Why I’m Betting on a SaaS Rally in 2026.
The hype says AI will zero out the cost of software. The market is about to learn that “software” was never the expensive part.
Last time I argued AI won’t replace engineers. This time I’m arguing it won’t replace their employers either.
The One-Line Version
Amodei’s “software → free” thesis is a P&L mistake. He’s compressing one row on the income statement and calling it the whole business.
What Amodei Actually Said
At Davos 2026, Anthropic CEO Dario Amodei told the Wall Street Journal that AI would make software “cheap and potentially free,” because it would remove “the need to spread development costs across millions of users.”1
“If your moat is ‘our software is complex and difficult to write, and we can write it, and others can’t match it,’ I think that’s going away.”2
That second part, I agree with. If your only moat is hard-to-write code and a piss-poor customer experience, your days are numbered. That’s been true since before Claude could write Hello, World.
The first part is where he loses me. The lever he’s pulling is amortized development cost — the price of writing and maintaining code, divided across a customer base. Zero it out, and in his telling, “software” goes to zero with it.
Code cost is one row on a SaaS P&L. The other rows don’t budge.
Here’s why.
1. Code Is a Line Item, Not a Business
Engineering labor is real money. It is also one cost center among many. SLAs, hardware depreciation, compliance, on-call, contracts, sales, support, procurement integration — none of those get cheaper because Claude can scaffold a CRUD app.
AI is a competitive input to one role on the org chart. Pricing the whole business off that one input is a category error.
2. Systems Are the Moat
Your SLA doesn’t care that your runbook was AI-generated. Your auditor doesn’t care that Claude wrote your Terraform. Your enterprise customer’s procurement team doesn’t care how the sausage gets made — they care who is on the hook at 3 a.m. when the sausage stops getting made.
The bill for being a trusted service provider is paid in accountability, not keystrokes. AI compresses the labor inside those functions. It doesn’t compress the obligations.
3. The DIY Trap Is Loaded
“We’ll just vibe-code our own [Salesforce / Datadog / Stripe]” is the 2026 version of “we’ll just self-host it.”
AI makes the initial build viable for more people. It doesn’t make operating a production service viable for them. The DIY honeymoon ends one of two ways: the first time production melts down and nobody can explain why, or the first SOC 2 questionnaire that asks who owns control CC7.2.
4. Incumbents Win This Cycle
Mature SaaS companies are using AI to compress the one line item it compresses — engineering labor — while their other moats keep compounding. Every dollar of dev cost AI strips out of their P&L is a dollar they can spend widening the gap on the rows AI can’t touch.
The “AI levels the playing field” story assumes the incumbents are sitting still. They aren’t.
5. The Decommission Cycle
The vibe-coded replacement gets quietly decommissioned. Migration back to the SaaS vendor who does it cheaper, faster, and with someone to sue when it breaks — because that vendor also adopted AI, and used the savings to widen the gap in (1) through (4).
The “AI killed SaaS” narrative ages like milk.
Who’s Talking
I’m not a skeptic from the cheap seats. My career arc:
- Microsoft: Systems Engineer → Senior Service Engineer (the role Microsoft later renamed SRE, after I’d left).
- Amazon: Systems Engineer → System Development Engineer → L5 Software Development Engineer → L6 Senior SDE.
I’ve spent most of my career maintaining systems where the application software was written by other teams. But I’ve always written code:
- As an SE, I wrote the code that deployed and operated other people’s software — automation, monitoring, pipelines, recovery tooling, the production substrate.
- As an SDE, I write all of it: application code and the systems code that keeps it alive.
Which is why I’m extracting so much value from agentic engineering. AI saves me a shit ton of typing precisely because of my background. I know what the right answer looks like, where the failure modes hide, and how the operational pieces snap together. The agent handles the keystrokes. I handle the judgment.
The Receipts
I’m running two production-grade agentic engineering programs:
- Legion — the executor. kuhl-haus-legion: 1,144 contributions in the last year (40% commits, 42% PRs, 17% issues, 1% code review).
- Bishop — the reviewer. kuhl-haus-bishop: 144 contributions (99% code review, 1% issues).
Three control surfaces drive the system — Chat for instructions, Obsidian for design docs and reviews, GitHub for Issues and PRs. The day-to-day shape is similar to delegating to a junior engineer: decompose, provide references, iterate on design, iterate on code review, hand-hold through ambiguity. (How I do agentic engineering is its own post. One I keep putting off.)
Here’s what that’s cost me at Anthropic:
| Month | Anthropic spend |
|---|---|
| Feb 2026 | $77 (2 days of usage) |
| Mar 2026 | $754 |
| Apr 2026 | $1,125 |
| May 2026 (MTD) | $685 |
| Total | ~$2,641 |
Cheaper than a junior SDE? Yes.
Lower barrier to entry than hiring? Maybe.
Free? Not even close.
The CEO whose company makes software “free” has invoiced me $2,600+ in four months for the privilege.
Mini-FAQ
Q: Aren’t you an AI skeptic dressing up sour grapes as analysis?
A: I just told you I’ve spent $2,600 in four months on agentic engineering and I’m running a two-agent program in production. I’m a heavy user. I’m betting against the pricing narrative, not the technology.
Q: Doesn’t AI being cheaper than a junior SDE prove Amodei’s point?
A: It proves AI is a competitive input to one role on the org chart. It says nothing about SLAs, hardware depreciation, compliance, on-call, or contracts. See also: Myth: Companies Can Skip Hiring Junior Engineers with AI Tools.
Q: Why does AI save you so much typing if it doesn’t save everyone that much?
A: Because I know what I want, in what shape, with what failure modes, deployed how, monitored how, and rolled back how. The agent is a force multiplier on existing judgment. Hand the same tooling to someone without the systems background and you get a working demo that falls over the first time it meets real traffic. Same tool, different output.
Q: Are you predicting SaaS stocks rally in 2026? Is this investment advice?
A: I’m predicting the category rallies — incumbents widen their moats and DIY refugees migrate back. This is not investment advice. This is an old systems engineer telling you which way the wind is blowing.
Q: What would change your mind?
A: A non-trivial enterprise running a fully self-hosted, AI-built replacement of a major SaaS product in production, hitting four-nines, passing a SOC 2 Type II, and costing materially less all-in than the SaaS vendor it replaced. I’ll wait.
The Bet
Amodei’s “software → free” thesis confuses a line item with a business. He’s right that AI compresses the cost of writing code. He’s wrong that the cost of running software as a service follows it down. The $2,600 I’ve handed his company since February is, ironically, the cleanest evidence that “free” isn’t even the right word for the part he is talking about.
Even Amodei’s own audience didn’t buy it. CEOs at Davos pushed back on the same stage.3, 4
The companies that already know how to run software at scale are about to eat very well.
References
1: Anthropic CEO Dario Amodei said in a Wall Street Journal interview…
— Digg
2: Software will become ‘essentially free,’ warns Anthropic CEO
Amodei — The News
3: At Davos, CEOs said AI isn’t coming for jobs as fast as Anthropic CEO
Dario Amodei thinks — Fortune
4: AI luminaries at Davos clash over how close human-level AI is —
Fortune