There is a lot of anxiety in early-stage software right now, and most of it rhymes with one phrase: AI is eating software. If a large language model can generate a CRUD app, draft the copy, and wire up the integrations, what exactly is a software startup worth? The fear is real enough that a popular sport in venture is now predicting how many of today's 'AI startups' will be gone within a year.
We find the doom framing lazy, but the underlying question is the right one — and we happen to have a clean way to answer it. Across our portfolio, a cohort of companies has now graduated from seed into priced Series A (and a few into Series B) rounds led by serious investors. So instead of theorizing about what survives the AI wave, we can just look at what actually cleared the bar. The pattern is unambiguous.
First, the Bar Really Did Move
Before the archetypes, the context. The Series A bar has risen roughly 40% since 2023. Where $1–2M of ARR could anchor a Series A a few years ago, the conversation in 2026 starts closer to $3–5M, with net revenue retention above 120% and months of consistent growth behind it. Investors are underwriting operational maturity that used to qualify for Series B.
At the same time the market has split. Horizontal software — the generic, cross-industry tools — is contracting as AI agents absorb coordination work natively. Vertical, industry-specific software is holding up. The reason is simple: AI makes commodity software less valuable and proprietary, domain-specific software more valuable. The same wave that erases one category strengthens the other.
What's Dying: Thin Wrappers
The casualties are clear. In early 2026 a Google VP who runs its global startup organization warned that two kinds of AI startup face extinction: the thin LLM wrapper (a UI on top of someone else's model) and the aggregator (a router bundling several models). When Google and Accel's accelerator screened roughly 4,000 AI applications, they rejected about 70% as shallow wrappers. The logic: once a wrapper shows traction, the incumbent ERP, CRM, or vertical platform bolts on an AI widget and matches 80% of its value in a month.
The test everyone now applies, explicitly or not, is this: if the frontier labs shipped a new feature tomorrow, would this company vanish? If yes, it is not a Series A company — it is a feature waiting to be absorbed. Every company in our cohort that made it through answers that question with a confident no. Here is why.
Archetype 1: Picks-and-Shovels — AI Infrastructure & Dev Tools
The most consistent winners are the companies selling to the people building AI, not competing with the models themselves. As the volume of AI and agent development explodes, the rails that development runs on become more valuable, not less. In our portfolio this shows up as agent frameworks, background-job and orchestration infrastructure, AI-native QA and testing, and developer tooling — companies like Giga ML (enterprise voice and chat agents built for security and compliance), Trigger.dev (open-source background jobs), Mastra (an agent framework), and Momentic (AI-native testing).
Their moat is structural: a better model is a tailwind, not a threat. More capable AI means more AI gets built, which means more demand for the infrastructure underneath it. These businesses are short volatility on model progress — they win in every scenario where AI keeps advancing.
Archetype 2: AI-Native Verticals in Regulated, High-Friction Domains
The second cluster is software for industries where being right matters, the workflows are gnarly, and the data is proprietary — procurement, legal, compliance, regulated finance, life sciences. Companies like AskLio (procurement), Solve Intelligence (patent drafting and prosecution), Kobalt Labs and Greenboard (compliance for financial institutions), and pharma-focused commercial intelligence tools sit here.
A general-purpose model cannot simply replace these because the moat is not the language capability — it is the proprietary workflow data, the regulatory surface area, and the domain depth accumulated by being native to the industry rather than to the technology. The AI is an ingredient; the defensibility is everything around it. As models improve, these products get better and more trusted, while the compliance and accuracy bar keeps would-be commodity competitors out.
Archetype 3: Deep Tech — Where the Moat Is Physical
The third group barely intersects the wrapper debate at all, because the moat lives in atoms or science rather than software. Quantum computing (Sygaldry), drug-discovery infrastructure (Tamarind Bio), and autonomous surgery (Andromeda Surgical) are companies where AI is a powerful input but cannot be the substitute. You cannot prompt your way to a quantum processor or a validated clinical workflow.
These take longer and cost more, but they are the most insulated of all from model commoditization. When software is being eaten, the companies whose value is anchored in the physical and scientific world — "atoms meets bits" — look increasingly like the safest place to be.
The Common Thread
Three different archetypes, one underlying property: none of them is an AI feature. Each owns something that a more capable model makes more valuable rather than less — the infrastructure AI runs on, the proprietary data and regulated workflows AI plugs into, or the physical and scientific reality AI accelerates but cannot replace. That is exactly the inverse of the thin wrapper, whose entire value is the model it sits on top of.
This is also why we run every prospective investment through a single screen we call the AGI-resiliency test: when far more capable AI arrives, does this company's value increase or evaporate? The companies in our portfolio reaching Series A are, almost by definition, the ones that pass it. 'AI is eating software' is true — but it is eating a specific kind of software: the generic, ungrounded, easily-absorbed kind. The companies clearing the Series A bar are the ones building on the parts of the stack that AI feeds rather than consumes. In the noise, that is the signal.
