The Configuration Collapse

The VC kingmakers are conspiring. a16z says it's Workday's last workday. YC is funding AI-native everything.

We've been writing about these companies since March, well before a16z and YC released their positions. Not out of disdain or spite, but as a wake-up call to something very alarming we are seeing with these legacy systems of record and how their architecture works.

Workday and ServiceNow, in particular, are the best examples here because they are the least diversified and have been impacted the most. They are also smaller than Oracle and SAP, so they have less sand in their hourglasses before time is up. But the apps' revenue of all of the SaaS giants could end up the same, from Salesforce to SAP.

The Systems Engineering Trap

Mark Cuban calls it the Innovator's AI Dilemma, a riff on Clayton Christensen's classic framework. He posted it on X on April 5, 2026, and the framing is sharp: public company CEOs face a lose-lose choice. They either need to tear everything down and rebuild from scratch as an AI-native company, or sit around and wait and do nothing while the stock price erodes. Cuban predicts two waves of shareholder lawsuits incoming, one against companies that opt for the teardown and crash the stock, and another against companies that fail to act and let AI-native startups eat their lunch. In his words: "I think most CEOs don't come close to understanding AI in enough detail to even begin to consider these decisions."

Boris Cherny, the creator of Claude Code at Anthropic, said the quiet part out loud at Sequoia AI Ascent 2026. The reason large companies can't move is not the technology. It's that "the large company has to evolve their business process. They have to evolve the way they work. They have to retrain everyone to use technology. They're going to face a lot of internal resistance to that." That's Cuban's lose-lose, viewed from the operator's seat.

It's surprisingly clear how many of these companies are choosing to sit around and wait. But they're not being complacent. They are actually spending more money on fake marketing and press releases for AI than they have actual AI to ship.

And it goes further than press releases. The fake marketing has moved into the org chart. Pull up LinkedIn profiles at ServiceNow and Accenture now carrying the title "Forward Deployed Engineer." Many show no public commit history, no open source presence, and no shipped software you can point to. The ones that do show technical work tend to show platform configuration, not engineering output. That raises a fair question about what is being engineered versus rebadged. Salesforce ran the cleanest example of this. They announced Headless 360 at TDX 2026, which was the right direction. Everything exposed as an API, MCP tool, or CLI. Native React. Coding agents with full platform access including Claude Code, Cursor, Codex, and Windsurf. That is a real software engineering roadmap. In the same news cycle they announced the FDE Partner Network, with Accenture and Deloitte as launch partners joined by Capgemini, Cognizant, IBM Consulting, KPMG, PwC, Slalom, Tata Consultancy Services, and more than 20 additional firms. The same SIs who have configured the platform for 20 years, rebadged. Accenture is not recruiting computer science graduates into these roles, at least not today. Those two announcements cannot both be true.

One of the most common practices we are seeing among Workday customers right now is building their own MCP servers because they cannot rely on Sana, and connecting them to their internal tools, usually Copilot or Claude, is just easier. We've seen demand for this use case higher than anything else we've released.

It's not that Legacy SaaS doesn't realize this. The infrastructure itself is unbelievably archaic to rebuild for AI, especially considering none of these companies has ever built applications this new way. Their current product teams exist to keep the lights on. And very little recruiting is going on at top-tier schools to get engineering talent that could even solve the problem.

These platforms were architected before the practices that define modern software engineering existed as a stack: distributed version control (Git, 2005), public code hosting (GitHub, 2008), CI/CD, microservices, infrastructure as code, API-first design, polyglot open-source toolchains, trunk-based development, pull request review.

Oracle Applications were built when "version control" meant patch sets shipped quarterly on physical media and applied with adpatch. Workday was built by founders who came from PeopleSoft and reproduced the same closed paradigm in a cloud wrapper: a proprietary internal language, no customer access to source or schema, customization through forms-based metadata tasks rather than commits to a repo. ServiceNow followed the same playbook with the Now Platform, a metadata-driven runtime where customer extensibility flows through scoped applications, business rules, and UI policies, all configured in forms behind a vendor-controlled UI. SAP layered another generation of complexity on top of ABAP, an internal language that has been carried forward since the 1980s mainframe era.

All four systems were architected around a "configure, don't code" paradigm where the vendor owns the engineering and the customer (or their SI) owns the configuration. Extensibility is through tables and forms, not pull requests against an SDK. There is no repo for a customer to fork. There is no service boundary you can replace. The platform is a closed runtime, and the customer's only lever is metadata inside a UI the vendor controls.

That is the architectural reality the AI-Native Services thesis is pushing against.

Boris Cherny put a finer point on it at AI Ascent 2026 using Hamilton Helmer's Seven Powers framework. Two of the moats that legacy ERP and ITSM vendors have relied on for decades are now eroding fast. Switching costs are going away because "you can just use the model and you can kind of port from one thing to a different thing." Process power is going away because Claude "is getting really good at figuring out process, and especially with 4.7 it can just hill climb anything." Switching costs and process power are exactly what Workday and ServiceNow have been selling for two decades. The creator of Claude Code naming both of them as dying moats is not a subtle signal.

The Leadership Problem

Sam Altman posted on X: "i am switching to polyphasic sleep because GPT-5.5 in codex is so good that i can't afford to be sleeping for such long stretches and miss out on working." Take that as seriously or as ironically as you want. The CEO of OpenAI is publicly signaling that his coding agent is too productive to sleep through.

Now contrast that with the leaders running Oracle, SAP, Workday, and ServiceNow. They are sales executives, not engineers. None of them are running polyphasic sleep experiments because their coding agent got too productive. AI-first development tools have not entered their personal workflow, and you cannot lead an AI-native organizational transformation from a seat that has never used the tools driving the transformation.

Architecture is one constraint on the legacy vendors. Leadership profile is the other.

The Automation Curve

So who cares, Brad? They can still just add AI on top of the legacy platform? Well, that might be true for the lowest-value use cases. But we are heading to a tipping point, and you can already see it when you look at the data from the true AI-native companies.

Anthropic is the benchmark, and the gap to everyone else is bigger than most people realize. The table below uses Anthropic as the reference point and shows what each company is doing with their headcount to keep up.

The "vs Anthropic" column shows how many times Anthropic's revenue per employee is larger than each company. Private-company figures (Anthropic, OpenAI) are estimates based on third-party reporting. Sources at the end of the document.

This is what people in the AI-native world call token maxing. Replace headcount with tokens. Replace people with model invocations. Higher token consumption equals higher leverage. YC's Diana Hu frames it as "token maxing over headcount." Jensen Huang has reportedly suggested developers should burn $250,000 in AI tokens per year on top of their base salary.

Screenshot of the Wikipedia 'Token Maxxing' entry or the YC Diana Hu 'How to Build an AI-Native Company' frame. Shows the standard definition: humans consuming more AI tokens as a productivity metric.

What Anthropic is doing is token maxing on steroids. T-max on roids, if you're typing fast. The standard frame is humans using AI more. Anthropic has pushed past that into AI using itself. Loops on loops. Cron jobs that fire coding agents overnight. Agents that review PRs while the team is asleep. Agents that build the next agents. The tokens are not being burned by people clicking keyboards. They are being burned by software that does not sleep. The model writes the code that builds the next version of the model. That is the part the legacy vendors cannot bolt on. They can buy more model access. They cannot rewire their codebase to build itself.

Boris Cherny, the creator of Claude Code, put it as plainly as anyone has at AI Ascent 2026: "We have no more manually written code anywhere at the company. All of the SQL is written by models. Everything is just built by the models." Boris personally crossed to 100% AI-written code in October and November of 2025. He now ships "usually a few dozen PRs every day" and hit "150 PRs in a day" recently, trying to see how far he could push it.

And here's the part that should keep every legacy vendor's product team up at night: the lead isn't the technology. The same models are available to everyone. Boris was explicit about this. "The place that we're ahead is not the technology, because the same technology available to us is available to everyone here. I think there's actually a far bigger lead in kind of the organizational structure and organizational process." That is the gap a 39-year-old metadata platform cannot close by issuing a press release.

Here's the rub: until these legacy companies tear down their existing platforms, and find leaders who actually use the tools driving the shift, they will never be able to build and write software the way that Anthropic and others are doing, which is AI-first. Over time, capabilities will exponentially scale in favor of companies that are building this way, and these legacy providers will either be forced to acquire or be acquired. It will be not just Workday's last Workday, but Oracle's, SAP's, and ServiceNow's as well.

The market is already validating the thesis. People are calling it the SaaSpocalypse. The better name is the Configuration Collapse. For 30 years, vendors built configurable platforms and SIs configured them at enterprise scale. The configuration era priced labor instead of software. AI flipped that math. Configuration is now the expensive path, and custom code is the cheap one. The platforms built around the old math, and the consulting workforce priced against it, are both structurally exposed. The layoffs are the leading indicator. The architecture is the lagging one.

YC and a16z are calling it early because they are investors of the future. Mark Cuban is calling it from the public-company seat. Boris Cherny is calling it from inside the lab that is building the replacement. Wall Street is calling it the SaaS-pocalypse”. We are calling it the Configuration Collapse because it names the specific thing that is dying, not just the category. Pick a name. The phenomenon is the same.