Is Software Dead? No. It Just Got a Lot Harder to Win. The SaaStr AI Deep Dive with Rory O’Driscoll

You already know Rory O’Driscoll from our weekly 20VC x SaaStr podcast, where he, Harry, and I go back and forth on exactly this stuff. He’s also been investing in software for 30+ years. Long enough, as he put it on stage at SaaStr AI, to have stayed employed but not so successful he could retire and buy a football team. So when the whole industry starts asking whether software is dead, he takes it a little personally. Then he does what a good investor does: narrows the question until it’s answerable.

His answer, after weeks of his firm arguing about the numbers every Friday: software isn’t dead. It got a lot harder to win, but it isn’t dead. The math is below.

AI Is in “Invest Mode,” and the Gap Is Enormous

One fact colors every other decision right now. In 2026, the hyperscalers alone are spending roughly $688B on AI capex. Think of that as all the money going in to make AI happen.

What’s coming out the other side is about $110B in revenue. Around $89B of that comes from the two foundation model leaders on a GAAP basis. Round up everything else and you get another $20-30B.

So the industry is spending on the order of half a trillion dollars a year more than it’s taking in.

Rory’s framing: roughly right beats exactly wrong. Argue about $10B here or there and nothing changes. The shape is a massive bet funded years ahead of the revenue, and every strategy has to account for it.

It Takes Until ~2032 for Revenue to Catch the Spend

Run the revenue estimates forward and the crossover, where the two model leaders finally out-earn cumulative capex, doesn’t hit until roughly 2031-2032, at around $1T of revenue.

Two takeaways for founders from that timeline:

One, this goes on for years. We’re five or six years from being out of invest mode. That’s a long runway of capital pouring in ahead of returns, and a long runway of opportunity on top of cheap, improving models.

Two, expect a hiccup. When you’re spending far more than you’re taking in, someone eventually wakes up and asks why. Even if the long-term story plays out, don’t be shocked if the market takes a hard breath and pulls back before 2032. Plan your burn like that moment is coming, because it might.

Why This Is Even Possible: The Knowledge Worker Wage Bill

For the foundation models to earn a trillion dollars, the money has to come from the largest pot there is: the knowledge worker wage bill.

To hit these numbers, AI has to take an appreciable slug of it. If all the spend landed in the US, you’d be talking about something like 15-17% of all knowledge worker dollars. For software developers specifically, well north of 25%. Put concretely: for every $200,000 developer, roughly $50,000 going to tokens.

That’s the bet American capitalism is making. The scaling laws showed that money in produces better models out. The ChatGPT moment showed people actually want to use it. Combine those two and the capital keeps coming. But it only pencils if AI eats a real share of what is currently paid to humans. Possible, maybe even likely, but not without bumps.

The Stack: “Making AI” vs. “Using AI”

Borrow the model Jensen uses: energy, chips, infra, models, apps.

The bottom three layers are all about making AI. This is the stuff it takes to build $688B of capex, and roughly 80% of it isn’t venture-backable. Half of it is chips, and the vast bulk of those dollars went to one company. Capex went from about $200B four years ago to about $600B today, which means someone is pocketing roughly $400B a year, mostly Nvidia. Energy is a big, hard block but a relatively small dollar amount. Infra ranges from plain buildings up to the labeling companies and RL training gyms the model companies buy.

The top layers are using AI: how software takes a model and turns it into value for a business customer. That’s where the rest of us live.

The money question: how does enterprise want to consume ~$1T of AI over the next five or six years? Do they buy it all direct from the foundation models, in which case we all go home and rent a very small conference room? Or is there room for a whole set of companies to buy AI from the model providers, add real value on top, get close to a specific customer, and ship an economically differentiated product?

Rory’s read, and mine: there’s room. What defends that room is the next question.

The Harness Is the New LAMP Stack

There’s a term going around, “the harness,” for the software layer on top of a raw model that turns it into a dependable system. Deciding what context the model sees, allowing actions, selecting outputs, logging what happened, escalating and managing models by task.

Rory doesn’t love the word, and neither do I. The better mental model: for 15 years we built B2B software on the LAMP stack. This, or something like it, is going to be how you build every application in the age of LLMs.

Every SaaS app ran on the same LAMP stack, and they were not remotely the same app. Same thing here. Most AI apps get built the same way, reasoning over a model and distilling it into something a business user can use, and they’ll all describe themselves the same way. But how they get instantiated varies enormously depending on whether you’re a legal product, a customer support product, a coding product, or something with hardware attached. Sameness of architecture is not sameness of business.

The Moats That Actually Hold

The real question underneath all of it: what can Claude or ChatGPT roll over, and what can’t they? Rory’s firm has been sorting this into moats. The ones that hold:

**Software plus sensors.**If your product combines vision, touch, or other real-world sensors with reasoning to drive a business outcome, the foundation models aren’t going to start shipping sensors. Eminently defensible.**Marketplaces and network effects.**If you’ve built a platform where every side wants to be there and more users make it more valuable to everyone, a single person with a Claude Code app can’t recreate that. A recruiting marketplace they recently backed is a clean example.**Proprietary, non-public data.**If your business is built on data the model has never seen and can’t see, you don’t even have to explain the moat. The model simply can’t do what you do.**Full-stack.**Instead of selling software to a business, you become the business. Rory has a portfolio company doing wealth management with LLMs. OpenAI isn’t getting into wealth management. They’ll hire wealth managers for their own engineers, but that’s different.

Then the two hardest, closest to the model providers, where you’re still fundamentally a software company:

**Data flywheel.**You start with a simple app, and as you learn how your users actually work, you compound into a specialist, differentiated product over time.**Forward-deployed engineers.**Sometimes the software only works with people on the ground to capture context. That can differentiate you, though note the foundation model companies are signaling they’ll do this too, partnering with PE on one hand and making clear they’ll build it on the other.

The takeaway: there are subtle architectural moats and obvious business-model ones, and they point the same way. The model companies start with a trillion dollars and will be fine. There’s still room for everyone else.

Three Versions of “Is Software Dead”

The phrase means three different things. Keep them separate.

Version one: will the foundation models take it all from the new AI companies? Nobody knows, and most VCs are betting both ways: writing checks into Anthropic at a few hundred billion pre, while also funding companies that only make sense if there’s room outside the model layer. You have to form a theory of the case on what survives.

Version two: is plain vanilla SaaS dead? For companies that just automate a workflow, this is the real threat. Move and survive, or get rolled.

Version three: what happened to everything built before 2022? This is the