Growth Is Only the Beginning: Who Captures the Returns in AI and Venture?

1. Guest Background

Harry Stebbings hosts this nearly hour-long 20VC conversation with participants identified in the supplied evidence as Jason and Rory. Their first names are established; their surnames, formal positions and organizational affiliations are not. The appropriate introduction is therefore based on their contributions to this discussion, rather than an inferred professional biography. This is an analysis of venture markets and the AI capital chain, with multiple companies serving as examples.

Jason brings together investing anecdotes and intensive product use. He describes an investment in an AI business that raised approximately $3 million and plans to stop fundraising. He also reports using Replit for roughly 14 to 15 hours a day and says he missed Modal’s financing after requesting a business day to consider a transaction completed over a weekend. These accounts help explain his attention to ownership, model costs and production workflows. They remain self-reported experience, rather than independently established evidence of market-wide outcomes.

Rory repeatedly tests the assumptions underneath investment enthusiasm. His contributions address the number of large exits available to funds, the relationship between compute construction and demand, and the incentives created by listing an asset manager. He also explicitly declines to offer stock recommendations on limited research. That restraint matters to the episode’s interpretation: the conversation supplies competing analytical frameworks, not a researched portfolio or a unified investment verdict.

2. What the Episode Covers

The episode begins with a financing puzzle. Harry describes an anonymous company with approximately $20 million in annual recurring revenue, rapid growth and profitability, yet little venture interest. The transcript’s additional phrase “700% profitable” is ambiguous and cannot establish a profit margin. Rory challenges the idea that investors have simply stopped taking risks: projects involving space-based data centers and quantum computing still attract capital. In his account, some specialized software businesses face a different problem—investors fear that AI could undermine their markets.

From there, the discussion examines where capital and risk sit within AI. CoreWeave introduces the relationship between infrastructure spending, debt and demand. Anthropic introduces the quality of model revenue. Harry cites third-quarter revenue of $16.8 billion, 45% sequential growth, an 85% gross margin and adjusted profitability, alongside a rumored IPO valuation reduction from $2 trillion to $1.5 trillion. These are figures relayed in the conversation. Rory’s subsequent exercise, assuming approximately $5 billion of additional revenue each quarter, is a conditional valuation scenario rather than company guidance.

Application platforms provide a more operational perspective. Jason interprets Replit’s default move toward OpenAI as a purchasing decision: roughly comparable output at a reported 30% to 40% cost advantage. He allows that routing could change again. The programming discussion then moves beyond generating a feature to getting it into production, encompassing review, testing, sandboxes and memory management. The question becomes how much work remains after software appears to function.

Modal opens the discussion of inference businesses serving open-weight models. Harry reports a $750 million financing at a $15.75 billion valuation and a revenue run rate exceeding $1 billion. The participants contrast this with an earlier scale of approximately $60 million to under $100 million. Their estimate of $5 billion to $6 billion in combined run-rate revenue across major inference providers, and their comparison using hypothetical supplier gross margins of 30% and 70%, illustrate opportunity rather than establish audited market economics.

The conversation also tests enthusiasm against price. Jason passes on the proposed Town investment, described as a roughly $108 million Series B at a $1 billion valuation, because he finds the available data insufficient. Subsequent remarks about enormous inference budgets are sarcastic, not verified operating plans. Bending Spoons offers a different model: selectively acquiring and operating software over long periods. The participants disagree over whether software can possess the durability implied by the Berkshire and See’s Candies analogy.

Cash distributions and capital recycling complete the frame. A claim that 37signals generated $370 million in profit since 2020 is challenged during the discussion and may reflect an inference from employee profit sharing. Harry’s account of Harvard’s endowment and SpaceX exposure leads to speculation that liquidity could return to venture funds. The closing ideas about a Tesla–SpaceX combination and shares for children likewise remain predictions or relayed proposals, including their supposed marketing effects.

3. Core Views: Reasoning, Examples, and Limits

The strongest argument concerns the distance between a successful company and a successful fund investment. Jason argues that funds above $5 billion need billion-dollar positions. His Listen Labs example makes the constraint tangible: a company worth around $2 billion would require ownership approaching half the business to supply that position. A founder’s excellent outcome can consequently be too small for a large fund’s economics. Rory challenges the scalability of this strategy, citing his claim that the US technology industry historically produces roughly 10 to 20 decacorns per decade. His objection is about how many funds can depend on scarce outcomes simultaneously. Neither the historical count nor the future supply of giants is independently established here, so the exchange identifies a structural tension rather than proving that large funds must fail.

The seed discussion applies the same reasoning to ownership. A $300,000 allocation at a $100 million valuation may remain insufficient to transform a fund even when the company later reaches $18 billion. Entry price, allocation and follow-on access interact; a spectacular company valuation chart leaves those variables invisible. Rory maintains that early risk should earn higher multiples over long periods, while the participants accept that recent years may have produced the reverse. Their agreement is a period-specific judgment, not a demonstrated conclusion about the permanent inferiority of seed investing.

CoreWeave exposes a related distinction between growing demand and demand growing fast enough. Rory describes a business that must build ahead: moving slowly disappoints customers, while moving too quickly leaves capacity vulnerable to even modest deceleration. Debt makes that timing problem less forgiving. His suggestion to combine OpenAI and Anthropic revenue is intended to reduce distortion from share moving between them. It is an indicator, not a complete measure of AI demand or infrastructure utilization. His remembered financial figures also cannot support a precise valuation. Jason’s preference for Nvidia’s core technology adds a competitive argument, but Rory correctly separates different supply-chain positions from the question of which security delivers a return at a particular price.

Model revenue brings customer choice into the analysis. Jason argues that developer switching makes recurring revenue less predictable than the label ARR might suggest. His example of Anthropic losing 10% of share is explicitly a possibility, with no clear distinction between a relative decline and percentage points. Rory extends the mechanism: slower end demand could encourage price cuts to defend share, pressuring growth and profitability together. Replit gives the argument an operational example. Under Jason’s account, a platform switches its default when quality is sufficiently similar and costs improve. That supports continued competition; it does not establish a permanent model ranking. His negative inference about Kimi from its absence as a universal default also leaves integration and reliability explanations unresolved.

In programming, the episode locates value in the unfinished work after generation. Jason’s description of a feature being 95% complete before three days of testing and repair is experiential, not a universal productivity measurement. Its analytical purpose is to distinguish a convincing demonstration from production readiness. Independent review products may benefit from one model checking another, yet Claude Code Review illustrates the platform-entry threat. Rory argues that proximity to foundation models creates both exceptional use cases and competitive exposure. The technical importance of review therefore does not establish that an independent vendor will retain pricing power or produce the largest investment return.

Inference services face a comparable question about who keeps the margin. Participants imagine neo-cloud operators extending from supplying compute into serving developers directly. Demand can expand while competition changes the economics of the intermediary. Fireworks’ suggested mid-30s gross margin and possible move toward 50% are speculative, and the conversation does not establish that Baseten, Fireworks and Modal will become a closed group of winners. The hypothetical 30%-versus-70% supplier-margin comparison only illustrates procurement incentives under comparable capability; it is not evidence of actual customer prices.

Finally, valuation requires a future operating path and an eventual multiple. Using the episode’s roughly $90 billion and 40-times-revenue figures for Anduril, the participants explain how repeated revenue doubling might justify a premium. The vulnerability is assuming both sustained growth and a favorable exit valuation. Jason’s concern that AI marketing businesses could converge toward lower sector multiples even while growing 60% to 70% is a risk scenario, not an observed outcome. Across these examples, strong execution can coexist with a disappointing investment if the entry price already consumes too much of the future.

4. Learning and Application

For founders, the practical starting point is investor fit. Ask what position value a fund needs, what ownership it could realistically retain and what company outcome would satisfy that combination. This makes rejection more informative than simply comparing growth rankings. Jason’s suggestion to approach around 100 investors is best treated as a tactic for widening the search, not a quota or a financing guarantee. Investors can reverse the exercise: calculate proceeds from the actual check, ownership, expected dilution and follow-on rights, then compare them with fund size. Company appreciation alone does not answer that question.

For infrastructure analysis, place demand assumptions beside the construction schedule and capital commitments. Test what happens when demand continues to grow but grows more slowly than planned. For model businesses, vary total demand, vendor share and price separately. Combining OpenAI and Anthropic revenue may reduce one source of share distortion, but cannot establish whether a particular operator’s entire capacity will be absorbed. Platform routing provides another observation point. A useful procurement exercise compares quality and total cost on the buyer’s own tasks while retaining integration and reliability as possible explanations for vendor selection. One default change cannot settle long-term model leadership.

Software teams can apply the build-to-production argument by measuring a complete delivery cycle. Track the effort after initial generation: review, testing, repairs and preparation for production. If the bottleneck lies there, evaluate whether cross-model checks, testing tools or sandboxes reduce total rework. Do not insert Jason’s 95% or three-day example into a project estimate as an established benchmark. Buyers need workflow evidence; vendors additionally need a reason customers will keep paying when a foundation-model company offers overlapping functionality. Solving an important problem and sustaining a profitable independent business require separate validation.

For expensive investments, write down the revenue destination implied by the entry price, then change both growth and exit multiples. Include earlier deceleration and convergence toward sector comparables. In inference, also test an upstream supplier entering the developer-service layer rather than assuming margins rise with revenue. The purpose is to reveal which conditions carry the return, not to turn the podcast’s verbal estimates into precise forecasts.

For cash-generating founders, financing should follow the desired distribution of value. Jason’s Zapier discussion makes complete self-funding attractive when a smaller team can operate without much outside capital. The participants also acknowledge that some founders need the initial million or two to begin. Buying investors out later is not a costless reset: at six times annual company cash flow, purchasing a 30% stake consumes approximately 1.8 years of all that cash flow, before considering other operating needs. Compare dividend arrangements, cash retention and borrowing costs before choosing between no financing, one round and later liquidity.

Fund partners and LPs can similarly distinguish investment performance from management-company growth. Rory argues that public investors give more credit to management fees than carry, potentially encouraging AUM expansion and additional products. His listing thresholds and cultural concerns are judgments rather than universal rules, but they offer useful questions about incentives. A strategy can benefit the management company without serving exactly the same objective as an individual fund or a founder seeking distributions. Those outcomes should be evaluated separately.

Source

• Original episode: The New VC Game | CoreWeave Sell-Off Creates Panic | Revenue Multiples Out of Control