“Run-fast” companies after LLMs.

Before LLMs, building software was slow and expensive. Duplicating a broad product cost millions, so product surface area was itself a moat. SaaS companies leaned on "running fast" to deter competition, accumulating enough product and share that entering the market as a copycat was unappealing. Durable moats like switching costs and network effects came later, layered on top of this early lead.

But now, LLMs spit out code and features at a pace that collapses the moats of any pre-AI company, and all SaaS companies are more vulnerable today than three years ago. 

That has led application investors to panic, questioning the continued effectiveness of the "run-fast" playbook overall. Every run-fast category, the nervous VC imagines, ultimately becomes a bloody mess of vendors who have all "solved" the problem. (A more grounded version of this fear: even where a winner emerges, pressure from the models, copycats, and bundlers erode pricing power.)

But the run-fast playbook isn't doomed, nor is the application layer. The upshot of cheap code isn't that running fast gets you nowhere; it's that you'll need to run a lot further before "running fast" can yield a compelling moat against the models or startup competition. And that distance requirement keeps stretching as models improve, cheapening the progress made to date.

To run further for longer, today’s companies have to claim bigger problem spaces than their SaaS predecessors did. A problem space that may have represented a midsized TAM in the old days is often now something that can easily be built and shipped as a feature from a competitor in an adjacent category (--one that might swallow this midsize TAM whole).

So today's application-layer winners often need to bundle, pulling together adjacent categories that share a buyer until the combined surface area is too wide to be fully solved soon. The catch is that bundling cuts both ways, and the best defense against bundling is often offense, which has led to the proliferation of “kingmaking rounds.” 

(Of note is that this all has a portfolio construction upshot for the VC. Since each bundled winner absorbs multiple former venture outcomes, the application layer still produces winners, and likely bigger ones, but certainly fewer.)

Better models also turn architecture into a moving target. An advanced architecture choice today puts a company ahead for the short term, but it may become obsolete over the course of months. And so winners will also need to regularly revisit what they have already built and reinvent themselves. Losers risk getting fiercely challenged when a new entrant gets to the new architecture faster.

None of this means the application layer is doomed. Running fast was always a first moat. Long-lived SaaS companies layered on more durable ones over time, like switching costs and network effects; application-layer winners will probably lean on those second-order moats even more heavily once they exist. But they take years of scaling to emerge, and a Series A investor won't see them yet.

Which leaves the practical question for the early stage investor: how do you diligence a “run-fast” Series A opportunity today? I’d propose four criteria:

Demand: A ripping wedge running on the current state of the art architecture. 

Many buyers just want AI for AI's sake right now, and producing a magic demo has rarely been easier. That means the bar for demonstrating demand is higher. The new floor for traction is $1M to $5M of ARR (or $0 to $1M in two quarters, followed by $1M to $3M in the two after).

The ‘state of the art’ bit may feel a bit of a throwaway line, but the point is to acknowledge that while AI is constantly shifting, the investor should ensure the startup’s technology is at least as up-to-date as it can be.

Category: Expansive product runway and the vision/aggression needed to win the bundling wars.

The startup must face lots and lots of runway ahead to build new stuff, and keep building new stuff. There is no greater horror than running out of runway for your product roadmap.

And the investor must look to adjacent, bundle-able categories as direct competition. Bundling is a magnitude more common and compelling today than in SaaS, and it is easy to find out too late that what you are investing in is a subset of another category.

Offense is often the best defense against bundling. Aggressive bundlers recognize this game, raising capital and hiring ahead of revenue to ship into adjacent categories early, and prioritizing surface area over perfection.

I’d also note that not all bundling makes sense. ElevenLabs is launching AI phone calling, which requires competing with a number of its customers and selling to a whole new ICP. This seems like a much more challenging expansion to pull off than, say, Listen Labs bundling in simulation (a la Simile/Aaru) on the same data they have already collected for an existing customer.

Winners: A team resilient enough to burn the boats of an old architecture, and dynamic enough to see ahead of the technological curve.

Model capabilities jump every twelve to eighteen months and reset what's possible. "Set and forget" is the most dangerous founder posture in this era, and the team that won't tear down and rebuild what they shipped last year is the team that gets overrun by the next entrant who will. Founders who have already done a teardown are a bullish signal, but the sample is small. A diligence question that travels: what product decisions would you make differently if you were starting from scratch today?

Longevity: A credible path to a durable moat. 

‍

Switching costs, data flywheels, network effects all accrue with adoption, so they won't be visible at the Series A stage. What's possible at the Series A is a plausible story for how the run-fast lead converts into one of these more durable narratives over time.

It is worth noting that switching costs in AI probably look quite different from those in SaaS. Today’s switching costs are likely about business logic and processes living within the system, versus years of data in an existing database that will be hard to exfiltate.

—

Appendix: Competition with the labs

There’s a new dynamic to startup competition in the AI era, which is competition with the labs in addition to competition with other startups. Competition with the model labs comes in two forms.

The first is direct: occasionally a lab will enter a category with intention, like Anthropic's Claude for Financial Services or Claude Design. This is rare and usually narrow. 

The second is incidental and far more common: a startup picks positioning that sits in the path of the next model release, and the model itself becomes the competitor without anyone at the lab thinking about them. Covering for model deficiencies can be a good wedge, but it cannot be the durable value prop. The companies that win position themselves for a future in which the models are much smarter, while winning today by compensating for current limitations with product they expect to be made redundant by the next release.

This offers us a fifth criterion for evaluating run-fast apps:

Model improvement: A value proposition that doesn’t bet against smarter models.

Products that exist mostly to cover a current model shortfall (eg context length, reasoning depth) can show real traction today and still get washed out in the next cycle. This is not to say that covering model gaps cannot be a part of a winning strategy (in fact, it often is), but it should not be the strategy.