Software Darwinism 2.0

How Intelligence (AI) Deflation is Impacting Software & Where the CSU family Stands.

Apr 10, 2026

I started this conversation on SaaS, AI, and the shifting market narrative two months ago with my first piece:

🔗 Software Darwinism: Does All SaaS Die in the Age of AI? And Where Does Constellation Software Fit?

And I think this is a conversation worth continuing. As investors, we’re watching a sector that’s evolving, and accelerating, as quickly as it has in years. Understanding where value accrues, where moats weaken, and where they strengthen is becoming one of the defining questions for software investing today.

There’s a debate worth paying attention to, one that some of the brightest minds in tech (including Jensen Huang) have been hinting at:

If intelligence (AI) is trending toward marginal cost zero, what happens to companies whose entire business model is acting as a toll-booth for that intelligence?

Some argue software may become an intelligence (AI) wrapper trap, others that the era of intelligence deflation will benefit software incumbents cost structures.

The thing is both sides of this debate may be correct, but they are missing an important point : their conclusions don’t necessarily apply uniformly across the broader software market. There will be winners and losers.

Just as the semiconductor industry went through a long consolidation that ultimately produced some of the most durable growth businesses in the world, something similar could happen in software. For decades after the dot-com crash, software benefited from a favorable structure: high margins, low CapEx intensity, and generous SBC-funded growth narratives.

That “advantage” may be changing.

📰 What’s New?

📚 Articles

💸 General Investing
💎 Hidden Gems Series
✨ Annual Specials – Annual Letters, Investing Philosophy, Industry Write-ups & Top Picks of the Year
📚 Resources for Investors
📢 Latest Stock News
💬 Join My Chat → [🔗 Learn more]

🔐 Paid Subscriber Exclusives

💼 Portfolio Corner – Holdings, valuation models, trades, performance & more!
🤫 Exclusive Sections – From select articles, Deep Dive Briefs & Hidden Gems

👀 Recent Releases!

✨ Top Picks: 2026
🔎 Deep Dive Brief: Cadence
💸 Software Darwinism → Why not all SaaS dies in the age of AI & Where CSU Stands
💎 Hidden Gems: Sixth Edition → A Primer on TerraVest Industries

👀 Coming Soon

🔎 Deep Dive Brief: Norbit ASA
💎 Hidden Gems: Seventh Edition

The SaaS Re-Rating is Real and Deserved ?

Let’s start with what’s happening in the SaaS market and what’s driving the SaaS-mageddon narrative, because this sets up everything else.

A growing share of OpenAI and Anthropic’s enterprise ARR is flowing through cloud platforms and SaaS products that embed their models. These businesses plug into the Claude or GPT API, slap a “Copilot” feature into their dashboard, and charge clients an extra $20/month per seat. They’re functionally reselling Anthropic’s intelligence with a markup.

It gets interesting when you look at the deflationary trend of intelligence (AI).

Foundation models are getting cheaper and better. That’s a fact, and Nvidia’s roadmap is hard proof of this and a glimpse into the future. So, if you are a software company and your entire value proposition is “we make it easy to use AI to write marketing emails,” your terminal value is trending toward zero. As underlying models get cheaper, barriers to entry collapse, competitors flood in, and pricing power evaporates.

The market increasingly believes this and it’s applying the same whip across the broader software industry. As a result, many horizontal and vertical SaaS companies have been heavily punished over the past 8–12 months.

The message from the market is becoming clearer: if you’re primarily a distributor of a deflationary asset (compute, intelligence), you shouldn’t trade at 15× ARR. You should trade more like a traditional distributor at a low double-digit EBITDA multiple. We’re not there (in most cases), but that’s the direction sentiment is moving.

The worst part of software investors is this re-rating does not seem to be a temporary market pessimism but a fundamental reassessment of what these businesses are actually worth, repriced one new Claude’s AI feature release at a time.

And when valuation frameworks shift at the industry level, not just the company level, multiples rarely snap back quickly.

The death of the GUI moat, and why horizontal SaaS deserves the scrutiny it is getting

There is a piece circulating from SemiAnalysis, 🔗 Claude Code is the Inflection Point, that has been generating a lot of discussion, and rightfully so.

It is a very good articulated framework for understanding what AI agents actually do to the software industry’s economics, not in a vague sense, but structurally. It is worth taking seriously, particularly if you own any software name.

The core of the SemiAnalysis argument is simple: AI agents do not rely on human-oriented workflows. For the better part of two decades, companies like Salesforce, ServiceNow, and Microsoft built part of their switching cost moats by forcing humans to learn their interfaces. An entire sales team that knows how to navigate Salesforce is a sales team that is not going anywhere. That is a real moat, and the market has priced it accordingly, sometimes generously so.

The problem SemiAnalysis points out is that the moat was never the software. The moat was the human in the loop.

Let me give you an example : If an autonomous AI agent becomes so good at its job that it’s able to query a database directly, synthesize the output, and deliver a finished analysis to the VP of Sales without a single human ever opening a CRM tab, then the $150/month seat license is not competing against a better CRM. It is competing against an API call that costs a fraction of a dollar. That is a different conversation entirely, and I think the market has been internalizing this. We are not there yet, but we are getting there fast.

This is why, if you’ve been following me for some time in my 🔗 Portfolio Corner, you’d notice that I don’t spend much time worrying about horizontal SaaS, even at current multiples. Thin wrappers, BI tools, and data-entry software built on 75-80% gross margins are facing a structural problem that no amount of AI feature-bolting will fully solve.

SemiAnalysis is basically saying: The foundational models are going to consume those margins, and the incumbents’ only defense is switching costs that are themselves being eroded.

When you put it this way, that is not a thesis I’d want to be long on. But, do I fully agree with this take ? Not really, I believe there are nuances within the industry.

Stress-testing VMS against the same framework

Where it gets more interesting and more contested is whether this framework extends to Vertical Market Software. SemiAnalysis specifically names Accenture’s use of AI agents in public sector and healthcare. Those are not random verticals. Those are the exact verticals that CSU and TOI have spent decades colonizing. So I think it is worth being honest about both sides of this.

The bear case is not unreasonable: the collapse in migration and maintenance costs.

One of the key reasons a municipality keeps paying its $50-100k annual maintenance fee to a VMS vendor is not because they love the software. It is because the alternative, hiring consultants to migrate decades of legacy data to a modern system, costs a few million dollars, takes months / years and some non-negligible operational pain. That switching cost is a function of human labor, and human labor is precisely what AI is making cheaper. If Accenture can use ‘almighty’ Claude Code, at some point, to build and deploy a custom billing system for that municipality in three weeks instead of months with supervised AI agents maintaining the new software autonomously, the $50-100k maintenance fee starts looking a lot less sticky. I would be intellectually dishonest if I dismissed this risk entirely. Again, we are absolutely not there yet, but the pace of innovation in AI suggests we might not be that far. The market believes this.

The bull case, however, rests on something that SemiAnalysis (and the market) mostly gloss over: the difference between information processing and regulatory reality.

VMS is not just software that processes data. At its best, it is software that encodes decades of hyper-local compliance requirements : union rules, tax codes, public sector procurement regulations, pension payout structures,… The code is not valuable because it is technically complex. It is valuable because it reflects bureaucratic reality that took years to accumulate and it’s deeply embedded in customer operations. No AI agent is going to hallucinate its way through safely. And even if an AI lab claims it can, there’s usually no decision-maker on the other side willing to delegate responsibility for those workflows to a probabilistic system.

There is also a liability dimension here that I think gets underweighted.

An AI agent generating the wrong slide for a marketing team? Embarrassing.
An AI agent miscalculating a 911 dispatch route or a pension disbursement? That’s a lawsuit and a news story.

The tolerance for hallucination error varies by vertical, and VMS tends to sit in the low-tolerance end of that spectrum.

Why the CSU Family is Different

If you own Constellation Software, Topicus, or Lumine, you should absolutely be monitoring this debate.

Now, let’s point out why I do think the CSU family is different and more resilient to this AI narrative :

1. The Code is No Moat. It is the Workflow, Switching Costs and Customer Trust

I feel like I’ve repeated myself so many times on this topic across the Substack that I’m starting to get déjà vu. So, I’ll keep this one simple.

VMS moats aren’t built on code complexity. They’re built on switching costs, regulatory compliance, embedded workflows, industry-specific know-how, and customer trust.

The software could be written in COBOL or generated by ChatGPT, the customer doesn’t really care. What they care about is that it works, remains compliant, and doesn’t introduce migration or reliability risk into mission-critical processes.

We’ve covered this before, so I won’t go deeper here.

2. CSU Owns Distribution and Data

CSU companies already have the customers and the proprietary data. If an AI feature truly adds value to a golf course management system, the CSU company will simply plug the Anthropic API into their existing software. Because they own the distribution (the customer relationship) and the data (20 years of golf course tee times, booking patterns, member preferences), they can defend their position.

They might not be able to mark up the AI feature aggressively, and they shouldn’t try, but they won’t lose the client to an upstart either. The customer relationship and embedded position are the assets that matter, not the sophistication of the codebase.

3. The First Paradox: Why intelligence deflation might actually be a tailwind for CSU family margins and cash flows

Here is the part of this thesis that I find underappreciated. The same AI forces that threaten VMS’s organic growth could simultaneously be the best thing that ever happened to Constellation’s margins and cash flow generation.

Why?

The CSU family owns thousands of legacy software assets that require ongoing developer resources to maintain, patch, and update. If one developer with access to modern AI tooling can now do what previously required a team, the internal R&D and maintenance burden on those assets drops considerably. And that matters, because R&D is one of the largest operating cost lines across the Constellation ecosystem. Of course, I don’t know exactly where margins land (nobody does) but the direction of travel seems fairly clear. EBITDA margins that currently run around 25% could organically expand without any organic revenue growth whatsoever. That would be an unusual situation.

4. The Second Paradox: Buying the “ashes” of the SaaS industry might actually be a tailwind for CSU and family growth runaway

The M&A angle is, if anything, even more interesting.

SemiAnalysis is essentially arguing that VC-backed horizontal SaaS companies are going to see their moats and margins compress. When that happens, their valuations compress with them.

CSU has spent 30 years building a database of acquisition targets and a repeatable process for deploying capital into software businesses at disciplined multiples. If the cost of acquiring a SaaS business drops from 8x revenue to 1x revenue because AI has structurally impaired its growth outlook, then the ROIC on Mark Leonard’s next wave of capital deployment could be meaningfully higher than the last. The M&A bonanza thesis requires the bear case to be directionally right about SaaS valuations (we are seeing it already in the public market), which is exactly why it is worth taking the SemiAnalysis framework seriously rather than dismissing it.

So, I say : let the SaaS-mageddon unfold while CSU and co. pick up the corpses and proof test its new strategy, PEMS.

How I think about this in portfolio terms

Put all of this together and I think the honest conclusion is that software deserves a reduced weight relative to where most long-term software investors have historically sized it, not because the cash flows are impaired today, but because the range of organic growth outcomes has widened in an unfavourable direction.

For VMS and a holdings companies like the CSU family, the upside scenario (AI drives internal margin expansion and a cheap M&A wave) is compelling. But, the downside scenario (switching costs erode faster than margins expand) also creates uncertainty. That uncertainty could warrant a smaller position, not a zero.

The other natural hedge here, which I am more inclined to take, is owning the companies actually causing the intelligence deflation : the semiconductor ecosystem and the hyperscalers. If AI agents are truly going to commoditize software labor and erode software moats, then the economic surplus has to land somewhere. It lands with the companies selling the picks and shovels. Owning both sides of this transition is not a hedge in the traditional sense; it is an acknowledgment that the direction of travel is clear even if the exact casualties are not.

The SemiAnalysis thesis does not make VMS uninvestable. What it does is raise the bar for what you need to believe to size it aggressively. And right now, I think that bar is higher than many long-term software investors are giving it credit for.

Thanks for following along,

—Nikotes

Unlock Premium Content – For just $0.39/day ($12/month) or $0.27/day ($100/year)!

🔗 Portfolio Corner - 🗓️ Monthly Updates (Last Update: 29-Mar-2026)

Discussion about this post

Ready for more?