Software after the SaaSpocalypse

Hi, it's Michael from firstminute capital. We are a $500m seed fund backed by 130 unicorn founders. Along with my brilliant colleagues Sam, Lorcan, and Adriana, I invest in European pre-seed and seed AI companies like Granola, Vocca and Alecto.

---

Last week I built an AI-native version of the Financial Times, a 600-person company that has been going for 138 years — and that I used to work for. I also built an AI-native consultancy firm, taking a bite out of larger businesses like Crayon and Proactive Worldwide. I built a labour marketplace, an ecommerce business, an operating system for VCs, a productivity tool, and for good measure an art project. I did this all in a week, powered largely by diet coke and vibes.

The process was blissfully fun. But it also made me terrified, and led to many sleepless nights. Because it made me worry that the commentators are right: that we are living through the largest existential crisis in the history of tech where the terminal value of the entire firstminute portfolio — and every other venture or PE software portfolio — is in question.

On the one hand, from on high, the foundational model companies are laying waste to whole software categories by moving from selling models to owning whole workflows. Anthropic erased $830bn in market value — nearly the GDP of Switzerland, or the entire market cap of Walmart — by announcing integrations around sales, finance, legal and HR. Thomson Reuters, LegalZoom, ServiceNow, RELX: all devastated in a single news cycle.

On the other hand, random people — like me, but better — can single-handedly vibe code category killers. Just look at how Peter Steinberger, the creator of OpenClaw (pictured below looking jacked), ushered in a new era of personal assistants for all at a stroke. Semi-technical workers at companies can now use tools like CodeWords or Claude Code to build internal software they would previously have paid for. At best, there is going to be significant pricing pressure on software companies as more challengers emerge.

Are we all doomed? I want to argue that — while we are living through a period of enormous upheaval — the discourse that "SaaS is dead" and "software has no moats" is entirely overblown. It is actually the greatest time ever to be an early stage software investor. The big model companies are not going to eat all the software. Vibe-coding 18 year olds are not going to replace Salesforce.

Our house view aligns with Nvidia's Jensen Huang (pictured below), who recently said that AI replacing all software "is the most illogical thing in the world". What has happened is that a monumental new opportunity has emerged to build disruptive software that eats into labour budgets, sells to industries that were previously impossible to crack, and harnesses entirely new consumer behaviours.

The possible scenarios

Before getting into it, it helps to map the territory. There are roughly five ways the future of software plays out.

1. Total LLM dominance. The foundational model companies go after every vertical successfully and everything flows through them. Companies, labour and software is just an API call into Anthropic or OpenAI. (Unlikely)

2. Incumbent adaptation. Existing SaaS companies build their own agentic layer and capture most of the transition value themselves. The per-seat model changes. The interface changes. But they adapt. This is the historical precedent. Oracle was supposed to die in the cloud transition. SAP was supposed to die. They thrived. The companies with distribution, proprietary data, and the balance sheet to move fast own the revolution. (Very likely)

3. AI-native disruption. New companies built to be AI-first replace incumbents in the categories they are too slow or too conflicted to defend. This will not happen everywhere at once — it is probably a 10-year structural shift — but enormous value will be created by new entrants. Fantastic for early stage investors. (Very likely)

4. Enterprise DIY. Companies build their own software using AI tools and stop buying from vendors altogether. This entirely overstates the technical sophistication of most companies. The ones that can already build their own software are mostly technology companies — and they were already doing this. (Unlikely)

5. AI bubble bursts. The technology is overhyped, something goes wrong, investment dries up. Possible — every cycle has a trough. But the underlying capability is real in a way that feels different from previous waves, and the capital committed is too large to evaporate cleanly. (Unlikely)

Why scenario two and three are most likely

Our view: LLM dominance and enterprise DIY won't happen. Incumbent adaptation and AI-native disruption are the two most likely scenarios. Here is why.

AI is a feature, not a product. The core mistake in the LLM dominance scenario is treating AI as a product when it is actually a feature. In most enterprise contexts, AI without software is useful for a narrow set of tasks — image generation, coding assistants, chatbots. To unlock AI's value across a full enterprise stack, you need software to orchestrate the interaction between AI and non-AI components: routing outputs, enforcing compliance, maintaining audit trails, integrating with legacy systems. Case in point: OpenAI and Anthropic are two of the largest Salesforce customers.

Compliance and safety matter more than people think. Enterprise software that governs payroll, clinical records, financial audit trails, or compliance workflows cannot run on a system that is probabilistic by nature. When Knight Capital's trading algorithm ran unchecked for 45 minutes in 2012, it lost $440m. When CrowdStrike pushed a faulty update in 2024, the damage ran to $5.4bn. No general counsel signs off on replacing a certified, auditable system with something that is "mostly right." Real-world enterprise tasks average 380 to 540 steps; at 99.9% per-step accuracy, end-to-end success rates fall below 70%.

Labour gets eaten, not software. What the model companies will eat is likely labour, not all the software. OpenAI and Anthropic are competing for a share of the multi-trillion global labour market, not the $800bn global software market. They will take enormous chunks of white-collar work: drafting, summarising, analysis, outreach, coding. That is a huge deal — and scary, depending on where you sit. It is not the same thing as replacing CrowdStrike, a company with fifteen years of proprietary data, a compliance moat, and contractual relationships with half of the Fortune 500.

You can't vibe code enterprise software. Building software is easier than ever. Distributing it, supporting it, securing it, and integrating it into a complex existing environment is not. One of our founders building agentic non-clinical workflows for healthcare is winning over impossibly hard customers because of her decade of industry relationships, her deep and specific knowledge, and her credibility in rooms a well-funded newcomer simply cannot access. The speed at which she built the software is great, but not the point. What matters is that she understands customers and can sell to them.

Enterprises are buying more software, not less. The share of enterprise AI solutions being built internally fell from 47% to just 24% in a single year. Companies that tried to build their own AI products mostly gave up and bought instead — AI products are hard to build in a way that is genuinely usable and economical to maintain. The enterprise DIY scenario is at the moment happening in reverse.

Incumbents will adapt. The per-seat model is changing. The interface is changing. Satya Nadella declared on the BG2 podcast that business applications "will probably collapse in the agent era," because at their core they are "CRUD databases with a bunch of business logic." Sam Altman was only slightly less brutal: "every company is now an API company, whether they want to be or not." IDC predicts seat-based pricing will be obsolete by 2028. Fine. But none of this means existing software companies die — it means they have to evolve their packaging. Oracle was not supposed to survive the cloud. SAP was not supposed to survive the cloud. Microsoft was supposed to die three times. The companies with distribution, proprietary data, customer relationships, and the balance sheet to move fast have consistently owned the transitions that were supposed to kill them.

We are just at the beginning. One in three Americans has used ChatGPT. A quarter-century after Amazon's IPO, only 20% of shopping happens online. Fifty percent of US households still have cable. Technological capability arrives decades before behavioural adoption catches up. We are on level two or three of a very tall building. Scenarios 2 and 3 play out over a decade, not a news cycle.

Yes, some businesses get killed. Others get stronger.

The above should not play down the existential crisis in many software categories. Some businesses will die. To understand which ones, you need to understand what has actually changed.

For years, the best software was a sophisticated filing cabinet — a system of record that stored, organised, and surfaced your company's data. But a huge proportion of SaaS built on top of that was really just plumbing: tools that moved data from one filing cabinet to another, or put a prettier interface in front of it for humans to use. Zapier connecting your CRM to your email platform. Fivetran piping your database into your analytics tool. A whole industry charging rent for being in the middle.

Two things have now happened simultaneously. First, MCP — the Model Context Protocol, adopted by every major AI lab — means agents can connect to any data source directly, making proprietary integrations far less of a moat. Second, people are increasingly accessing software through AI interfaces rather than traditional UIs — querying their CRM through Claude rather than logging into Salesforce. If you don't own the pipes and you don't own the interface, what exactly do you have?

So what does this mean? And who should be worried?

Interface businesses are most exposed. These are tools that exist purely to help humans do tasks — file storage, document signing, basic dashboards, form builders. When an agent does the task directly, the interface becomes a toll booth that gets routed around. No user opening the app means no upsell and no renewal.

Pipe businesses also need a rethink. In an MCP-first world, generic connectivity is free. Light wrappers — thin UIs on top of someone else's API, with no proprietary data underneath — are the most acutely exposed. When the underlying model can be accessed directly, the wrapper has nothing left to sell.

Systems of record are durable, for now. SAP, Workday, Epic — pulling one out is a multi-year, nine-figure undertaking. These companies are not safe forever, but they have space to act. The ones moving aggressively on agentic integration will capture the transition. The ones that drift will erode quietly: seat counts compressing, NRR declining.

New businesses and new models will arise. The opportunity is not to attack fortified platforms head-on. It is to go where incumbents are weakest: industries that missed SaaS entirely, categories where incumbents are too conflicted to go AI-native, and the surface areas AI itself creates. Healthcare administration, construction project management, agricultural supply chains — markets where an AI-native product built by a genuine domain expert faces almost no incumbent resistance. The total opportunity is growing: Gartner forecasts global software spending hitting $1.4 trillion in 2026, and Forrester projects SaaS revenues rising from $318 billion today to $576 billion by 2029.

What this means for venture investing in 2026

The same forces killing the weakest businesses are opening up new categories that did not exist before. Here is where we are focusing.

01. AI-native services. The most underrated category. An AI-native services business delivers comparable output to a traditional consulting firm at a fraction of the cost, charges near-consulting prices, and runs at software margins. The moat is not the code — it is the workflow, the brand, the named analyst with a memory. These businesses sell outcomes, not seats.

02. Agentic systems of action. Founders with deep industry knowledge can wedge in with an agentic product that eats a labour budget, builds a data flywheel, and backs into becoming the new system of record in categories where incumbents are too slow or too conflicted to move. The agentic layer is the wedge. The data is the moat.

03. Regulated and under-digitised industries. Insurance claims, legal workflows, construction, logistics, agriculture — industries that missed the SaaS revolution not because they were unsophisticated but because building for them properly required domain expertise that pure-play software companies never bothered to develop. AI makes it possible to build for these industries for the first time, with almost no incumbent to fight.

04. AI-native cybersecurity. AI massively expands the attack surface. Every AI agent is a potential ungoverned entry point. Identity has replaced the network as the foundational security boundary. Demand for cybersecurity goes up when AI proliferates, and it shifts — away from securing code at build time toward protecting running systems at runtime. That is a new category.

05. Hardware-enabled software. When hardware gets smart — cameras that understand context, sensors that model physical environments, robots that operate in the real world — the software layer on top becomes extraordinarily defensible. The moat is not the model. It is the dataset built from real-world deployments that no competitor can replicate from a standing start.

06. Industry 4.0 and multimodal. The digitisation of physical industrial processes — manufacturing, energy, logistics — accelerated by multimodal AI that can process video, sensor data, and language simultaneously. Proprietary data, embedded physical workflows, enormous switching costs.

07. Deep tech and hard tech. The safest category in a world of cheap code is the one where the code is the least of it. Defence, robotics, energy infrastructure, regulated engineering — durable regardless of what happens to software margins. A vibe coder who built an AI newsroom in a day cannot build a certified avionics system in a year. The barriers are physical, regulatory, and institutional.

—

A friend forwarded me a piece of BusinessWeek from 1991. In it, a programmer named Eric Bergerson describes a new technology that will transform software forever. Business people, not just programmers, will build software directly.

He predicts a world of interchangeable, plug-and-play components where you shop for pieces separately and compile your own custom software. "Instead of the programmers," he writes, "the business people can make changes to the software."

The technology he was describing was object-oriented programming.

We are still here. Software did not die. My guess is that in 2057, someone will forward this essay as an artefact from a time when people thought AI was going to make software disappear. It didn't. It just made it change and evolve.