š Hi, Iām Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI.
Upcoming events:
DDVC Investor Summit at Bits & Pretzels 2026 Sep 28-30
DDVC x Bits & Pretzels Breakfast 30th Sept in Munich
Brought to you by Exa ā The Search Engine Built for AI
ExaĀ turns the web into live, structured data that your agents and pipelines can act on.
AI-forward firmsĀ use Exa to source companies by thesis, enrich target businesses programmatically, and monitor signals across the market in real time.
Sign up with code EXA50DATADRIVEN for $50 in free credits and bring real-time web data to your AI agents via Claude, Codex, Cursor & more!
Welcome to another Data Driven VC āInsightsā episode where we cover the most interesting research and reports about startups, GPs, LPs, AI & automation.
Extremes of Power Law in VC
Peter Walker at Carta shared new data on 400+ US venture funds from the 2016, 2017, and 2018 vintages, illustrating how extreme the power law in VC really is.
Bottom Half of Funds Return Almost Nothing: The 25th percentile fund sits at 1.06x net TVPI and the 50th percentile at 1.38x, both described in the data as close to a total miss on invested capital.
Only the Top Decile Clears "Good": The 75th percentile reaches 2.09x, the 90th percentile 3.37x, and the 95th percentile 4.72x, a sharp jump once you cross into the top quartile.
The 99th Percentile Is a Different Business Entirely: Top 1% funds post 13.12x net TVPI, an order of magnitude above the median.
āļø KEY TAKEAWAYS
LPs allocating to a median-quartile manager should expect close to breakeven, based on this dataset. Manager selection into the top decile is what actually drives venture returns. That distribution is unlikely to shift meaningfully as these fund vintages mature further.
What Got Hyped vs What Got Built
Sequoia's Konstantine Buhler ranked the most hyped topics on Hacker News for every year since 2007 and mapped each year against the most valuable company actually founded that year.
The Top Company of a Given Year Is Rarely the Hyped Topic of That Year: Airbnb (2008) launched while Google dominated the conversation, Uber (2009) arrived amid low-level programming chatter, and Anthropic (2021) was founded while crypto consumed the internet.
New Trends Announce Themselves 5-6 Years Early: LLMs first cracked the top 15 in 2016 and didn't hit #1 until 2022; AI coding entered at #12 in 2021 and topped the list by 2026; crypto entered in 2011, well before its 2017 and 2021 peaks.
Sustained Attention Is a Durable Signal: The "Musk-verse" has held a top-15 spot for 14 consecutive years, the only topic with that kind of staying power in the dataset.
āļø KEY TAKEAWAYS
The practical signal for sourcing is to watch which subcommunities discuss a topic 5-6 years before it becomes consensus. Recent years' "hyped" topics haven't produced a clear winner yet, per Sequoia's own data, so early attention alone isn't enough to pick the outcome.
Join 1914+ investors in our free Slack group as we automate our VC job end-to-end with AI. Live experiment. Full transparency.
How VCs Decide: Jockey or Horse?
Stanford GSB's Ilya Strebulaev revisits the "jockey or horse" debate using survey data from over 1,000 VCs plus a study tracking 50 VC-backed companies from business plan to IPO.
The Case for the Jockey (Teams): 47% of surveyed VCs rank team as the single most important factor, though 95% call it important or very important when multiple answers are allowed.
The Team Giveth and Taketh Away: 55% cite team quality as the top driver of both investment success and failure.
The Evidence Favors the Horse (Business), at Scale: Across 50 companies tracked from business plan to IPO, fewer than three-quarters of CEOs at IPO had held the role at initial investment, while each company's core business line stayed stable.
āļø KEY TAKEAWAYS
Founders raising earlier pushes team weight up, likely explaining recent team-first wins. It also raises the odds of a mismatch between the founding team and what the company later needs, showing up as turnover or contract terms built for that risk. Think now about what scenario you can foresee, and whether your fund is prepared for it.
AI Harness' ARR Multiples
Tomasz Tunguz asks how the market values an "AI harness", using Harvey, Legora, Sierra, Ramp, and Decagon as the data points.
Three Companies Crossed $100M ARR Within Nine Months of Each Other: Harvey, Legora, and Sierra each announced crossing $100M in annual recurring revenue within nine months of one another, with Ramp (~$1.4B ARR) and Decagon (~$35M ARR) bracketing the group at the high and low end.
The 100x ARR Multiple Is Back: The article notes a 2021 piece on the "100x ARR multiple" as a historical premium; five years later, multiples are back at similar levels, but reported growth rates underneath are running roughly 3x faster.
Multiples Accelerated Again in January 2026: Multiples typically compress as a company scales, but Legora, Sierra, and Ramp saw theirs accelerate starting January 2026, driven partly by sustained growth and a more favorable fundraising market.
āļø KEY TAKEAWAYS
These multiples require growth roughly 3x faster than a normal SaaS company, which is what these five names are actually posting right now. If a portfolio company can't show that same growth premium, a 25-125x ARR multiple doesn't apply to it, no matter how similar the product category looks.
Upgrade your subscription to access our premium content & join the Data Driven VC community
The Work-From-Anywhere VC Stack
Yohei Nakajima posted his personal AI operating setup, a hardware-plus-agent stack that lets him work on almost anything from anywhere, hands-free.
A "Chief of Staff" Control Plane, Reached Hands-Free: Voice or text tells the Chief of Staff what's needed; it assigns the task to the relevant project manager, who assigns it down to a builder/doer.
Projects Are Just Repos, and They Nest: Each project (work, research, or personal) lives in its own GitHub repo holding its instructions and knowledge, and any project can itself manage other projects by spinning up a new repo underneath it.
Roles Are Skills Any Model Can Run: Chief of staff, manager, and builder are role definitions that Claude, ChatGPT, Codex, or Claude Code can each execute, which is what lets a task started on one AI resume on another without losing context.
āļø KEY TAKEAWAYS
This is what a sophisticated modern VC OS increasingly looks like: one person pairing general-purpose AI models with GitHub, replacing what used to require a stack of dedicated SaaS tools.
Can Agents Use a Computer Yet?
Fabrizio Serafini, Seema Amble, and Eric Zhou at a16z surveyed a year of computer-use deployments and found the frontier has shifted from clicking the right button to reliably doing the job.
OSWorld-Verified Jumped From 42% to 85% in a Year: The best model scored 42% a year ago; today's leader, Claude Fable 5, scores 85%, above the ~72% human baseline.
Agents Undercut Offshore BPO on Cost: Inference runs $6-8/hour typical ($3-15 range) versus ~$10/hour for offshore BPO and $30-45/hour for US back-office labor, pencilling out to 70-80% gross margin against US labor even at the high end.
The Moat Moved From Model to Context: As raw computer-use capability becomes commoditized, buyers pay for the specific, hard-to-replicate knowledge of how one company's workflow actually runs.
āļø KEY TAKEAWAYS
For a VC evaluating a computer-use agent startup, the thing to check is how much company-specific operational knowledge the product has already captured, since that's what determines defensibility once the underlying model becomes a commodity.
Thatās it for today!
Stay driven,
Andre
PS: Join the DDVC Investor Summit at Bits & Pretzels virtually or physically in Munich during Oktoberfest Sep 28-30
