The Platform Papers

The Day the Seats Disappeared

How a folder of markdown files eraased $285 billion in market value within four days — and what the last forty years of computing say happens next.

📍 Atascocita, TX 🕐 22 min Feb 2026

1. Background

At 9:31 a.m. Eastern on Thursday, February 6, 2026, the opening bell had barely stopped vibrating when the iShares Expanded Tech-Software Sector ETF — ticker IGV, the benchmark that tracks the combined weight of every major SaaS company in America — began its eighth consecutive losing session. By the close, the ETF had posted its sharpest weekly decline since the 2008 financial crisis. Bloomberg, CNN, and CNBC were all running variations of the same headline. The word they kept reaching for was unprecedented.

It wasn’t. The pattern was forty years old. But most of the people selling that morning had never seen it before, because the last time it happened, the thing being destroyed was the mainframe market.

If you manage a software budget, you watched your vendors’ stock prices collapse and wondered whether your renewal negotiations just changed. If you run a PE portfolio, you watched your SaaS holdings lose a third of their value in a week and wondered whether the multiple you paid is recoverable. If you write software for a living, you watched an AI system demonstrate that it could replicate your most billable workflows and wondered what that means for the next five years of your career. And if you’re an enterprise architect — someone whose job is to understand how the technical pieces connect — you watched a research preview from Anthropic and understood, perhaps before anyone else, that this wasn’t a product launch. It was an architectural statement.

This is the first installment of a five-part investigation into what happened on February 6, why it happened, and what it reveals about the structural transformation underway in enterprise software. Each installment examines the event from a different angle. Each stands alone. Together, they tell the story of a pricing model that assumed humans would always be the unit of work, and the week the market realized they wouldn’t be.

Transparency Notes

View the sources for Part 1 (This Post) → Trust but verify.

This content was produced with the help of AI. I believe it is important to declare I had help in the same way I should cite my sources.

2. Problem Statement

The entire SaaS industry — $285 billion in lost market capitalization, thousands of companies, millions of seats — is built on a single assumption: the human user is the irreducible unit of software consumption. Every contract, every renewal, every expansion deal is denominated in seats. Salesforce charges per user. ServiceNow charges per user. Microsoft 365 charges per user. The seat is the atom of SaaS revenue, and it has been since Marc Benioff first put a CRM in a browser in 1999.

On January 12, 2026, Anthropic released Claude Cowork as a research preview — an AI agent with read, write, and create access to local file systems. On January 30, Anthropic followed with Cowork Plugins: eleven open-source starter packages covering legal, finance, sales, marketing, customer support, and other high-value professional workflows. On February 2, the legal plugin was demonstrated processing contract reviews and NDA triage — work that enterprise legal departments currently pay for by the seat across platforms like Thomson Reuters, LegalZoom, and ContractPodAI.

The market’s reaction was not to the product. It was to the architecture underneath it.

What Anthropic had released was not a chatbot with new features. It was a system called Skills — organized collections of files that package composable procedural knowledge for agents. In a presentation titled “Don’t Build Agents, Build Skills Instead,” two Anthropic engineers, Barry Zhang and Mahesh Murag, described what amounts to a redefinition of the unit of software work. A Skill is a folder. It contains markdown instructions, Python scripts, and metadata. It can be versioned in Git, shared in Google Drive, or zipped and emailed. It is, by design, the simplest possible container for expertise — and that simplicity is the point.

Because when the unit of work is a folder that any agent can load at runtime, the unit of work is no longer a human being. And when the unit of work is no longer a human being, the per-seat pricing model has no foundation.

The question this series investigates: Is the February 6 selloff the first visible crack in a façade where disruption has already occurred, but has yet to be fully reflected in quarterly earnings calls?

3. Impact

The numbers are unambiguous. Between February 3 and February 6, 2026, approximately $285 billion in market capitalization evaporated from the software sector. The S&P Software Index posted its worst stretch in nearly two decades. Jefferies analyst Jeffrey Favuzza coined the term that stuck: SaaSpocalypse. He described the selling as “get me out” style liquidation — not a rotation, not a rebalancing, but a categorical rejection of an entire business model.

The damage was not distributed evenly, but it was distributed broadly:

Thomson Reuters fell 15.83%. LegalZoom dropped 19.68%. HubSpot declined 39% year-to-date. Atlassian fell 35%. Asana lost 59% over twelve months. DocuSign shed 52%. Even the companies executing well were punished: ServiceNow, which reported a 98% renewal rate and 21% revenue growth in Q4, saw its stock drop more than 10% on what management called “the strongest quarter in the company’s history.” Oracle, which grew its cloud business 52% year-over-year, fell 30.3% year-to-date. Salesforce, which reported $1.4 billion in Agentforce ARR and was actively building the agentic future, still dropped 26%.

View the interactive Selloff Spectrum → Nine companies. Four days. Color-coded by exposure layer.

The median SaaS company was trading at 5.1x enterprise value to revenue by December 2025 — down from 18–19x at the 2021 peak. That’s a 70% compression in valuation multiples over four years, and the February selloff pushed the number lower. Private SaaS multiples told the same story: from roughly 41x in Q3 2021 to 4.5–4.7x by early 2026.

The international contagion was immediate. In India, where a $300 billion IT services industry employs 5.67 million people on a labor arbitrage model — billing Western clients for human hours — the Nifty IT index plunged nearly 9% in its worst week in years. TCS, Infosys, and Wipro collectively lost trillions of rupees in market value. The threat was existential: if AI agents can complete tasks 55% faster than humans, the billable-hour model that built Bangalore’s skyline faces a structural reset.

But the most revealing data point was not a stock price. It was a survey. Gartner polled 2,501 CIOs and found that 91% were increasing their generative AI funding, with a mean budget increase of 38%. Only 1% were cutting. And the only category of IT spending that was declining? On-premises software. The money was already moving. The selloff was the market catching up.

4. First Attempt

The market’s first attempt at processing the selloff was the crudest possible instrument: sell everything with a per-seat model. This was understandable. It was also wrong — not in direction, but in granularity.

In the days following the selloff, a loose consensus formed among the analysts and investors trying to make sense of the damage. The optimists pointed to fundamentals. Bank of America’s Vivek Arya argued that the market was “pricing mutually exclusive scenarios” — simultaneously punishing hyperscalers for spending too much on AI infrastructure and punishing SaaS companies for being disrupted by that same infrastructure. Both things couldn’t be true at the same time, Arya reasoned. Gartner’s official position was that “predictions of the death of SaaS are premature,” noting that AI agents automate “task-level knowledge work” but won’t replace “core systems” overnight. Janus Henderson described 2026 as a “sorting mechanism, not an extinction event.”

The pessimists pointed to architecture. Jason Lemkin at SaaStr laid out the seat-compression arithmetic in plain language: “If 10 AI agents do the work of 100 reps, you need 10 seats, not 100 — that’s a 90% reduction.” Fidji Simo, CEO of OpenAI’s enterprise division, described existing SaaS tools as “reinforcing silos upon silos” and argued that Frontier — OpenAI’s enterprise platform, launched the day before the selloff — treated AI agents like employees with their own identity, onboarding, and permissions. The Register reported Forrester analyst Charlie Betz offering a characteristically measured take: “The economics of SaaS may change, but the idea that it evaporates? Nah.”

The first attempt, in other words, was a binary debate: SaaS is dead versus SaaS is fine. Neither camp examined what specifically about the Skills architecture made it different from previous AI product launches. The chatbot era had produced dozens of “Copilot” announcements without triggering a $285 billion selloff. What changed?

To answer that, you have to look inside the folder.

5. Lessons

The reason the Skills architecture triggered a market event where previous AI launches did not is a property that enterprise architects recognized immediately and that financial analysts mostly missed: progressive disclosure and runtime composability.

In their presentation, Zhang and Murag explained that Skills are “progressively disclosed.” At runtime, the agent sees only lightweight metadata for each available Skill — a name and a short description. When the agent determines it needs a specific Skill, it reads in the full instruction set. Everything else — scripts, templates, reference data — is organized in the file system and loaded on demand. This means a single general-purpose agent can carry hundreds or thousands of Skills without saturating its context window. It loads expertise the way a human opens a reference manual: only when the task requires it.

This is not how enterprise software works. Enterprise software encodes expertise in application logic that requires a human to navigate an interface. Salesforce doesn’t give you a folder of sales procedures; it gives you a CRM with fields, workflows, and dashboards that a trained human operates. ServiceNow doesn’t give you a folder of IT service management processes; it gives you a ticketing platform with SLA engines and approval chains. The value of these systems is inseparable from the human using them — which is why they charge per seat.

Skills decouple the expertise from the interface. When a legal compliance Skill contains the procedural knowledge for contract review — the same knowledge that a paralegal learns over years and executes through a Thomson Reuters interface — the agent doesn’t need the interface. It needs the data. The application layer, the thing you pay for per seat, becomes optional. Not useless — the data it stores is still valuable. But the interaction layer, the thing that justifies the seat price, is commoditized.

Zhang and Murag made an analogy to computing history that deserves close attention. They compared models to processors, agent runtimes to operating systems, and Skills to applications. “A few companies build processors and operating systems,” they said, “but millions of developers have built software that encoded domain expertise and our unique points of view.” The implication: the application layer of the agentic era won’t be controlled by the same companies that control the application layer of the SaaS era. The expertise is being repackaged in a format — a folder — that anyone can create, version, and distribute.

This is what the market priced on February 6. Not a product demo. A platform shift. The same pattern that IBM survived (barely), that Microsoft navigated (eventually), and that dozens of once-dominant companies did not survive at all.

6. Adjustment

The lesson from the first week of the February selloff is that the selloff was directionally correct but categorically blunt. It punished every company with a per-seat model equally, without distinguishing between two fundamentally different types of SaaS value.

Interface-layer SaaS derives its value from the user experience — the dashboards, the drag-and-drop builders, the workflow visualizations that humans interact with. When an AI agent can bypass the interface entirely by accessing the underlying data through a protocol like MCP, this value evaporates. Examples include document management platforms, basic CRM functionality, email triage tools, scheduling software, and simple analytics dashboards. These companies saw the sharpest declines: LegalZoom (-19.68%), Asana (-59% over twelve months), DocuSign (-52%).

System-of-record SaaS derives its value from the data model, the compliance infrastructure, and the integration depth that makes it the authoritative source of truth for a business process. AI agents don’t replace this — they depend on it. An agent executing a contract review still needs the document management system’s repository. An agent processing a financial reconciliation still needs the ERP’s general ledger. The system of record becomes infrastructure that agents consume, not an interface that humans navigate. Companies with deep system-of-record positions — Veeva in life sciences, Palantir in defense, Workiva in regulatory compliance — showed relative resilience.

The adjustment, then, is analytical: stop asking “will AI disrupt this company?” and start asking “where does this company’s value reside?” If the value is in the interface layer, exposure is high and immediate. If the value is in the data layer, the company has a transition window — but not immunity, because AI-native data architectures are also emerging. The question becomes one of timing and execution, not of whether disruption arrives.

This is exactly the analytical error that was made in every prior platform transition. And the history is instructive.

7. Solution

The February selloff is not the first time a dominant software pricing model collapsed under the weight of an architectural shift. It is the third. And the first two provide a precise template for what happens next — including the timelines, the casualties, and the survivors.

Each transition follows the same two-beat structure: an infrastructure moment that makes the shift technically possible, followed by a product moment that makes it culturally inevitable. The infrastructure moment is for engineers. The product moment is for everyone else. And the gap between the two has been compressing with each successive wave.

View the interactive Wave Compression timeline → Three waves. Same pattern. The gap is the clock.

Wave 1: Mainframe to Personal Computing (1977–2002).

The infrastructure moment was Intel’s 8080 processor in 1974 — the first chip capable of powering a general-purpose microcomputer. The product moment came three years later, in 1977, when three machines launched simultaneously and broke computing out of the hobbyist garage: the Apple II, the Commodore PET, and the Tandy TRS-80. Historians call it the “1977 Trinity.” The Apple II was the one that stuck — not because it was the most powerful, but because Steve Wozniak designed it with color graphics, expansion slots, and a floppy disk drive that made it useful to people who had never touched a mainframe terminal. It was promoted, in its debut year, as “an extraordinary computer for ordinary people.”

In 1981, IBM held 62% of the mainframe market and dominated enterprise computing. By 1984, desktop computer sales ($11.6 billion) exceeded mainframe sales ($11.4 billion) for the first time — the crossover point that signaled the old model was losing its structural advantage. By 1992, IBM reported an $8.2 billion annual loss — at the time, the largest in American corporate history. Its stock fell from $43 to roughly $10. Its workforce was halved, from 405,000 to 220,000. Digital Equipment Corporation, which peaked at 125,000 employees, was acquired by Compaq in 1998 for $9.6 billion — a fraction of its peak valuation. Wang Laboratories filed for bankruptcy in August 1992.

IBM survived because Lou Gerstner, who took over in 1993, had the intellectual honesty to recognize that the mainframe business model was dying and the operational discipline to rebuild around services and middleware. He grew the company’s market cap from $29 billion to $168 billion. Most of IBM’s competitors did not have a Gerstner.

The wave crested and broke with the dot-com bust of 2000–2002. The NASDAQ peaked at 5,048 on March 10, 2000 and fell 78% by October 2002. A majority of dot-com companies ceased trading. The index did not recover its peak for fifteen years — until April 2015. But the bust was not just a speculative correction. It was the burial marker of the client-server era. The companies that emerged from the wreckage — Amazon, which had launched AWS infrastructure internally; Salesforce, which had launched its cloud CRM in 1999; Concur, which pivoted from packaged software to pure SaaS in the crash’s aftermath — were already building the architecture of the next wave. The seeds of Wave 2 were planted in the dying soil of Wave 1.

Wave 2: Client-Server to Cloud-Native SaaS (2007–2022).

The infrastructure moment was Amazon Web Services, which launched its Elastic Compute Cloud in March 2006 — the first commercially viable on-demand cloud computing platform. The product moment came less than a year later, on January 9, 2007, when Steve Jobs walked onto the Macworld stage and introduced the iPhone. Sixteen days later, on January 16, Netflix launched its streaming service. A Moody’s analyst later described the convergence: “The timing finally worked for all of these things that had been percolating and talked about for over a decade. It finally started coming together in 2007 where you had a combination of tech available to make that device portable and able to do everything that Steve Jobs envisioned.”

2007 was not the year cloud computing was invented. It was the year cloud-native delivery became the only rational business model. The iPhone proved that software could be consumed anywhere, on any screen, without installation. Netflix proved that subscription-based delivery at consumer scale could destroy an entrenched physical-media incumbent — Blockbuster filed for bankruptcy three years later. Together, they established the cultural expectation that software should be always-on, always-updated, and billed by the month. SaaS companies built their empires on this expectation.

The consequences for the incumbents were devastating, but slow. Microsoft under Steve Ballmer stagnated at roughly $300 billion in market capitalization for fifteen years. Ballmer had all the resources, talent, and market position in the world, but he could not cannibalize the Windows and Office licensing cash cows to pursue a cloud-native architecture. Satya Nadella, who replaced Ballmer in 2014, could. He drove the stock from $32 to over $500 — a 15x return — and grew the market cap from $300 billion to over $3 trillion by pivoting the entire company to cloud-first. Oracle spent more than a decade rebuilding for cloud. SAP is still migrating: more than 23,000 companies need to complete the transition to S/4HANA by a 2027 deadline.

Wave 2 ended on November 30, 2022, when OpenAI launched ChatGPT. Within two months it became the fastest-growing consumer application in history, reaching 100 million users. The enterprise response was immediate: Microsoft pledged billions in OpenAI investment; Salesforce, Amazon, Google, and Oracle rushed to announce their own AI initiatives. TechCrunch called it the moment AI “took over the world.” It was — but not because of a chatbot. It was because ChatGPT demonstrated, to a mass audience for the first time, that natural-language interfaces could replace graphical ones for a wide range of knowledge tasks. The per-seat model depends on humans navigating graphical interfaces. ChatGPT showed that the navigation itself was becoming optional.

Wave 3: SaaS to Agentic (November 2024–present).

The infrastructure moment was Anthropic’s release of the Model Context Protocol on November 25, 2024 — an open standard for connecting AI agents to external data sources and tools. MCP solved what engineers called the “M×N problem”: if you have M AI applications and N tools, you previously needed M×N custom integrations. MCP reduced this to M+N. The protocol was adopted within twelve months by Anthropic, OpenAI, Google DeepMind, and Microsoft. In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. The New Stack wrote that it is “difficult to think of other technologies and protocols that gained such unanimous support from influential tech giants” at comparable speed.

MCP may or may not survive as the permanent protocol for agent-to-tool communication — early internet protocols were also superseded. But it established the paradigm: a standardized interface layer that allows any agent to access any data source without requiring a human to operate the application in between. That paradigm is the structural equivalent of HTTP for the web era. The protocol is debatable. The pattern is not.

The product moment followed fourteen months later, in January 2026, when Anthropic launched Skills and Cowork — packaging composable procedural expertise in folders that any agent could load at runtime. The gap between infrastructure and product compressed again: three years for Wave 1 (Intel 8080 to Apple II), less than one year for Wave 2 (AWS to iPhone), fourteen months for Wave 3 (MCP to Skills). The compression itself is part of the evidence base for acceleration.

The data confirms the acceleration is real. AI startups are reaching $5 million in annual recurring revenue approximately 1.5x faster than the top SaaS companies of the 2018 cohort. Y Combinator reports a 60% decrease in MVP development time compared to 2022. Enterprise spending on generative AI grew from $1.7 billion to $37 billion in less than two years. The AI agent market is projected to grow from $7.84 billion in 2025 to $52.62 billion by 2030, a 46.3% compound annual growth rate. Bain & Company’s technology practice wrote the line that best captures the acceleration: “In three years, any routine, rules-based digital task could move from ‘human plus app’ to ‘AI agent plus API.’”

If the three-wave pattern holds, the SaaS-to-agentic transition will produce the same three categories of outcome. Gerstners — companies that recognize the shift early, accept the revenue cannibalization, and rebuild around the new architecture. DECs — companies that execute too slowly and become acquisition targets at a fraction of their peak valuation. And Wangs — companies that fail to transition at all and go to zero.

Wave 1 gave you 25 years from the Apple II to the dot-com bust. Wave 2 gave you 15 years from the iPhone to ChatGPT. The evidence from Wave 3 suggests the window is 5–7 years from the infrastructure moment. If MCP was the starting gun — November 2024 — then the window for strategic positioning closes somewhere between 2029 and 2031. The companies that audit their position in 2026 will have time to pivot. The ones that wait for the quarterly earnings to confirm the thesis will be making the same mistake that IBM made in 1985, that Ballmer made in 2008, and that Blockbuster made in 2007.

8. Three Takeaways

Takeaway 1: The unit of value in enterprise software has shifted from the user to the task. (From the problem.)

The per-seat pricing model is not being disrupted by a better product. It is being disrupted by a redefinition of what software is for. When Anthropic packaged expertise as a folder of markdown files and Python scripts — and when that folder could be loaded by any agent at runtime, composed with other folders, and executed without a human navigating an interface — the unit of software consumption shifted from “the person who uses the tool” to “the task the tool performs.” This is not a pricing adjustment. It is an ontological change in how enterprise value is created and captured. Every SaaS contract denominated in seats is now denominated in a unit that is structurally shrinking.

Takeaway 2: Each platform wave follows the same two-beat pattern, and the market always gets the first beat wrong. (From the learning.)

In Wave 1, the market kept buying IBM at $43 while desktop sales were already surpassing mainframe sales. In Wave 2, the market valued Ballmer’s Microsoft at $300 billion for fifteen years while Nadella’s cloud-first vision was already proving out at Amazon. In Wave 3, the February selloff punished ServiceNow — a company with 98% renewal rates and deep system-of-record positioning — with the same ferocity it applied to LegalZoom, whose entire value proposition is an interface layer over legal document templates. This is the hallmark of a first-wave panic: the market correctly identifies the direction of change but cannot yet distinguish between the exposed and the insulated. The analytical work of the next six months is distinguishing between interface-layer SaaS (high exposure, immediate risk) and system-of-record SaaS (transition window measured in years, not quarters) — and acting before the market refines its aim.

Takeaway 3: The infrastructure-to-product gap is compressing, and the compression is the clock. (From the solution.)

Intel 8080 to Apple II: three years. AWS to iPhone: less than one year. MCP to Skills: fourteen months. But the more important compression is the full wave duration. Wave 1 ran 25 years (1977–2002). Wave 2 ran 15 years (2007–2022). If the pattern holds, Wave 3 runs roughly 7–10 years from its infrastructure moment — placing the window’s close somewhere around 2031–2034. The companies that moved early in each prior wave (Gerstner at IBM, Nadella at Microsoft, Netflix over Blockbuster) captured generational returns. The companies that moved late (DEC, Ballmer-era Microsoft, Blockbuster) became case studies in strategic obituaries. The window for Wave 3 positioning is open now and will narrow faster than any previous transition. This is not a prediction. It is a pattern that has executed twice before within living memory, and the pace data says it is executing again, faster.

9. Call to Action

If the analysis above is even directionally correct, it demands two responses — one technical, one financial — and neither can wait for the next budget cycle.

The Technical CTA: Audit your software stack against the Skills architecture.

For every SaaS tool your organization pays for by the seat, answer one question: does this tool’s value come from its interface or from its data model?

If the value is in the interface — if the tool is essentially a dashboard, a workflow builder, a visualization layer that a human navigates — then an AI agent with the right Skill can bypass that interface entirely by accessing the underlying data through a protocol like MCP. That tool is in the high-exposure category. It doesn’t mean you cancel the contract tomorrow. It means you flag it for parallel evaluation: begin testing whether an agent-based workflow can replicate the outcome at a fraction of the seat cost.

If the value is in the data layer — if the tool is a system of record with compliance infrastructure, proprietary data structures, or regulatory certification that no agent can replicate — then the tool is in the lower-exposure category. But “lower” is not “zero.” The regulatory moats that protect healthcare IT, defense, and financial services create a 2–5 year buffer, not permanent immunity. Use that buffer to evaluate whether your system-of-record vendors are building MCP-native agent access into their platforms. If they are, they’re positioning themselves as infrastructure for the agentic era. If they aren’t, they’re assuming the moat is permanent — the same assumption IBM made about mainframes in 1985, and the same assumption Blockbuster made about physical stores in 2007.

The deliverable is a three-column inventory: tool name, annual seat cost, and whether value resides in the interface layer or the data layer. This exercise takes a day. It will inform every software procurement decision your organization makes for the next three years.

The Business CTA: Model your seat-cost exposure as a percentage of operating expense.

Calculate what percentage of your total software spend is billed per-seat versus per-usage or per-outcome. If seat-based spend exceeds 60% of your software budget, you have concentration risk in a pricing model that IDC, TSIA, and Gartner all project will be fundamentally refactored within three years. IDC forecasts that 70% of vendors will shift to consumption or outcome-based pricing by 2028. TSIA’s analysis concludes that “AI makes user-based pricing unsustainable.” Gartner projects that 30% or more of enterprise SaaS will include outcome-based pricing components by the end of 2025.

Flag this number for your next board or leadership review. Frame it not as a crisis but as a planning input — the same way a CIO in 1985 might have flagged mainframe dependency as a percentage of the IT budget, or the same way a media executive in 2007 might have flagged physical distribution costs in the quarter the iPhone launched. Netflix modeled its streaming economics while it was still mailing DVDs; by the time Blockbuster understood the shift, the window had closed. The companies that modeled their exposure early in each prior wave captured the cost savings of the transition. The companies that deferred the analysis became DEC — acquired at $9.6 billion, a fraction of peak value — or Wang Labs — bankrupt, assets auctioned, name forgotten.

The February selloff was not a single event. It was a signal — the market’s first, crude attempt to price a transition that is already underway. Three waves of platform disruption have now followed the same pattern: an infrastructure moment that engineers recognize, a product moment that everyone else recognizes, a market panic that punishes indiscriminately, and then a sorting that separates the Gerstners from the DECs from the Wangs. We are in the early innings of Wave 3. The question is not whether the per-seat model is dying. The question is whether you will be a Gerstner, a DEC, or a Wang. The audit and the financial model are the first steps toward answering that question.

Part 2 of The Platform Papers — “The Protocol Is the Product” — examines how the Model Context Protocol is commoditizing the application layer, and which verticals break first when agents bypass the UI entirely.