Summary

For the first time in David George’s career, capital itself compounds a company’s advantage — making the power law more extreme than it has been over the last 10–20 years of tech investing. His mechanism: the old failure mode was throwing money at headcount (“hire a thousand people” and drown in coordination issues); now “you can throw dollars at compute and compute can make products and the businesses better,” fueled by what Aram Verdiyan calls “unlimited demand for inference.” The three frontier companies — SpaceX, OpenAI, Anthropic — represent $3.5–5T of potential enterprise value; Jen says many LPs and the broader institutional allocator community had little exposure to SpaceX before it went public.
Of 3,000 US VC firms, only 20 achieved consistent 3x net returns over two decades — while the Cambridge average venture return over the last 10 years is 1–2x. David’s LP construction implication: concentrate in those 15–20 firms; a 50–70-fund LP portfolio can’t beat the average, and at 1–2x “you’ll do better in private equity” without the 10-year lockup. The consistent winners all had access to category-defining companies every vintage — and sized them at 5–10%+ of late-stage funds so a single company can return the fund.
AI’s TAM is labor, not software — the US spends roughly 40x more on labor, so calling AI “the next evolution of software is far too limiting.” AI hit $100B in revenue in four years versus SaaS’s fifteen, attacking $30T of GDP at once; Aram admits “I’ve been chronically wrong about how big these outcomes can get.” George refuses layer-war framing — “I think everything might work” — but within any category the power law holds and “second place is playing for scraps.”
Traction has never been harder to parse: zero-to-$5M-ARR-in-a-month companies with no renewal cycle, cohorts selling to each other, “not even ARR but multiplying by 12.” For every 49 of those there’s 1 real one — Cursor raised a ~$400M round on ~$3M ARR and was still being called dead “even the morning of the acquisition announcement” of its sale to SpaceX for $60B. George’s test isn’t financial analysis but the post-it on his screen: “is the market demanding more of your product?”
The pre-ChatGPT software cohort is stranded: 2021–22’s $200–300B of software LBOs at 25–32x EBITDA are “worth probably half that,” which is driving redemptions in private credit. Only 15–20 public SaaS names still trade above 10x revenue, and one point of growth is worth three points of EBITDA. The fix — Intercom’s founder-led AI-native rebuild — amounts to “suiciding your existing business,” an n-of-one so far; bolting on AI via an operating partner “just doesn’t work.”
LP and GP incentives are diametrically opposed: a GP gets fired for missing the next Facebook; “you don’t get fired for investing in IBM if you’re an LP” — or often for not investing at all. Missing the frontier models leaves an allocator only slightly below benchmark and still employed, which explains the exposure gap. Aram’s case: AI should be “core or a super core,” not a satellite — CalPERS is now making up lost time, moving public markets from 91% to 58% and venture/growth from 9% to 43%.
Diffusion is the bull case: the median US company spends $12 per employee per month on AI; the top 1% spends $7,000. George’s steelman for legacy SaaS is that coding may be “a head fake” — perfectly documented, verifiable, simulatable, attributes “most tasks in business do not share” — but the adoption gap plus the fastest-growing companies they’ve ever seen, adding more revenue per month than megacap tech companies on perhaps 10–30M users, makes him “super super super bullish.”
The next $100T of market cap comes from robotics (“bigger than the language stuff,” within 10 years), autonomy (fewer than 10,000 Waymos live in the US), healthcare (18% of GDP, barely scratched), and solving the supply bottleneck. Today’s chatbot is “the skeuomorphic version” of consumer AI. Aram: demand isn’t the constraint — “the US doesn’t have a problem with energy generation. It has a problem with speed to power”: permissioning and transmission, where other countries deploy 10x more renewable capacity a year.

Deep dive

1. Capital now compounds advantage — the power law went systemic

Jen Kha’s framing of the episode: power law “used to be just a feature of a cottage industry in venture capital and now it’s systemic,” with the three frontier model companies — SpaceX, OpenAI, Anthropic — representing “somewhere between three and a half to 5 trillion dollars of potential enterprise value,” to which many LPs and the broader institutional allocator community had little exposure before SpaceX went public.
David George’s mechanism for why the power law is now more extreme than in 10–20 years: increasing returns to scale always existed, but “for the first time in my career, you can take capital and throw it at a company and it compounds their advantage.” The classic way to ruin a startup — throw money at it, hire a thousand people, drown in coordination overhead — no longer binds: “you can throw dollars at compute and compute can make products and the businesses better.”
Aram Verdiyan’s demand-side explanation: “unlimited demand for inference.” AI reached $100B in revenue in four years versus SaaS’s fifteen, “and we’re not even close” on penetration. Outcome sizes have repriced accordingly — top-decile outcomes went from ~$10B to ~$40B, “soon to be probably 100 billion by the time Anthropic and then OpenAI come out” — and the last cycle’s $25T of new market cap should be exceeded by this one.

2. The TAM is labor, not software — and it’s not zero-sum

George on scale: the US economy spends something like 40x more on labor than software, so “to equate it to software and say oh it’s the next evolution of software is far too limiting.” Labor won’t disappear — it gets reinvented. Aram’s healthcare version: healthcare IT is $60–100B a year, but AI hits the tasks themselves — claims, billing, administration, a trillion-dollar industry — so AI’s TAM can be 10x+ traditional SaaS. His confession: “I’ve been chronically wrong about how big these outcomes can get.”
Aram’s expansionary example, as told by their legal counsel: “I love Harvey. All my clients think they’re lawyers now… my billable hours have only gone up with the advent of AI.”
On which layer of the stack wins, George’s honest non-answer: “I don’t know, the market is going to be so big. I think everything might work.” He explicitly rejects zero-sum reads (open source winning ≠ labs losing) — but within a category the power law is brutal: winners take the vast majority of share and “second place is playing for scraps.” Loss tolerance is the corollary: ~60% loss rates in their best early-stage funds, 10–20% at growth — “if we’re not losing money… we’re not taking enough risk.”

3. 3,000 firms, 20 consistent winners: access, sizing, and the death of the middle

Aram’s dataset is the episode’s anchor: of 3,000 US venture firms, only 20 — under 1% — had consistent 3x-net-TVPI performance, requiring three to four 3x-net-TVPI funds over a 20-year span. The consistent ones “consistently had access to the category defining companies every vintage.” Cambridge data puts the average venture return over 10 years at 1–2x: “you’ll do better in private equity. You’ll definitely do better in the public markets. You don’t need to lock up your money for 10 years.”
The logo alone isn’t sufficient: early-stage funds must own enough; late-stage funds must size the best company at 5–10%+ so one position can return the fund. “Fund returning math in late stage didn’t exist before. It now does.”
“Death of the middle,” per the discussion: hyper-specialized early-AI funds with deep domain experts have done well, and full-stack platforms (seed through IPO) work — “everything else in between… struggles to compete.” Eddie’s tweet, read aloud: interest in big VC funds “has been driven by founders, not LPs” — founders want the brand that can scale, be a life-cycle investor, and land customers and hires. George’s flywheel: domain expertise wins the deal, 700 employees of operating resources (fees reinvested) bend the outcome, and killer references create persistence of returns.

4. Late-stage franchises are built on early-stage ball control

George, self-described as “very biased”: “our business starts and ends with early stage” — the growth fund’s access, information, and relationships all flow from it. Aram agrees from the LP seat: “it’s really hard to come in as a de novo late stage firm and write a $500 million check”; the 5–10% concentrated late-stage position exists because the early franchise built the founder relationship years earlier.
Jen’s proposed coexistence theory for pre-seed: at sub-$20–40M valuations with sub-$100M funds, small firms can win a round or two before the big platforms, which rationally wait among seven look-alike AI startups until they can lead the A or B of the category winner. Aram endorses coexistence, citing healthy seed relationships, their own chunkier seeds, and Speedrun. He says founders “very much are hoping to stay in the orbit” of the brand, and the firm “may not actually do the investment but we have to at least understand the landscape” to make informed later-stage decisions — the ball-control rationale.

5. Traction fog: real vs. misleading ARR

Aram: “AI is actually making our jobs harder than ever before.” Rounds are larger and faster, and the traction is confusing — a company out of an accelerator claims “zero to five million ARR in a month” with no renewal cycle, sometimes selling to its own cohort, “and it’s not even ARR, but they’re multiplying by 12.” For 49 of those there’s one special company doing a couple million of actual ARR that becomes the next Cursor.
Cursor is the specimen: roughly $3M ARR raising a ~$400M round, widely mocked — “even the morning of the acquisition announcement people were still saying that Cursor is dead. I’m like, they just announced that they were going to be acquired by SpaceX for $60 billion.”
George’s method when a company has sold for only a couple of months: “you’re not going to be able to do it with financial analysis” — it’s founder judgment plus customer texture. His screen post-it: “is the market demanding more of your product?” Harvey is the worked example: strong early commercial logos but “the usage was not very good… mediocre” — then post-reasoning models “that totally flipped,” from fear of hallucinations to “every client is actually demanding the law firms use the product.” “Everyone can do cohort analysis… but understanding the texture of the market and what the customers actually want — that’s how you make the decision.”

6. LP incentives, concentration, and the liquidity question

Jen’s structural point on why allocators lag: a GP “can get fired for missing out on the next Facebook, the next Uber” — omission is fireable — while “you don’t get fired for investing in IBM if you’re an LP,” and potentially not for failing to invest at all. Miss the frontier models and you’re merely slightly below benchmark, still employed. Hence Aram’s allocation stance: AI “is not a satellite position. You should be core or a super core.”
The construction case: with 20 winners out of 3,000, LPs should concentrate in 15–20 firms. A 50–70-firm portfolio can’t beat the average. Aram’s sizing point is equally critical, and David’s example shows the failure mode: an LP finds the right fund and puts 1% in — “Great. You 10xed it. It returns 10% of your fund. It does not move the needle at all.” CalPERS, having “famously lost out on billions,” has shifted public markets from 91% to 58% and venture/growth from 9% to 43%.
The liquidity pushback and its counter: unicorns stay private 10+ years and an IPO isn’t a distribution — 12–24+ months more, especially owning 10–15%. But “would you have wanted to sell Stripe, Databricks… three, four years ago? The answer is unanimously no.” Anthropic, first funded in 2021, is “about to go public 5 years later.” George’s Fund I story: at year 16 they offered every LP liquidity on their seed-stage Stripe position — “every single one of those LPs said no, we’d rather let this continue to compound” — and the fund finally exited at year 17.

7. The pre-ChatGPT reckoning: stranded SaaS, LBOs, and private credit

George’s public-market read: only 15–20 SaaS companies trade above 10x revenue — “it used to be dozens and dozens” — and nearly all show AI-driven growth acceleration. His firm’s data: “1 percentage of growth in the public markets is equivalent to three percentages of EBITDA,” a reversal from the 2021 profitability focus.
The stranded cohort: 2021–22 saw $200–300B of software LBOs with $200B+ of debt at 25–32x EBITDA average; “those companies today are worth probably half that. The reason you’re seeing redemptions in the credit markets in private credit is exactly that.” Aram’s hypothetical to George — a 2016–21 vintage company growing 30%, marked 10–20x on venture books, “Silver Lake has no interest in that company anymore. They would have a year ago” — draws an honest “it’s very TBD… we have a lot of exposure to those companies too,” though ~95% of the firm’s NAV sits in the accelerating cohort.
The turnaround template is brutal: Intercom brought the founder back, rebuilt AI-native, and scaled — “it’s almost like you’re suiciding your existing business, which in private equity is really hard to do.” George’s high-five to the founder: “You did it, man… it’s an n of one right now.” George’s warning against AI-washing PE: “just because you put Sears on a website didn’t make it Amazon” — bolt-on AI customer-service agents without workflow churn customers, NPS drops track revenue drops, and the spiral compounds under debt. Aram: “you can’t just throw an operating partner at the company and say let’s put AI on it.”

8. Diffusion at 1%, and where the next $100T gets created

George’s steelman for slow change — worth keeping: coding may be “a head fake, right? Coding is perfectly documented… it’s verifiable and it’s simulatable. Most tasks in business do not share those three attributes,” so diffusion into other knowledge work could take much longer. Yet the same data makes him “super super super bullish”: the median US company spends $12 per employee per month on AI, the top 1% spends $7,000; cutting-edge banks are at maybe 1% of headcount cost. These are “the fastest growing companies we’ve ever seen… of all time,” adding more revenue per month than megacap tech companies on the back of perhaps 10–30M users, against 150M workers in the US.
Asked for the next $100T company, George demurs (“that’s probably two tech cycles away”) but maps the whitespace: consumer AI’s chatbot is “the skeuomorphic version” — the native version will be proactive and do work on our behalf; “we are nowhere on robotics, but I think robotics is going to be bigger than the language stuff” within 10 years; fewer than 10,000 Waymos are live in the US; healthcare at 18% of GDP has barely been scratched on both care delivery and drug discovery.
Aram’s closing addition — the bottleneck is supply, not demand: energy, grid, and data centers upstream of chips and models. “The US doesn’t have a problem with energy generation. It has a problem with speed to power” — permissioning, transmission, regulation, while other countries deploy 10x more renewable capacity a year. That’s where “not $10 billion, $50 billion, but $100 billion-plus opportunities” can be created, and why “this is not the dotcom or Covid — the traction is real and it’s not ephemeral revenue.” Jen’s sign-off: “It’s time for machine age. Let’s bring the machines.”

Coinbase’s Everything Exchange: Agentic Finance, Stablecoins & Tokenization with CEO Brian Armstrong