Notes on the future: AI, SaaS, and the markets that will bleed
These started as private notes in Portuguese, a thesis I kept rewriting to myself about what AI is actually doing to the software market. This is the cleaned-up English version, opinionated on purpose.
In short, before I lose you: per-seat pricing is dying, the small-team paid DevTools mid-tier is going to bleed because engineers are the one group not inside Plato’s cave about their own tools, and early-stage seed and Series A in pure technical SaaS is going to bleed because bootstrapping a small excellent team is genuinely viable now. None of these segments disappear. They just lose a real slice of the market, and the shape of what’s left changes.
The thesis
AI is not going to kill SaaS the way a lot of people are claiming. Two specific things will happen. Per-seat pricing breaks, because agents don’t take seats. And a new category, AI SaaS or Agent SaaS, grows alongside traditional SaaS, not in place of it.
The two niches that bleed, meaning lose a real slice but don’t disappear, are DevTools and analytics tools in the single-dev and small-team segment, and venture capital in seed and Series A of small technical startups. In both cases what’s actually shifting is the boundary of who needs whom, not the existence of the segment.
Why SaaS stocks are bleeding
The current drawdown is a pricing and valuation crisis, not a death certificate.
Valuing a SaaS used to be the easiest thing in software: take ARR, multiply. The dominant pricing model was per seat. With AI, that stops making sense. Why would you buy 50 seats of Slack when one agent touches the whole database through one seat? Investors don’t know how to price this transition, so the market is volatile.
The intuition has a name now: the SaaSpocalypse. In February 2026 a 48-hour selloff erased about $285B in SaaS valuations, and the broader correction has erased somewhere around $2T since the start of 2026. Median revenue multiples fell from over 7x to under 5x. Forward P/Es from around 35x to around 20x, levels you haven’t seen since 2014. The root cause is exactly the one above: agents don’t need seats, they execute thousands of tasks without taking up a single one of the systems they operate in.
Pricing will migrate to per-usage over the next couple of cycles. The startups that figure out how to rebuild GTM and pricing around agents first will lead the next one; investors arrive later, as they tend to. Gartner projects that at least 40% of enterprise SaaS spend migrates to usage, agent, or outcome-based by 2030, and live examples already exist: Intercom at $0.99 per resolved ticket, Zendesk at $1.50 to $2.00. Plenty of companies will fail in the transition. Most will not.
Why traditional SaaS doesn’t die
There’s a narrative going around that software has gone to zero, that building software is free now because of AI. The first part is true. The second part skips a step that turns out to matter quite a lot.
The cleanest evidence of the missed step came in May 2026 from Uber. COO Andrew Macdonald said publicly that if you can’t draw a line between AI spend and useful features shipped to users, the trade gets harder to justify.
The numbers are brutal. Uber rolled out Claude Code in December 2025. By March 2026 84% of engineers were “agentic coding users”, up from 32% in February. Around 95% of engineers use AI monthly. Around 70% of committed code is AI-assisted. And: they burned the entire 2026 AI budget in four months, and the spend didn’t translate into more useful features at the customer end.
The dynamic ends up looking like this. Before, sales asked for ten features a month and engineering shipped them in two months, so there was always a queue. After, sales is still asking for ten features a month and engineering is shipping them in a week. Three weeks of capacity sit empty, and engineers fill the gap with features that nobody outside the building asked for.
In an org with a sales function, engineers don’t have a direct discovery channel to the buyer. Sales does. Engineering builds, sales sells the half-built product as if it were finished, and the product chases. You don’t build the whole thing in private and reveal it, or you spend months on features the customer doesn’t value when you finally show them. It’s basically Plato’s cave: engineers inside the building see dashboards and tickets and metrics, which are shadows of the actual business, while outside the cave is the customer, and the person talking to them is sales.
I want to handle an apparent contradiction before it becomes one. Later in this post I’ll argue you can build an internal tool in two days now. That doesn’t contradict the cave. AI gives real productivity gains, but the gains only convert to customer value when there’s a discovery channel between the people building and the people using. In internal tools the engineer is the user, so there’s no discovery bottleneck and the gain converts cleanly. In external products the bottleneck sits in sales, and the engineering gains pile up behind it instead.
To keep this honest, I should name a counter. If AI also accelerates sales, with AI SDRs, automated demos, AI-assisted discovery, then the ten-features-a-month ceiling lifts and part of this argument weakens. My thesis assumes sales scales more slowly than engineering, which has been historically true but isn’t guaranteed to keep being true.
There’s a wider productivity point underneath all of this. Companies measure productivity in lines of code or commits, which is a fallacy. Anthropic published research about how Claude has changed work inside the company and openly admitted that productivity is hard to measure precisely. The industry is moving from input metrics, LOC and commits, to outcome metrics like change failure rate and deployment frequency. Real productivity is more customer-requested features being accepted by customers, not raw output. AI doesn’t speed up sales cycles, transactions, demos, or discovery, and those run at the same speed they always did.
The mirror case to Uber is Klarna. In 2023 they replaced 700 customer-service agents with an OpenAI chatbot. In 2025 CEO Sebastian Siemiatkowski walked it back: they had focused too much on efficiency and cost, the quality had dropped, it wasn’t sustainable. They went hybrid.
In both cases the bottleneck was never engineering. It was the part where someone has to talk to the customer and understand what the answer actually means, and that’s not where AI helps.
The first bleed: DevTools and analytics
These tools existed because verticalising your own analytics, logs, and dashboards used to be too expensive and too slow. Paying $30 a month for Datadog or PostHog freed engineers up to focus on the product they actually cared about.
Two things changed. The first is that engineers have time freed up now, for the reasons in the Uber section above. The second is that DevTools are the one part of their world where engineers are not inside the cave: they are the customer of their own DevTools, they know exactly what they want, and the cost of building it has collapsed to a VPS, a database, and an agent maintaining the code over time. Five years ago I’d have shipped my analytics data to an external platform; today I build the analytics tool I want in two days and the agent picks up the maintenance.
It’s worth being precise about which segment actually bleeds, because the broad framing gets it wrong. A lot of single-dev tools live on free tiers and convert almost nobody to paid, and those aren’t really in the bleed because they were never collecting revenue from this audience in the first place. The bleed is in the small-team paid mid-tier, the $100 to $500 a month range, where the cost of the subscription finally comes within a couple of hours of agent-orchestrated work to replace it. That’s where the math flips.
This doesn’t extend to the whole DevTools market. The tools that survive and grow have at least one of three properties. Data network effects, like Sentry and Datadog, where anomaly detection and benchmarks come from millions of apps and you can’t replicate the dataset locally. Organisational scale: observability for 200 engineers, distributed CI, security and compliance, SOC2. The cave argument for the individual engineer doesn’t cover the tradeoffs at CTO or Head of Platform level. Operational risk: when observability fails in production the cost is large, and “an agent maintains it” only works for tools that aren’t load-bearing.
What’s actually bleeding is the single-dev and small-team segment. Team and enterprise scale holds or grows. The global market shrinks. The mix shifts.
The second bleed: VC in early-stage technical funds
The old logic ran like this: to scale, you have to hire a lot of people fast, to hire fast you need outside capital, and therefore VCs were unavoidable for anyone with growth ambitions.
AI changed this slice in two ways. You don’t need a large team anymore, you need an excellent one, because a good engineer with AI produces more than two or three average engineers. And the shape of the role has shifted underneath. The ideal engineer is not writing code, they are orchestrating agents, with attention split across backend, frontend, database, and infra, and most of the day spent taking fast decisions that unblock the agents already in motion.
The consequence is that bootstrapping has stopped looking small-time for small-to-medium technical products. It used to mean slow and unambitious. Now it’s a viable route to $1M to $10M ARR with a team of two to five, no equity dilution, no handing control to investors. More technical startups skip the first round, so VCs lose deal flow and returns in that specific slice.
This doesn’t apply everywhere, and it’s worth being honest about where it doesn’t. VC capital was never just about paying engineers, and it’s still required where AI doesn’t substitute: enterprise GTM where you’re funding a sales org and brand and events and partnerships, winner-take-all markets where speed still gets bought with capital, heavy capex like fabs and hardware and biotech, paid acquisition in consumer where CAC still needs eight-figure cheques. The bleed is specifically VC capital aimed at seed and Series A of pure technical SaaS. Late-stage, growth, enterprise GTM, deep tech, and consumer all hold. What gets lost is an archetype of startup, not the sector.
A note to myself
The argument that verticalising is cheap with AI applies to my own product. Whoever buys my SaaS can run the same logic back at me. The only defence is the same set of things that protect the DevTools that survive: aggregated data the customer can’t replicate, organisational complexity that makes verticalising more expensive than buying, or operational risk the customer doesn’t want to take on. Without at least one of those, what I’m selling is a single-dev DevTool dressed up in a logo, and it will bleed for the same reasons everything else in this category does. I try to remember that when I get carried away.
Where this leaves us
The short version is that AI isn’t killing software. It is breaking per-seat pricing, eating a chunk of the small-team paid DevTools market because the engineers using those tools are also the people who can now build them, and eating a chunk of early-stage technical VC because the kind of small team that used to need an outside cheque doesn’t, anymore, in the cases where the product is mostly software. The rest of the market keeps going. Mostly what shifts is the shape of who pays whom for what, not the existence of any of the players.