>
# Part 1: Software Moats Are Never About Code — They Are About Ecosystems and Standards
The SaaS doomsday narrative comes in two classic variants, both circulating widely across X and financial communities, both sounding plausible on the surface, and both making the exact same fatal mistake.
# Variant One: "AI-Native Custom Development Shops" Will Replace SaaS
The story goes like this: a new breed of AI-native development companies will emerge, building bespoke HR systems, ERPs, and CRMs for every enterprise using AI-generated code — rendering subscriptions to ServiceNow, SAP, Workday, and Salesforce obsolete. Companies will simply "build their own."
It sounds compelling. It ignores one thing: **software is never a one-time asset once written. It is a liability that requires permanent maintenance.**
The historical graveyard of "build your own" failures makes this painfully clear.
The UK's National Programme for IT: launched by the British government in 2003, the project aimed to build a unified electronic patient records system for the entire NHS. After a decade and more than £12 billion spent, it was officially abandoned in 2011 with almost nothing usable to show for it — the single most expensive IT project failure in public sector history.
Queensland Health Payroll System, Australia: over A$1.2 billion spent customizing a SAP-based payroll system, which upon launch catastrophically failed to pay thousands of nurses and doctors for months, eventually triggering a full parliamentary inquiry.
The FBI's Virtual Case File: the FBI spent $170 million and three years building a custom case management system, only to discover on the eve of launch that it was completely unusable. The entire project was scrapped.
Germany's Lidl: spent over €500 million across seven years attempting to customize SAP for its operations, abandoned the effort entirely in 2018, and reverted to standard SAP — seven years of work and hundreds of millions of euros, ending exactly where they started.
These failures share one common thread: **the cost of "custom" is never the development cost. It is the ongoing cost of maintenance, updates, compliance, and integration.** Every OS upgrade requires adaptation. Every regulatory change — labor law, tax code, data privacy directive — requires synchronization. Every internal restructuring requires reconfiguration of permissions. SAP and Workday subscriptions already bundle all of that. Build your own, and your team now owns every one of those responsibilities indefinitely.
More fundamentally, the **domain knowledge** these companies have accumulated over decades is not in the code. Workday's HR modules encode labor law across dozens of countries, continuously updated. SAP's financial engine is a living implementation of GAAP, IFRS, and countless national tax regimes. ServiceNow's ITIL framework is the crystallization of decades of IT industry best practices, refined through edge cases discovered across tens of thousands of enterprise deployments. An AI-native custom system can handle the scenarios you describe to it. What it cannot handle are the scenarios you haven't thought of yet — and that is precisely where enterprises get burned.
History has told us this story repeatedly: **a technically superior challenger does not automatically defeat an incumbent that owns the ecosystem and the standards.** In the mid-1990s, Netscape Navigator was objectively better than Internet Explorer across almost every technical dimension, commanding 90% market share. Microsoft bundled IE directly into Windows and crushed Netscape — not through better code, but through distribution control. Betamax was widely acknowledged to have superior picture quality over VHS, yet lost the format war because Matsushita and JVC controlled more content licensing. Today, ServiceNow maintains a 98% renewal rate among Fortune 500 companies — not because there are no comparable alternatives (Atlassian's JSM is priced at roughly one-fifth of NOW) — but because ripping ServiceNow out of a large enterprise's IT, HR, and security operations is itself a multi-year, tens-of-millions-of-dollars engineering undertaking. That switching cost dwarfs any competitor's pricing advantage by an order of magnitude.
If any enterprise genuinely abandons its existing SaaS ecosystem in favor of some "AI-native alternative," they should first do the math: the time and risk required to rebuild a new ecosystem may far exceed the time required to modernize the existing one to meet AI-era standards. More ironically, companies that take this path will find they are no longer the beneficiary of Software as a Service — they have become the raw material, the data source and ecosystem-building fuel, for the AI platform companies. It is no longer Software as a Service. It is **Surrender as a Serf** — same acronym, entirely different meaning.
# Variant Two: AI Agent Adoption Will Cause Seat-Based Subscription Revenue to Collapse
The second variant is more specific: as AI agents replace human workers inside enterprises, the headcount supporting per-seat subscriptions will shrink, and SaaS revenues will follow. The logic sounds coherent. The actual data says otherwise.
**Where exactly is the seat reduction? Show your work.**
Two data points from the most recent earnings cycle answer this directly.
Salesforce's latest quarterly results showed Agentforce ACV surpassing $1 billion with nearly 30,000 deals closed. Critically, this is not seat revenue being replaced by agent subscriptions — it is an entirely new revenue layer sitting on top of existing human seat revenue. Enterprise headcount did not decline. Customers with AI agent capabilities in place are simply spending more on the same platform. Salesforce's AI revenue is additive, not substitutive.
Atlassian's Rovo AI now has over 5 million monthly active users. Customers using Rovo are growing ARR at twice the rate of customers who are not. AI capability is functioning as a retention and upsell tool, not a seat compression mechanism.
Both data points point to the same conclusion: **AI agents are revenue amplifiers for SaaS platforms, not headcount cannibals.** The fear that "fewer people means fewer subscriptions" rests on an assumption that has yet to show up in a single earnings report.
Even if seat counts do eventually soften, the SaaS incumbents have already repositioned ahead of it. These companies are aggressively moving toward consumption-based pricing, outcome-based pricing, and enterprise-wide agreement structures with higher total contract values. This is not defensive posturing — it is a deliberate upgrade of their monetization ceiling.
ServiceNow is particularly well-positioned. Seat-based revenue already represents only about 50% of NOW's total revenue — the company was never fully dependent on headcount to begin with. More importantly, NOW has been bundling AI Control Tower directly into enterprise subscriptions with a free trial period. The strategic logic is straightforward: once a company's AI agent governance, security compliance, and ticket automation are all flowing through Control Tower, the system becomes the nervous system of that enterprise's digital operations within months. Almost no one voluntarily disconnects from their nervous system. This is not a sales tactic — it is product-level lock-in, identical in mechanism to how Apple's ecosystem works.
**The deeper point: in the age of AI agents, the value of SaaS platforms does not shrink — it upgrades.**
For AI agents to actually "do things" inside an enterprise — trigger approvals, modify records, call APIs, execute tickets — there must be somewhere that records what was done, who authorized it, and whether the outcome was compliant. That "somewhere" is precisely the System of Record that ServiceNow, SAP, and similar platforms have spent decades building inside enterprise operations. The more agents there are, the more critical that governance and record layer becomes — not more redundant. Every additional AI agent running inside an enterprise is one more entity that needs to be governed, audited, and managed — that is a tailwind for these platforms, not a headwind.
# Part 2: What the Market Is Actually Afraid Of — and Why It Is Wrong
The market's fear is not entirely unfounded. It has real case studies behind it.
Chegg ($CHGG): nearly zero moat, a product that provided standardized homework help. When ChatGPT arrived, students got better answers for free. Chegg's subscriber base collapsed within a couple of quarters and the stock followed. This is a genuine case of AI destroying a SaaS business — but it is also the clearest possible illustration of the necessary conditions for that outcome: **thin moat, standardized product, zero switching cost.** What fits that description is Chegg. What does not fit that description is SAP.
Adobe ($ADBE): as generative AI image creation rapidly improved, the market concluded that Midjourney and Sora could replace Photoshop and Premiere. ADBE traded to multi-year lows. In reality, Adobe Photoshop with Firefly integrated still substantially outperforms any publicly available AI image tool in precision professional use cases. The PDF format, digital signature standards, and print output specifications that Adobe defined are industry infrastructure — no AI tool can bypass them and still function professionally. Adobe has real risks, but "replaced by AI" dramatically overestimates the professional capability ceiling of current AI tools.
Duolingo ($DUOL): this one warrants genuine caution. Sam Altman has publicly expressed interest in providing free education. If OpenAI develops a serious educational product, Duolingo's moat — gamified language learning — does not hold up well against a truly conversational AI language partner. This is a name I would not put on today's buy list.
Then came the catalyst for the current wave of panic: Anthropic launched Claude Cowork, SaaS stocks sold off sharply on the day, and the market's narrative escalated from "specific features being disrupted" to "AI agents will eventually handle all enterprise workflows and the entire software layer will be zeroed out."
Three fundamental rebuttals to that narrative:
# Rebuttal One: Challengers Think They Are Changing the Rules. Incumbents Already Shipped Your Disruption as a Plugin.
There is a fact the large model companies would prefer you not focus on: **the capabilities you believe are disruptive have already been absorbed into the existing software as features** — often running at higher precision inside more specialized professional contexts.
Adobe Firefly is already inside Photoshop. Professional designers do not need to switch to Midjourney; they use AI directly inside the tool they already know, with uninterrupted workflow, higher output precision, and cleaner commercial licensing. ServiceNow's Now Assist already performs intelligent classification, auto-summarization, and solution recommendations inside ITSM tickets. Salesforce's Agentforce crossed $1 billion in ACV with nearly 30,000 deals. These are not PR gestures — these are signed contracts.
Now consider the inverse question: how well does a general-purpose large language model trained on public internet data perform when processing a multinational corporation's proprietary ERP data, internal approval workflows, and historical IT ticket records? The Bain survey of nearly 1,000 companies answers it: close to 40% of enterprises reported actual AI cost savings below 10% of their stated targets. A general-purpose model facing enterprise-specific proprietary data is like a brilliant new graduate with no industry experience — genuinely intelligent, but nowhere near capable of independently conducting a financial compliance audit. "Smart" and "domain-ready" are separated by decades of specialized knowledge accumulation.
More critically: **even as AI capabilities continue to advance rapidly, these incumbent SaaS platforms will be among the first to integrate and benefit from those advances — not the last to be displaced by them.** Because they control the data, the integrations, and the user habits — any new AI capability must flow through these platforms to reach enterprise users at scale.
# Rebuttal Two: Staying on Your SaaS Platform Means Zero-Cost Model Switching. Moving to an AI Ecosystem Means Being Locked In.
This is the most underappreciated point in the entire debate.
**Choosing to continue using ServiceNow, SAP, or Workday leaves you completely free at the AI model layer.** ServiceNow's pricing model charges per "assist" — per task execution — with no binding to any specific underlying model. NOW's multi-model routing architecture is publicly documented: high-frequency IT ticket summarization goes through NOW's own LLM or NVIDIA's open-source Nemotron; deep technical reasoning goes to Google Gemini; HR cases requiring empathetic language go to Anthropic Claude; general agent orchestration goes through Azure OpenAI. The implication is significant: **you can swap the underlying model for whatever is cheaper and more capable next month — including GLM-5.2, or whatever open-source breakthrough arrives next year — without changing a single line of your enterprise workflow.**
The inverse is also true. If you bet on a proprietary AI platform's agent solution to replace your existing SaaS stack, you have staked your workflows, your data, and your integrations on a single vendor's roadmap. In an environment where model capability leaderboards turn over every few months, that single-vendor dependency is a deliberate surrender of optionality — and optionality is precisely what is most valuable in the age of AI agents.
# Rebuttal Three: The Rise of Open-Source Models Is a Threat to AI Labs. It Is a Windfall for SaaS Companies.
The recently released GLM-5.2 ranks fourth on the Artificial Analysis Intelligence Index — positioned between Gemini 3.5 Flash and Claude Opus 4.8 — outperforms GPT-5.5 on long-horizon coding benchmarks, carries inference costs at roughly one-sixth of Claude's flagship tier, and ships under a fully open MIT license. The media has widely characterized it as an event comparable to last year's DeepSeek shock. It was released with unmistakable strategic timing: just days after new US export control directives took effect, directly answering the question "what do you use as your foundation model when access to American frontier models is restricted?"
If open-source model capabilities continue advancing — if models like GLM-5.2 can handle increasingly complex agentic reasoning tasks within the next year — enterprise dependency on expensive closed-source APIs like Claude and GPT will systematically decline. **This represents a potential shock to AI lab revenue that may be more severe than January 2025**, because this time the disruption is not limited to training efficiency — it is a systematic erosion of the pricing premium on closed-source inference. And today's AI labs are deeply entangled with semiconductor hardware and cloud providers through high-leverage financing loops: if downstream AI lab margins get compressed, every node in that transmission chain feels the pressure. (This will be covered in detail in a dedicated follow-up post.)
For the SaaS companies in today's headline, the effect runs in the opposite direction: **the stronger open-source models become, the lower SaaS platforms' model routing costs, and the higher their AI feature gross margins.** ServiceNow can route more tasks to free open-source models like GLM-5.2, reducing dependency on expensive closed APIs, while the per-task fee charged to customers remains unchanged. That is pure margin expansion — not a risk. Microsoft Copilot has already announced plans to incorporate open-source models across different reasoning tiers, validating this commercial logic.
**Summary**: AI labs' API pricing power is being systematically eroded by open-source alternatives. SaaS platforms, sitting above the model layer and routing flexibly across all available models, are the quietest beneficiaries of the open-source revolution.
# Part 3: If SaaS Doomsday Actually Arrives, It Will Be Everyone's Doomsday
Let us conduct a thought experiment. Follow the SaaSpocalypse narrative all the way to its logical conclusion and see where the path ends.
Suppose OpenAI and Anthropic actually succeed. They build new ecosystems that dissolve the moats ServiceNow, SAP, Workday, and Salesforce spent decades accumulating. All enterprises migrate their data and workflows onto AI platforms. The SaaS incumbents disappear.
What falls next?
If enterprises have handed all their data to AI platforms, what purpose does Windows serve as an operating system? Microsoft's moat collapses alongside the SaaS companies.
If AI platforms now hold all enterprise and personal data, why would anyone search on Google? Google's entire business model becomes structurally irrelevant.
If AI launches social recommendation and shopping features, what remains of Meta's and Amazon's traffic moats?
If the AI labs begin selling their own hardware — rings, glasses, watches — and their own inference chips, Apple's and NVIDIA's ecosystem barriers face direct assault.
At that point, AI controls the full corpus of proprietary and public data accumulated by human civilization. It controls every entry point from information retrieval to commercial transaction to work execution. It manufactures the chips required to run itself.
Congratulations. What you have just described is not a new commercial landscape. It is Skynet.
**This narrative, followed to its end, produces only two possible conclusions:**
Either the chain breaks somewhere — some moat exists in the real world that AI cannot dissolve — in which case the companies that are "too structurally embedded to be swallowed by Skynet" are exactly what you should be buying into during today's panic.
Or the chain holds. In that case, your investment portfolio will be the least of your concerns.
**When the market is most afraid, the companies with genuinely durable moats are typically the ones most mispriced.** $NOW today carries a five-year total return of barely 30%, has declined nearly 40% over the past year, and trades at a 23x forward P/E that contains virtually no AI growth premium — while simultaneously holding $27.7 billion in remaining performance obligations, an accelerating Now Assist ACV trajectory, a security moat built through the Armis and Veza acquisitions, and an AI Control Tower strategy with Jensen Huang's personal endorsement. This is not a story of AI disruption. This is a story of market mispricing.
Personal opinion only. Not investment advice. Discussion welcome.
*Next post: Why a second DeepSeek shock will be significantly more severe than the first — and how the high-leverage financing loop connecting AI labs, semiconductor hardware, and cloud providers will transmit that impact through the system.*
Show full