👋 Hi, I’m Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI.
Join our free virtual roundtable “The Compounding Data Layer: Building an Edge When Everyone Uses the Same AI Model” with Eight Roads, Offline Ventures, and Kruncher in less than 24 hours here
Brought to you by Exa — The Search Engine Built for AI
Exa turns the web into live, structured data that your agents and pipelines can act on.
AI-forward firms 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 “Insights” episode where we cover the most interesting research and reports about startups, VCs, LPs, AI & automation.
Power Law Is Everywhere
Peter Walker from Carta shows that market concentration is far more extreme outside the US than most investors assume.
Concentration doubled: The top 10 stocks in the S&P 500 made up roughly 17% of total US market cap in 2015. Today that figure is 36%.
US ranks near the bottom globally: Despite the doubling, the US is one of the least concentrated major markets by top 10 share. Japan sits lowest at 29%.
75%+ is common elsewhere: Countries including Spain, Poland, Israel, Denmark, Singapore, Norway, and several others see their top 10 stocks take 75% or more of total index weight, with a handful at 100%.
✈️ KEY TAKEAWAYS
The "just be in the right 5 companies" logic that mega-fund VCs use is really a description of how public markets already behave everywhere else. The US is the outlier for having less concentration, not more. For LPs benchmarking VC power-law returns, this is a useful reminder that concentration risk is structural to markets in general, not unique to venture.
The Sick Man of Private Markets
The Odin Times argues venture megafunds are overdue for the same fee and structure reset that private equity went through after 2008.
1.5% to 1.75% fees: After the Global Financial Crisis, larger PE funds moved off linear fee scaling, stepping down to 1.5 to 1.75% and shifting the fee base from committed to invested capital.
10+ co-investments outperform: Research cited shows PE LPs building portfolios of at least 10 co-investments net-outperform putting the same capital straight into the fund. The VC equivalent takes about 20 deals.
+0.42x MOIC, +2.6% IRR: Applying PE-style fee step-downs and stronger co-investment rights to VC megafunds is estimated to lift net MOIC by 0.42x and net IRR by 2.6 percentage points.
The piece's core claim is that VC megafund underperformance versus PE is an incentive problem, not a power-law problem. LPs already have a working template from PE's post-GFC reforms. The obstacle is coordination: no single LP wants to be first to push back on a top-tier fund's terms.
a16z's New Media: What’s Left for Everyone Else?
Refinery Media's Laurie Owen tracks how a16z's media operation has scaled over the past year, and where it is pulling back.
Fellowships double as deal sourcing: a16z isn't really hiring for its New Media, FDE, design engineer, and growth fellowships. It's building early relationships with people the piece expects to be "founders pitching them in 3 years," years before they'd enter a normal firm's pipeline.
Media as a closing tool: The "Bear Hug" pitch reframes the fundraise meeting around 20 concrete things a16z will do in week one, with New Media as a headline promise.
Four gaps that survive at scale: single-vertical depth, unscalable one-to-one work, genuine personality, and independent voice not read as the firm talking its own book. All get harder for a16z to hold as it grows, and are where smaller funds can still compete.
✈️ KEY TAKEAWAYS
a16z's scale advantage in distribution is real and getting harder to match. But the piece argues personality, unscalable one-to-one relationship work, and single-vertical depth are the gaps that get worse, not better, as a media operation industrializes. That is where smaller, focused funds can still compete.
Join 1,808+ investors in our free Slack group as we automate our VC job end-to-end with AI. Live experiment. Full transparency.
Private Equity Underestimates AI
Varick Agents breaks down why most PE firms deploying AI are only capturing part of the value creation equation.
100+ deployments from 4 portcos: Workflow fragmentation across regions, teams, and systems means a portfolio with just 4 companies can require more than 100 separate AI deployments.
From 100+ down to 3 agents: A two-step consolidation approach, standardizing workflows within each company, then grouping companies with shared logic, cut one example portfolio's accounts payable transformations from over 100 to 3 reusable agents.
Three IRR levers, not one: Beyond EBITDA expansion, standardized operations support higher exit multiples with growth-oriented buyers and shorter holding periods, since transformation patterns can be redeployed faster across the portfolio.
✈️ KEY TAKEAWAYS
Most PE-AI deployments today only touch the EBITDA margin lever, ignoring exit multiple and holding period entirely. The argument is that consolidating workflows before deploying AI, not automation alone, is what unlocks all three levers and drives IRR within PE's fixed 5 to 7 year window.
The 4 Layers of AI Engineering
Alex Prompter lays out a framework for how AI usage matures from typing prompts to running unsupervised loops.
Layers 1 and 2, prompt and context engineering: Layer 1 is wording and instructions in the chat window, where most people start and stop. Layer 2 adds system instructions, files, and history, since a mediocre prompt with great context is said to beat a great prompt with none.
Layer 3, harness engineering: The code around the model, covering tool routing, verification steps, retry logic, and structured outputs. This is framed as what makes AI reliable instead of just impressive.
Layer 4, loop engineering: The system runs itself against a goal and stop condition, prompting, checking, and adjusting without manual triggering, removing the human as the bottleneck entirely.
✈️ KEY TAKEAWAYS
This is a useful mental model for any team evaluating AI tools on prompt quality alone. The compounding gains sit in context, harness, and loop layers, which is exactly where operational and diligence workflows built on AI-native tooling should be investing next.
A Field Guide to Fable 5: Finding Your Unknowns
Anthropic's Thariq Shihipar argues that with more capable models, output quality is now bottlenecked by the operator's ability to name their own unknowns.
4 types of unknowns: Known knowns, known unknowns, unknown knowns, and unknown unknowns. Reducing them is framed as the core skill of agentic coding.
Blindspot pass and interviews: Concrete techniques include asking Claude to run a "blindspot pass" to surface unknown unknowns, and to interview you one question at a time on unresolved ambiguity.
The launch video proof point: Fable's own launch video was edited entirely by Claude Code, combining transcription, ffmpeg cuts, Remotion prototypes, and Claude-taught color grading.
✈️ KEY TAKEAWAYS
As models get more capable, the constraint shifts from model quality to how well an operator can define and close their own unknowns. A repeatable toolkit for surfacing those unknowns is becoming the actual differentiator in agentic workflows.
That’s it for today!
Stay driven,
Andre
PS: We’ve been the first-ever customer of Exa and their structured web access is a true super power! Test it yourself with $50 credits here
