying a house
  • managing HR

over activities that previously required a fragmented collection of parties

  • unknown booking platforms
  • unreliable delivery service
  • complicated buying process
  • payroll, device management, insurance

These businesses are not hard for their technology but for their development and GTM. Most startups just do one thing – these startups aggregate many things, most of which are in the real world and not accessible by API.

Notably, none of these businesses face the margin pressures of the B2B application layer. They all benefit from AI and software getting cheaper. Their value stems from their trusted brands and integration of fragmented services – AI for them is more of a cherry on top.

As the realities of easy mode venture set in, the hard mode ventures – of hardware, consumer, marketplaces and deep-tech, that easy-mode VC eschewed – are now suddenly attractive.

Sometimes the best businesses are complex melting pots of all the things.

Kirsten Green shared this week on the launch of Oura Ring 4.

Following Kirsten, consumer hardware with integrated services may be the ultimate expression of hard mode: infrastructure, consumer trust, and integrated services all wrapped in a physical object that can’t be cheaply replicated with a prompt or ordered overnight from Shenzhen.

Hard mode clearly comes with its own tradeoffs: you trade SaaS margin compression for hard mode’s capital-intensity, inventory costs, and consumer adoption challenges. The bet today is that AI’s deflationary effects on software and its accelerating effects on hardware development make hard mode increasingly attractive relative to middleware software plays.

New economics of doing hard things Link to heading

The physics of the game have changed. The factory line of venture capital has shut down.

Anthropic hasn’t shared the prompt they used to have Sonnet 4.5 generate the Claude web app, but it seems directionally safe to bet that with improving models a lot of software is a prompt and a days worth of compute away.

You have this entire decade of [SaaS] companies and it turns out that it was completely wrong both on durability, margins, etc., … And who is the winner? It still ends up being the people have the fundamental hardware infrastructure.

Delian explained on TBPN in August. With the software layer becoming free, why pay a middleman rents for services you can spin up yourself to vertically integrate and capture the surplus yourself?

In this world, value accrues to the messy world of atoms, infrastructure, trust, and coordination.

In this age of AI, durable businesses either own the model layer or vertically integrate in the physical world - middleware gets squeezed from both sides.