👋 Hi, I’m Andre and welcome to my newsletter Data Driven VC which is all about becoming a better investor with data and AI.
Welcome to another Data Driven VC “Insights” episode where we cover the most interesting research and reports about startups, GPs, LPs, AI & automation.
Valuations Need a Correction
Nnamdi Iregbulem at Lightspeed builds a selection-adjusted index of Seed valuations, arguing that headline prices only capture the companies strong enough to raise.
Only Winners Get Priced: A private valuation is a financing price set at the moment a company raises. Since stronger companies are more likely to raise, observed deal prices sit above what the typical startup is worth.
The Series A Bar Keeps Rising: In the early 2010s, the median company raising a Series A came from the top 10-15% of active Seed startups by valuation. Today it comes from the top 3%.
New Rounds Make Poor Comps: New Seed rounds are priced at about 4x the average and 10x the median existing Seed startup. Until recently, Iregbulem estimates the median Seed startup was worth no more than about $1M, far below the headline deals.
✈️ KEY TAKEAWAYS
Rising Seed prices can hide a weaker market, since only the strongest companies raise. VCs pricing new deals or valuing older holdings should look past recent rounds: only about 4% of Seed startups raise a Series A each quarter, so many holdings go months or years without a fresh price.
Building a VC Firm Like an Intelligence Agency
Lauren DeMeuse at Theory Ventures explains how the firm borrowed Palantir's forward-deployment model for portfolio support, sending operators to work inside companies as an extra pair of hands.
Advice Stops Short of Execution: DeMeuse, who worked on Palantir's forward-deployed team in 2012, argues that capital and advice get a startup most of the way, while the hardest work, from first sales hires to executive searches, happens inside the company.
Operators Build Sales Teams and Hiring Plans: One operator designs how the company sells, from which accounts to target to how deals get closed, while another plans each hire, including recruiters, agency contracts and pay. Both work on every founding salesperson and senior sales hire.
Ollama Builds an Enterprise Sales Team: At Ollama, a platform for running open-weight AI models, Theory worked inside one day a week, setting up sales tools, co-designing enterprise trials and scoping hires. The CEO says the hiring and sales plan took under 90 days.
✈️ KEY TAKEAWAYS
VCs copying this model need (1) a concentrated portfolio, which DeMeuse says makes it possible, (2) operators who have done the job themselves and (3) internal tools that adapt to each company. Sales and hiring are a natural place to start, since that is where the piece places the hardest work.
Upgrade your subscription to access our premium content & join the Data Driven VC community
Long-Running Agents Win the Relay Race
David Haber at a16z argues that agents able to work for weeks can take over high-value customer journeys such as legal cases, insurance claims and mortgages, from first contact to final result.
Agents Can Now Own the Whole Journey: Many high-value processes run for weeks and risk losing value at every handoff. Long-running agents can now carry one from start to finish, keeping context and following up as often as needed, as Haber's plaintiff-law case study shows.
Outcome Pricing Grows the Market: When an agent drives the whole journey, its results are measurable and easier to attribute, so vendors can charge for outcomes. Haber expects this to expand AI's market from a share of labor budgets to a share of revenue.
Three Ways to Go to Market: Vendors can (1) sell the agent as software to existing teams, (2) take over a service companies already outsource, or (3) go full-stack and run the business themselves, each with a different trade-off between sales speed, pricing power and operating complexity.
✈️ KEY TAKEAWAYS
Investors deploying their own agents should give them whole journeys, such as a deal from first call to close, and judge them by outcomes. The leaks Haber describes show up in venture too: slow replies lose deals, follow-ups stop too early and data-room details go unread.
Startups Are “Spikes” or “Explainers”
Aaron Harris, founder of Magid and former YC partner, sorts startups into two types based on how quickly investors understand why they matter.
Spikes Sell on One Signal: A Spike draws investor excitement from one narrow trait, usually a team that built a unicorn or came from a frontier AI lab, or traction that stands out while the bar is very high.
Explainers Do the Teaching: Explainers lack an instant signal and must walk investors through what they do. Harris finds they are often the most valuable long-term, with lower pricing and fewer competitors.
New Investors Reveal Your Type: Existing backers are biased, so founders learn their type from how net new investors react. If investors chase you, you are a Spike; if you chase them, you are an Explainer.
✈️ KEY TAKEAWAYS
Spikes should raise fast, since Harris finds investor attention fades on an unpredictable timeline. Explainers should build investor relationships well before a raise, and investors willing to do that work may find them at lower prices.
Join 2,015+ investors in our free Slack group as we automate our VC job end-to-end with AI. Live experiment. Full transparency.
Agents Shift AI From Inference to Execution
Jaya Gupta and Ashu Garg at Foundation Capital argue that AI's next phase is about turning intelligence into completed work, with power moving to the companies that build agents and decide what to buy.
Most Agent Work Is Execution: Inference (the model thinking, on GPUs) is a small part of agent work. Most is execution on CPUs: scheduled checks, storage and retries that keep running after the chat ends. Intel sees the CPU-to-GPU ratio moving from about 1:8 in training to 1:1 for agents.
Assemblers Control the Mix: Companies assembling agents route each step across hundreds of models, caches and runtimes, so suppliers compete per request. Browsers, devices, clouds and AI labs are fighting to become that coordinator and make the rest their suppliers.
GPU Financing Rests on a Shifting Mix: Neoclouds like CoreWeave financed GPU capacity against expected demand, but assemblers can cut inference with smaller models, caching or saved code. CoreWeave, whose interest costs ate about a quarter of Q2 revenue, warns customers may need fewer of its services.
✈️ KEY TAKEAWAYS
Gupta and Garg compare today's GPU buildout to Nortel and Lucent: internet traffic kept rising while carriers cut equipment orders. For investors, their bet is that AI can boom while some of the investments funding it lose, as value shifts to whoever decides how the work gets done.
VC Value-Add Needs Proof
Laurie Owen at Refinery Media argues that most firms market value-add they cannot prove, and shows how the strongest firms back their claims with case studies.
Value-Add Is a Marketing Problem: Every firm claims value-add and almost none prove it, so founders tune it out. Owen points to Y Combinator as the payoff: its value is demonstrated so publicly that top founders accept its unusually expensive terms.
Narrow, Measured Claims Earn Trust: Craft Ventures placed an operator-in-residence inside Replit as interim marketing lead, then claimed only a nearly 7x rise in Replit's online software reviews, even as Replit's ARR passed $150M. Owen argues that narrow claims like this make the rest believable.
Let Founders and Systems Tell It: The strongest proof comes from founders, like Rippling telling how Greenoaks signed a $500M term sheet within 12 hours during the SVB crisis, or from systems like 776's public scoreboard of intros, response times and hires.
✈️ KEY TAKEAWAYS
Firms that prove one thing clearly a few times stop needing to say "value-add", because founders say it for them. Owen finds no firm publishes the full story of a hire: who, which role, how fast and what changed after. Should VCs start sharing theirs?
That’s it for today!
Stay driven,
Andre
