The 4 arguments for the death of software, ranked from worst to best

Our investors ask me two questions on most calls right now.

Is AI another bubble?

Is software dead?

The short answer to both is no, but the reasoning matters more than the verdict.

I spent a decade at Google building high-availability systems before I co-founded SignalFire with Chris Farmer. Engineers don't trust vibes. We trust data, systems, and outputs we can predict. Venture capital ran on the opposite operating model for most of its history, mainly on gut feel, warm intros, and pattern-matching on founder archetypes. We built SignalFire (and Beacon, our AI platform) on the belief that the industry would eventually be forced to operate on hard data. That moment has arrived, and it's arrived for the software industry as well.

The bubble question is the wrong question

Demand for AI is still outrunning supply. Some of our portfolio companies can’t serve all their customers because they don’t have enough compute. That demand has tens of billions of dollars in real revenue attached to it, not just eyeballs or MAUs. In my opinion, bubbles don't really look like this.

What worries me is sectors like humanoid robotics, where capital is flooding in at late-stage valuations. While I’ve seen backflips and fun dance moves, I have yet to see a robot do economically valuable work. The deeper problem is one any engineer would flag immediately: LLMs work because the open internet existed to train them on, but robotics has no equivalent corpus. Where does the training data come from? Nobody has a clean answer yet, which makes it a research problem, and research timelines don't map cleanly to venture timelines. Robotics will have its ChatGPT moment, but I’m not sure when, and neither does anyone raising funding on the premise that it's happening soon.

The four arguments for software's death, ranked

The top concern I hear from founders and LPs is about the $285 billion that was recently wiped off legacy software market caps. People want to know if SaaS is structurally finished. I've heard four arguments for that thesis, and since we at SignalFire put a lot of capital into the application layer, I thought I’d share my read on the so-called SaaSpocalypse.

Here are the four most common arguments for software's death, ranked from worst to best:

#4: Everyone will vibe code their own software

The pitch I hear is: "Why pay Salesforce when you can vibe code your own CRM in a weekend with Claude?"

Anyone who has shipped and operated a production system knows the answer to this question. Generating a codebase is not the same as running a mission-critical service. When the person who vibe-coded your CRM leaves, what happens to the codebase? Code generation does not address issues like SOC2 compliance or hallucination control. It does not solve integration with a SQL database someone wrote in 1998, and it most definitely does not solve uptime accountability when the dashboard goes dark at 4 a.m. and revenue stops moving for customers.

Enterprises buy trust, not code. Code parity is easy now with AI, but trust parity is still very hard to achieve.

#3: Agents like Claude and ChatGPT will swallow enterprise apps

A better argument than vibe coding, but I'm skeptical it will work for any workflow where being wrong is expensive. While frontier models are impressive, LLM systems are non-deterministic and prone to hallucination. A normal software bug is reproducible. You file the ticket, find the line of code, and ship the fix. An agent failure is more like a flaky test that passes 98% of the time and fails on the one run that mattered. 

That tradeoff is fine for low-stakes work like drafting an email, summarizing a doc, or writing marketing copy, but the first time you lose a six-figure deal because the agent forgot to populate a required field, or it confidently logged the wrong contract value, you're going to want a system with a dedicated product surface that enforces the rules.

I've watched this play out in our portfolio. Teams experimenting with pure agent-driven workflows almost always end up rebuilding constraints around them, including validation layers, approval steps, rollback mechanisms, and audit logs that the LLM can't touch. By the time you've added all that, you've rebuilt a SaaS app around your agent. Instead of replacing the app, you've nested an agent inside a vertical application. The application owns the data model, permissions, audit trail, and customer relationship. The agent operates inside that envelope and gets more powerful over time, but the envelope is what gets sold, supported, and renewed.

This gets more interesting as model reliability improves, but we're not there yet.

#2: As seat-based pricing goes away, it breaks the SaaS model

This one is closer to something more real. Traditional SaaS has three layers: data, business logic, and UI. We're now stacking a fourth onto it, “the agentic layer.” If you used to sell 50 seats to humans, and now two agents do that work without the UI, what happens to your pricing power?

Here's where I split from the doomsday version. If agents are doing more work inside your software and the customer is getting more value, you don't have an existential problem. What you have is a pricing and packaging problem. Vendors who figure out how to price to value, through tokens, outcomes, or hybrid usage models, will be fine. Vendors who cling to per-seat pricing while their customers automate the seats away will not.

But pure outcome-based pricing is hard to operationalize cleanly across every category, which is why I expect hybrid models (usage + outcome) to dominate for the next few years.

#1: Code is now cheap, which erodes feature moats

This is the argument I take most seriously. The moat for companies like SAP, ServiceNow, and Salesforce was, in large part, engineering headcount that compounded over decades. Every conceivable feature, integration, and report was built, debugged, and absorbed into a codebase that a startup couldn't realistically catch up to. AI now compresses that timeline dramatically.

If your product is a pure workflow layer, and your defensibility story was "we built the feature first," you are in trouble. Business intelligence and creative content generation are where I see the thinnest moats today, as LLMs are extremely good at exactly the work those products do.

Where the moats actually live now

If feature velocity isn't a moat, then what is? Based on what we're underwriting at SignalFire today, here are three things that hold up:

1. High-accuracy workflows where the error budget is near zero. For example, financial infrastructure, healthcare, and regulated compliance. Vibe coding does not survive a HIPAA audit or a reconciliation break. The cost of being wrong is the moat here.
2. Proprietary data feedback loops. Products that get meaningfully better as customers use them in ways a competitor cannot replicate by spinning up the same foundation model. The data, not the model, is the asset here.
3. Deep systems of record. Software that embeds into legacy operations, owns the source of truth, and creates high switching costs. These companies should be running toward AI, not hiding from it. The agentic layer actually makes their data far more valuable.

State of the AI stack: Where we're investing and where we're not

Saying "we invest in AI" is now about as differentiated as "we invest in software" was in 2012. Here's how we think about the AI stack at SignalFire.

There are four layers, and they do not deserve equal capital.

1. Hardware: Compute is still the binding constraint in this cycle. We see it playing out at companies like Neolabs, which are waiting on POs that should have cleared months ago. The demand is real, the supply is gated, and the winners at this layer are mostly public or already at scale. A seed check does not change the outcome here.

2. Models: In my opinion, frontier models are a capex business, not a venture business. The cost to train one really well rivals the GDP of a small country, and the firms that can run that play, like OpenAI, Anthropic, and Google, are already running it. A model-layer startup trying to attack them head-on is picking the wrong fight. The few companies we've backed at this layer, like Sciforium, are doing something different in shape with novel architectures, novel training and inference approaches, and novel problem framings. If you're pitching us a "frontal assault” on ChatGPT, that’s tough. If you're outflanking the model labs with an approach the incumbents can't easily copy because it would cannibalize their core, we're all ears.

3. Infrastructure: This is where AI breaks the old SaaS assumptions in genuinely interesting ways. Traditional SaaS was read-heavy, where you stored a row and queried it a million times. AI workloads invert that. Training pipelines, agent memory, vector stores, and eval harnesses are write- and update-heavy and operate on data shapes the legacy stack was never designed for.

Here, we're investing actively on both sides:

- The data layer itself, with companies like PlanetScale and Greybeam, which are rebuilding the OLTP and OLAP database primitives for this shape of workload.

- The data generation layer, with companies like Preference Model, Moody Pines, and Terac, that supply clean, structured, labeled data, which is the actual bottleneck once compute is solved.

4. Applications. This is where most of our capital goes today. The application layer is where value actually reaches a buyer. The surface area is enormous because AI is now feasible in workflows that were previously untouchable, such as medical coding, freight optimization, legal review, sales motions, and clinical operations. Competition is fierce, and pricing is under pressure (see the seat-based argument above), but that's a feature of any layer with this much addressable demand. The winners will be the ones who pair an AI-native workflow with one of the moats above: a high-accuracy environment, a proprietary data loop, or a deep system of record.

A note for founders building in this market

If you're trying to figure out where your startup sits, the test is simple. Hardware and models are mostly closed to new entrants without a structural edge. Infrastructure is wide open if you've internalized that AI workloads are not SaaS workloads with a chatbot bolted on. And applications are where the volume is, and where the bar is rising fastest.

If you're at seed or Series A, here's what I'd actually do. Stop pitching your moat as "we ship fast," as that is now table stakes. Instead:

- Pitch the data you accumulate that no one else can get.

- Pitch the workflows you're absorbing that the customer cannot rip out.

- Pitch the accuracy your system delivers in environments where being wrong is expensive.

If you're a feature sitting between two larger applications, you're in the dead zone, and you should know it.

Software is not dying, but the bloated utility version of it is being repriced in real time, and what replaces it will be harder to build than a weekend project with a model API. That's the part I find genuinely exciting. The bar for what counts as a real company just went up.