Most Venture Funds Are Playing Momentum Games | Michael Dempsey

Summary

• Dempsey’s core warning: a lot of venture funds are playing momentum games, and momentum is uniquely lethal in venture because there is no scaled bid. In public markets a fading momentum trade comes down 10–20% into a live bid; in illiquid, dilution-heavy private markets, “when the bid starts to soften, it doesn’t soften 10 or 20%, it softens 100%.” Investing “looks easy in certain 36-month windows and usually you find out it’s hard in 48-plus months.”
• He rejects the “pay up for legible founders” seed strategy on data: everything but the bottom quartile of seed valuations produces the same distribution of top-decile outcomes. Loss ratios on high-priced seeds aren’t meaningfully better than cheap ones, and Compound’s two largest companies — Runway and Wave — had founders who “even today would not be considered super legible.” With AI lowering the barrier to creation, he finds it strange to expect talent to narrow rather than disperse.
• Maximal obviousness makes a sector uninvestable for Compound: venture is “the lowest conviction asset class in finance,” so consensus themes see model collapse and identical companies. José suggests AI is now roughly twice as obvious as its “maximally obvious” 2024 reading; Dempsey would rather wait for the dust-settling window he pegs at 2028–2030, when cheap intelligence commoditizes much of the market while frontier intelligence creates new opportunities.
• Compound’s edge is decaying by design and must be rebuilt: the 2016–2022 edge of simply understanding academic research is “largely gone” now that anyone can put a paper into ChatGPT. The replacement is prescriptive first-, second-, and third-order-effects thinking — something AI still can’t do because “nothing shows me that AI is able to think out of distribution” — plus understanding how narratives drive multiples and late-stage flows, which models are “horrible at.”
• The public-markets fund targets venture-like IRRs through dislocations the market misreads: software facing AI-driven margin compression can re-expand as inference costs fall — Pinterest’s fine-tuned open-source models reportedly run at 8% of frontier-API cost. Bio longs are undervalued data assets and left-for-dead small caps; shorts are narrative-driven deep tech where the first sign of invalidation can send stocks down “40 to 60% in 4 to 6 weeks.”
• Mag 7 views: Meta is “probably best positioned to monetize AI regardless of whether they build on the frontier”; Google’s structural advantages — including GCP, models, and unusually cheap debt — make him bullish; Apple is well positioned. He is less high-conviction on Nvidia and memory stocks. His unchanged 2024 “AI obvious basket” of about 30 names (ASML, Micron, hyperscalers, Alibaba) “has just so drastically outperformed,” and he thinks theme baskets à la Citrini may be the future of active management.
• Robotics was the big mistimed call — 2016–18 bets on generalizable deep-learning policies were “8 to 10 years too early” — but “this time is really different,” and he’s backing a stealth vertically integrated frontier lab plus vertical players like Alquist. On the $0-to-$10 trillion humanoid claim that José says he has heard Andrew Kang make, Dempsey urges a more measured view: humanoids make most sense where the relevant cost is human life — military or industrial — and industrial robotics is already the most penetrated area.
• Macro stance is “rolling bubbles,” not a single top: CoreWeave went to 79 and back over 100, while Palo Alto fell more than 40% and recovered toward highs — whiplash is structural because everyone’s stop is the same 20% drawdown. On selling discipline, Compound’s data says exiting hyper-reflexive names at all-time-high NTM revenue multiples beats selling on a 25% retrace — though “you might miss the leg 3 years later.”

Deep dive

1. Momentum dies at 100% drawdown in venture

• Dempsey’s definition: momentum is “doing something that you are doing because you know that there are marginal buyers at incremental price increases in the near term, not because you believe that long-term it’s durable.” All risk asset classes have been poor relative to simply buying the Nasdaq or S&P over the past decade, so you must be top decile — and “momentum is a hard place to be top decile in.”
• The structural difference from public markets, where Soros-style reflexivity trades work: public momentum fades into a bid that comes down 10–20% at a time. In venture there’s no scaled liquidity, plus constant dilution and preferred stacks — “when the bid starts to soften, it doesn’t soften 10 or 20%, it softens 100%.”
• He carves out one momentum-adjacent strategy as genuinely good: the “Thrive strategy” of buying the best assets and believing their terminal value exceeds market belief. It’s access-constrained, economies-of-scale, big-check investing — “actually might be one of the most durable strategies in all of venture from the past 5 to 10 years.”

2. The legibility fallacy: expensive seeds aren’t better seeds

• Against the pay-anything-for-hot-names seed game: “it’s a fallacy to believe that the highest price things are the best things at the early stage… that’s just quantitatively not true.” The data he cites: top-decile exits are evenly split across seed valuations — “everything but the bottom quartile of valuations creates the same distribution of outcomes” — and loss ratios of high-priced seeds don’t meaningfully beat low-priced ones.
• On founder legibility: “I just don’t fundamentally care if you’re like an amazing 20-year-old person who went to Stanford.” Compound’s two largest companies — the Runway team and the Wave founder — “even today would not be considered super legible,” and the same holds for their newer winners. Second-time versus first-time founder success data shows “pretty much no correlation.”
• His sharpest structural argument: everyone simultaneously claims AI is lowering the barrier to creation and raising minimum viable intelligence — so “it feels very strange to me that the talent will actually narrow in where it comes from. I don’t know any other paradigm in which barrier to entry has come down and talent has not been higher dispersion.”

3. Talent pools, king-making, and why Hummingbird’s game isn’t his

• José pushes back that legibility now means the math-olympiad archetype — the Citadel intern class producing the Hyperliquid founder, Alexander Wang, and Scott Wu. Dempsey concedes pockets of talent create floors for windows of time: mobile engineers acquired at about $10 million each during the Facebook era, self-driving teams from five places or the DARPA Grand Challenge taken out at $5–15 million per member, and now AI and the Neo labs — but “whether or not the investors get hosed on that is still to be determined.”
• The question that actually matters in the legibility debate, per Dempsey: “do you believe that king-making is actually possible within venture capital or not?” — because if legibility grants excess capital advantages, his skepticism is moot. “Elon is the best example of that.”
• On Hummingbird: it’s a real, different strategy — not “this cohort is interesting” but a shared way of sourcing people for asymmetry, historically finding illegible and very low-priced people, though they are no longer as cheap now that others orient toward them. “That’s a strategy… it’s not how I will make money,” and the legible signal “is almost like a lagging indicator.”

4. Compound’s edge: predicting the future, humans included

• The firm’s self-description: research-centric, thesis-driven investing — get “closest to the metal” on academic and R&D-group research, then form prescriptive views on first-, second-, and third-order effects. For 10 years everyone in venture told him “you cannot predict the future… let the founder show you the future”; his counter is that reasoned views matter more as volatility and company replacement rates rise.
• José connects this to Dempsey’s “On Humanity and Human Beings” essay and its critique of world-first thinking — technology building a beautiful model of the future and slotting humans in last. Dempsey agrees that’s where Compound diverges from Silicon Valley: closer to the Tyler Cowen view that “AI is not going to diffuse nearly as aggressively as all these 30%-plus GDP people think because humans just don’t want it to,” versus San Francisco’s “it doesn’t matter what humans want, the AIs do it all.”
• The same human-first lens applies to public markets: “a lot of people have a very simplistic view of like, oh that thing’s dead, software’s dead — and it’s like, well, that’s not really how these things work with humans making decisions.”

5. Maximally obvious means uninvestable — for Compound

• In 2024 he called AI “maximally obvious”; José says it is perhaps twice as obvious now. The mechanism: venture is “the lowest conviction asset class in finance,” so when things become obvious, prices scale exponentially, everyone builds identical companies, and picking the best of five clones at seed “is not where we have edge or alpha… I’m not going to try to pretend it is.”
• The second force is “model collapse” in tech itself — talent flowing into a singular dogma. “There are actually very few ideas that I’ve heard in the past 3 years where I’m like, wow, I never could have thought of that… AI for legal is like the idea people have been talking about for 30 years.”
• The discipline underneath: “so much of venture is just understanding what is your singular advantage and pushing on that advantage over and over again,” not playing the game LPs want you to play. The house rule mirrors his founder advice: “make the decisions that if your company fails, you can sleep at night having made.”

6. All firms decay; the ChatGPT-proof edge is nth-order thinking

• His default view: “all venture firms are constantly decaying,” so you fight the decay. Compound 1.0 (2016–2022) was built on understanding hard-tech academic research others wouldn’t underwrite at seed. “Now any person in the world can take an academic paper, put it in ChatGPT, and say explain this to me like I’m 16… that edge is largely gone.”
• The replacement: not “generative AI will be a thing because we read the early GAN papers in 2016” — a thesis detailed for its time, thin by today’s standard — but full cascades: early customers, competition, adjacent businesses, whether a company is even the best expression of the future you believe in.
• Why AI can’t yet do this itself, despite Compound actively trying, including embedding-space experiments to flatten idea distributions: “I have seen nothing that shows me that AI is able to think out of distribution” — by definition a model hands you the maximally obvious or viable ideas. Models are also “horrible at” how narratives impact public multiples and late-stage private flows — which is partly why Compound publishes its research.

7. No universal “great founder” — and the nerds are getting commoditized

• José relays Grad Capital’s Abhishek’s claim that there’s “one year left of investing in nerds” as AI commoditizes hyper-technical skill. Dempsey’s partial agreement: “it’s never been that the smartest founders are the best ones,” minimum viable intelligence is rising, but “I’m so pro-human that I think humans are so unmodelable… we have a while to run before we commoditize humanity through AI models.”
• His contrarian founder framework — in stated conflict with some of the best investors: “we don’t believe there’s just these things called great founders. We think there are great founders for certain types of businesses.” Supply-constrained, BD-heavy spaces need “more of a savage” — José’s example is neocloud teams securing gigawatts of powered land rather than merely optimizing kernels; heads-down businesses need something scarcer — “maybe just the ability to not get distracted doing random things on the side is the main thing people should be looking at.”

8. Portfolio construction as pseudo-religion: 1-in-12, 1-in-20

• José says Dempsey once told him that a fund needs roughly 30 seed bets to catch one or two good ones. Dempsey’s own rates are that “one out of every 12 times we make a pretty good decision, and one out of every 20 times we make an incredible decision.” He told LPs that “our portfolio should feel riskier and riskier as time goes on,” because LPs are not always leading indicators of where the world is going. Dogmatic 10-company concentration means “you’re going to be unemployed in 5 years.”
• The psychological architecture is the point: venture is all inputs with no output data for years, so year-five investors who never had conviction in their inputs drown in volatility. “You can’t come to a high-risk game played at very high stakes with tons of fear and uncertainty internally.” The fix: reduce noise and run an input system you deeply believe — “this is pseudo-religion” with some bounds of understanding.
• Against José’s rule-breaking instincts, the goal defines the heuristics: some LPs would love 10x fund one then 2x–2x; Dempsey wants “consistent top decile funds with the ability to have top-5% or 1% funds.” Related selling data: on hyper-reflexive names, exiting at all-time-high next-12-month revenue multiples beats selling on a 25% retrace — “but you might miss the leg 3 years later. I don’t know.”

9. Where the money’s going: bio, materials, and AI in 2028–2030

• Fund two, which he described as deploying from 2021 through the end of the period discussed, is mostly bio, some AI, and materials science — the last being what he’s most excited about: full-stack businesses synthesizing novel materials and shipping end products. Case study Orbital Materials: a chemistry foundation model, state-of-the-art in chemical synthesis, that shipped its material as a modular data-center cooling product for next-generation GPUs. Energy is the adjacent, admittedly consensus area.
• The AI timing call: “a really interesting time to invest in AI in 2028 to 2030 once we see dust-settling dynamics” — maximally viable, incredibly cheap intelligence and the commoditization of everything, while the furthest-frontier intelligence should create new companies.
• On commoditization mechanics: a subset of tasks becomes effectively free, possibly local, in 24–36 months; frontier intelligence stays economically premium for perhaps 5–10 years among a few compute-rich players — but “you only need one player who is not economically motivated by the core monetization of the core thing to really drive the price down.” Google’s structural position, including debt priced as less risky than U.S. Treasuries, is “one of the most interesting things in the world to me.” Hating Google right now is “a pretty stupid thing to do.”

10. Why a public-markets fund isn’t scope creep — and what’s long

• He accepts the charge head-on: “this is the fairest criticism anyone can have of Compound, outside of the fact that we might just be a really overpaid research organization that never monetizes.” The defense: in tech the most innovative companies are the largest ones, so understanding the entire flow and stage of technology companies — for example, what Tempus does in diagnostics — directly sharpens seed underwriting. “The only performance-oriented asset classes in technology investing are early-stage venture and public markets at this point.” The vehicle is closed-end, with money locked up for many years; longs are concentrated, shorts smaller and broader, and the fund targets venture-like IRRs — mostly not through standard Magnificent Seven holdings.
• The software thesis: usage-based transitions compress margins short-term — passing through inference can drop 80–90% gross margins toward 50–60% blended — which quarterly-focused markets hate but which can signal inference adoption inside the customer base. The Pinterest datapoint: the CEO said their fine-tuned open-source and first-party models run at 8% of the cost of frontier APIs — “pretty incredible” — setting up margin re-expansion as AI costs fall. Management that says AI is “a sham” would likely be unattractive to them.
• Bio longs: companies “sitting on data assets that are meaningfully undervalued” that will monetize far sooner and at a far larger scale than analysts appreciate, plus an uncovered long tail left for dead; themes include reaccelerating experimentation and biomanufacturing. Explicitly excluded: single-asset clinical-trial trading — “not something we have any edge in.”

11. Shorting the narrative machine in a market of rolling bubbles

• The short book examines the “next decade is about market-cap destruction” line José attributed to Dempsey: narrative-rich deep tech, historically SPACs and now broader, companies Compound passed on privately years ago — “we had a view that they weren’t good investments then and we have a view they’re not good investments now” — and companies they do not think will exist within two to four years. Rules: always a clear catalyst, express the short multiple ways, never stack correlated factors; risk controls are tight and individual shorts are small, because “we don’t try to be heroes shorting things over and over again.” The payoff asymmetry: the first sign of invalidation can send these names down “40 to 60% in 4 to 6 weeks.”
• Against Burry, Dalio, and Grantham bearishness, his frame is “perpetual rolling bubbles”: nobody has a real picture of 2028 earnings per share, so confidence collapses and rebuilds in whiplash. The neocloud round-trip — Leopold Aschenbrenner’s Situational Awareness, Value Aligned, and other funds heavily exposed to the same factor; CoreWeave to 79 and back over 100; Palo Alto down more than 40% on a random “Mythos” announcement and back toward highs — illustrates it. The reflexive kicker: “if everyone has a view that you sell when it draws down 20%, it’s not going to draw down 20%. It’s going to draw down a lot more.”
• On indexing versus picking: in 2024 he built an “AI obvious basket” — about 30 roughly equal-weighted names including ASML, Micron, hyperscalers, and Alibaba — “and that thing has just so drastically outperformed and I literally haven’t changed it in two years.” Maybe theme baskets are the future of active management, Citrini being somewhere in the middle — though he notes ARK was a major accumulator of capital on this thesis and “also lost more money than anyone in human history.”
• On the Magnificent Seven, Dempsey says Meta is probably best positioned to monetize AI regardless of whether it builds frontier models, because its products have so much surface area and are deeply embedded. He is bullish on Google’s structural advantages, including GCP and its Gemini Flash models, and thinks Apple is well positioned; he is less high-conviction on Nvidia and memory stocks. He expects meaningful dispersion among next-generation neoclouds based on competence, execution, financing relationships, and how much they verticalize.

12. Crypto: tokens and equities converge, and AI raises crypto’s floor most

• Compound still runs long/short crypto in the fund: bullish on “a small subset of projects that will be able to ascribe fees to the tokens,” expecting “far more power-law-looking things in crypto over time” — call it DeFi, DePIN, some DeSci, and protocols built on top of L1s — plus still a Bitcoin bull. Venture-side honesty: “we still don’t have a flow of talent that’s aggressively coming, but my hope is that will change.”
• The long-term convergence view: “on a long-term time horizon, public equities and crypto tokens look like the exact same asset” — and going public, becoming liquid, reporting to shareholders, and having responsibility beyond a small group of price makers and buyers can be good for companies, even as many founders want no part of it.
• His most interesting crypto claim: AI raising the floor of intelligence “should be significantly more valuable for crypto than any other area,” letting people without traditional credentials build durable companies with frontier intelligence alongside them rather than relying on “a Discord of degens yelling at them.” The sector is held back for now by regulatory weirdness and low status in the industry.

13. Robotics redux, the humanoid overreach, and the drone blind spot

• The honest post-mortem: VR was what he was most purely wrong about (“I love VR and it just never came through”); robotics was mistimed — 2016–18 investments expected deep learning to deliver generalizable on-policy learning, “an incorrect thesis,” making them “by definition 8 to 10 years too early.” There was some salvage: Hyphen in food and SparkAI, acquired by John Deere. The cruel dynamic: relationships built when robotics was dead became the field’s most important people — who then raised very large seed rounds Compound couldn’t touch. Now “this time is really different”: a stealth, vertically integrated frontier lab betting on scaling internal data and sample-efficient, low-cost training, plus verticals like Alquist — retail, data centers, and semiconductors — and sub-$5,000 task robots that “give leverage to humans.”
• On Andrew Kang’s “$0 to $10 trillion” humanoid call, which José says he has heard Dempsey call insane: humanoids make sense where the thing being replaced is not labor cost but “the cost of the human life” — military or industrial — and industrial is robotics’ most heavily penetrated market already. Dempsey says Kang is running a playbook seen many times before, particularly if the incentive is to draw people into a basket of companies.
• On why drone warfare remains underdiscussed, José supplied the Ukraine examples: strikes roughly 3,000 km into Russia at orders-of-magnitude cost advantages and a possible shift in military spending. Dempsey’s “super stupid thing to say” that isn’t: “sci-fi people don’t write about one-way drone strikes very often… art imitates life which imitates art, and that’s just not in the lexicon.” Dempsey said Compound had looked at drone defense; José said Delphi had looked at it, had no strong view, and had no investments. “There’s only so many things people can be excited about at once.”

14. Calm obsessiveness: introspection as investing infrastructure

• On therapy and psychedelics making him a better investor and firm-builder: touching “the edges of your life experiences” preserves the humanity lens — how humans actually progress versus “the maximally autist, highly structured, obviously intellectual way in which humans will move through the world.” In response to José’s Marc Andreessen framing that inner work might kill hunger: “certainly not… you need to operate from a place of calmness, but calm obsessiveness” — otherwise you burn out, and that’s where bad investing mistakes happen.
• The moral thread from his recent writing on tech underinvesting in legitimacy: people with front-row privilege keep deferring — “when we’re 50 and 60 is when I’ll think about taking care of the world” — while tech becomes “more insular… that’s not where change happens.” Treat life as an infinite game instead. José adds his own shift: acceleration-era tech-hero pandering has made him question whether “economic efficiency is the only objective and a justification in itself.”
• His writing advice, worth stealing: “You should default to feeling like everything you read from your past self is embarrassing in some way. And if it’s not, it probably means you’re not progressing enough” — people around you “are just looking to get a little more resolution always,” and some thoughts only come out on a page “because there’s no one there to stare at you and react to it.”