Why the AI premium poisons non-AI comp sets.

In 2026 AI-labelled startups raise at roughly double the valuation of non-AI peers at the same series, and the gap is published rather than inferred. The PitchBook-NVCA Venture Monitor Q2 2026 printed median pre-money valuations for both cohorts as of 30 June 2026: Series A $83.0M for AI companies against $44.0M for everyone else, Series B $267.0M against $134.0M, Series C $648.0M against $397.6M. For a founder building a non-AI company the consequence is operational, not academic: every unstratified “2026 median” you will be quoted sits well above the number your own cohort actually transacts at.

The premium by stage

All four pairs are printed figures from the PitchBook-NVCA Venture Monitor Q2 2026 (“YTD median AI and non-AI VC pre-money valuation by series”, as of 30 June 2026). No edition prints a split below Series A, so pre-seed and seed show none. The Series C+ row uses the Series C pair; D+ is listed separately because it behaves nothing like it. The Q3 2026 edition draws the split for the third quarter alone without printing the values, so the 30 June pairs remain the latest published. What Q3 does print is the AI-only median through 30 September: seed $21.0M, Series A $78.0M, Series B $266.3M, Series C $694.5M.

What this looks like in dollars

At 30 June the blended Series A pre-money median for 2026 was $64.0M, and it was a number almost nobody actually raised at: the AI median was $83.0M and the non-AI median $44.0M. A non-AI founder who benchmarks to the blend is asking for 45% more than their own cohort's median, which is a gap no traction story closes in a first meeting. The arithmetic runs the other way too. An AI-native founder who benchmarks to the blend is leaving $19M of pre-money on the table before the conversation starts.

This is also why a comp set pulled straight out of Crunchbase or PitchBook without a cohort filter is close to useless. Ten Series A comps drawn at random in 2026 will contain both populations, and because the two medians sit almost 2x apart the naive median across all ten lands wherever the AI/non-AI ratio in your particular sample happened to fall. You are not measuring the market, you are measuring your sampling accident. At Series B the two medians are $267.0M and $134.0M, so the same problem is worse a round later.

How to filter AI from comp sets

1. Read each comp's positioning. If the company describes itself as “AI-native”, “AI-powered”, “LLM-first”, or includes generative-AI as core product, it's an AI deal even if the underlying SaaS category isn't.
2. Check the round announcement language. AI deals consistently lead with model-capability claims, evaluation benchmarks, or training infrastructure. Non-AI deals lead with ARR, growth rate, or customer metrics.
3. For ambiguous deals (a SaaS company that added an AI feature), check whether the marketing positioning treats AI as the core product or a feature. Core-AI = AI premium; AI feature = non-AI premium.
4. If your comp set ends up with 6 AI and 4 non-AI deals, separate them. Report “AI median: $X. Non-AI median: $Y.” Then state which group you're benchmarking against.

If you ARE AI-native

The $83.0M Series A figure is itself a median across a very wide cohort. Carta now splits that cohort: its State of Private Markets Q1 2026separates applied-AI companies from foundational-model companies, and says a foundational-model startup at Series A “might be raising at a $300 million median valuation” while a non-AI startup at the same stage is at $55 million. Carta's figures come from companies on its own platform, so they will not match the Venture Monitor's, but the shape is the point: a handful of model builders sit far above everyone else, and an application-layer AI company is not in their market. The same thing happens later on: at 30 June the Series D+ AI median pre-money was $4,250.0M against $644.0M non-AI, which tells you about the largest model labs and almost nothing about a normal company. Pull comps from your own AI tier and your own round size, and treat any aggregate AI median above Series B as decoration.

Practical recommendation:When VCs ask for your benchmark, lead with the split rather than the median. “Our eight Series A comps split three AI and five non-AI. We are not an AI company, so we are benchmarking to the non-AI subset. The last published non-AI Series A median, at 30 June 2026, was $44.0M and our ask sits against that, not against the $64.0M blend.” Naming the blend and rejecting it yourself pre-empts the objection the VC was going to raise, and it signals you read the underlying data rather than a headline.