TheChek News Flow

I contend most AI investments will again lose money. But, again, given the size of the opportunity in AI transformation, more money will be made than lost, even more asymmetrically than in the past technology transitions. The bigger the transition, the bigger the opportunity and the bigger the bubble. The key is to be invested in one or more of the “megacorn alpha dogs”. And what matters is not the exit valuation but the multiple on investment, weighted by the dollars invested at each multiple.

There are many companies today with valuations, even seed valuations, in the billions.

Just in the western world, mostly the U.S., the major AI models: OpenAI, Anthropic, Sakana, Cohere, Adept, Character, Inflection, Scale, xAI, Mistral, Stability, Perplexity, Thinking Machine Labs, SSI …. the major coding companies: Cursor, Cognition, Replit, Lovable, Windsurf, Poolside, … the major robotic companies: Figure, 1X, Pi, SkillD, FieldAI, Rhoda, Agility,… the major self-driving companies: Waymo, Waabi, Aurora, Cruise, … the early enterprise, media and healthcare winners: Abridge, Sierra, Decagon, Glean, Sword, Hippocratic, Ambience, Runway AI, Eleven Labs, Synthesia, Databricks, Andruil, … And then there are the >$100 million “entry” valuations galore. In a market with 95% losing investments and 5% winning valuations, one grossly needs to get to 40X multiple on a big win to get a 3X net or 4X gross on the fund. Usually, the higher the valuation the higher the dollars invested by venture funds so weighted multiples are even harder to get with larger late stage investments. Usually later stage investments have a lower “fail” ratio so the math works. In AI the historical late stage “fail ratio” might not be true.

What does this say about valuations? Capital concentration is extreme: the largest rounds are in foundation models, compute/infra, defense, and (more recently) humanoid robotics, with valuations that rival or exceed historic IPO peaks. There will only be a few winners. Healthcare, coding, and enterprise search are the next densest clusters of unicorn/decacorn deals. I suspect there will be more winners but they will be smaller in IPO valuations or M&A exit valuations than their current valuations. Many of their later rounds may not get returns.

Power-law reality check: Empirical datasets show that a small fraction of investments drive the majority of returns (e.g., ≈6% of deals generated ~60% of VC returns in Horsley Bridge’s long-run sample). AI investments, especially multiple, may be even more lopsided given the high entry valuations for investors.

For venture fund limited partners (LPs) there is the urgent matter of DPI vs TVPI.

DPI (Distributions to Paid-In Capital) is the “How many dollars have we gotten back for every dollar we’ve put in so far?” TVPI (Total Value to Paid-In) equals distributions to LPs plus net asset values divided by capital paid in by limited partners. Many limited partners have discovered that DPI is truth-in-cash. It can’t be marked up. It only moves when assets exit and proceeds are distributed (cash or in-kind stock at distribution-date value). In-kind public stock or cash distributions lock DPI at distribution date; post distribution stock price movements are outside DPI so they don’t impact DPI. TVPI is a mostly hypothetical mark, especially in an inflated valuation environment — exactly where cycles (like 2024–25 AI) can inflate numbers. If DPI is low but TVPI is high, returns are still largely fictional paper returns and can decline dramatically.

Limited partners during the last two years are focusing mostly on DPI and clamoring for it. Very few funds are meeting LP thresholds on this metric and that is bad news for the venture business, both for limited partners and general partners. As (and if) the cycle normalizes over the next ten years, venture capital allocations will decline, venture fund raises and investments will decline, and innovation will be starved of capital (as is currently happening in biotech investments). Enthusiastic and naive efforts like $100 billion “venture funds” and abundant tens of billions of Middle East money funded easy “AI fund” raises will do a second take. Yes sir!

Yet, society will see massive transformation. Trillions of wealth will be created and AI-driven societies, those that choose to transform, will be much better off, especially the bottom half of the global population.

So what is an investor to do, be they a venture firm, a limited partner or an individual investor?

How does one underwrite paying up? Bet on the long run and long-term value creation, not short-term TVPI. Access to the best investment opportunities will be through the best, yet more cautious, venture firms. These are hard for investors to get into except when they raise super large funds. Be cautious. For every billion dollar pre-money valuation investment that is worth making, there are many many billion dollar valuation investments that are not worth making.

Highly selective and nuanced decision making on these assessments are skills hard to come by. It is a very hard skill to get right and we continuously doubt our own decisions. Firms that are highly selective but not overly valuation sensitive in what they invest in will do well, not those who are only looking for bargains. Bottom line: Bargain hunting will be a poor bet in AI. But firms “putting money to work” who invest in all high valuations indiscriminately will do poorly and that is the majority of investing that is happening. This danger will come home to roost in five to seven years in realized returns and DPIs.